2 Linux Ethernet Bonding Driver HOWTO
4 Latest update: 27 April 2011
6 Initial release : Thomas Davis <tadavis at lbl.gov>
7 Corrections, HA extensions : 2000/10/03-15 :
8 - Willy Tarreau <willy at meta-x.org>
9 - Constantine Gavrilov <const-g at xpert.com>
10 - Chad N. Tindel <ctindel at ieee dot org>
11 - Janice Girouard <girouard at us dot ibm dot com>
12 - Jay Vosburgh <fubar at us dot ibm dot com>
14 Reorganized and updated Feb 2005 by Jay Vosburgh
15 Added Sysfs information: 2006/04/24
16 - Mitch Williams <mitch.a.williams at intel.com>
21 The Linux bonding driver provides a method for aggregating
22 multiple network interfaces into a single logical "bonded" interface.
23 The behavior of the bonded interfaces depends upon the mode; generally
24 speaking, modes provide either hot standby or load balancing services.
25 Additionally, link integrity monitoring may be performed.
27 The bonding driver originally came from Donald Becker's
28 beowulf patches for kernel 2.0. It has changed quite a bit since, and
29 the original tools from extreme-linux and beowulf sites will not work
30 with this version of the driver.
32 For new versions of the driver, updated userspace tools, and
33 who to ask for help, please follow the links at the end of this file.
38 1. Bonding Driver Installation
40 2. Bonding Driver Options
42 3. Configuring Bonding Devices
43 3.1 Configuration with Sysconfig Support
44 3.1.1 Using DHCP with Sysconfig
45 3.1.2 Configuring Multiple Bonds with Sysconfig
46 3.2 Configuration with Initscripts Support
47 3.2.1 Using DHCP with Initscripts
48 3.2.2 Configuring Multiple Bonds with Initscripts
49 3.3 Configuring Bonding Manually with Ifenslave
50 3.3.1 Configuring Multiple Bonds Manually
51 3.4 Configuring Bonding Manually via Sysfs
52 3.5 Configuration with Interfaces Support
53 3.6 Overriding Configuration for Special Cases
55 4. Querying Bonding Configuration
56 4.1 Bonding Configuration
57 4.2 Network Configuration
59 5. Switch Configuration
61 6. 802.1q VLAN Support
64 7.1 ARP Monitor Operation
65 7.2 Configuring Multiple ARP Targets
66 7.3 MII Monitor Operation
68 8. Potential Trouble Sources
69 8.1 Adventures in Routing
70 8.2 Ethernet Device Renaming
71 8.3 Painfully Slow Or No Failed Link Detection By Miimon
77 11. Configuring Bonding for High Availability
78 11.1 High Availability in a Single Switch Topology
79 11.2 High Availability in a Multiple Switch Topology
80 11.2.1 HA Bonding Mode Selection for Multiple Switch Topology
81 11.2.2 HA Link Monitoring for Multiple Switch Topology
83 12. Configuring Bonding for Maximum Throughput
84 12.1 Maximum Throughput in a Single Switch Topology
85 12.1.1 MT Bonding Mode Selection for Single Switch Topology
86 12.1.2 MT Link Monitoring for Single Switch Topology
87 12.2 Maximum Throughput in a Multiple Switch Topology
88 12.2.1 MT Bonding Mode Selection for Multiple Switch Topology
89 12.2.2 MT Link Monitoring for Multiple Switch Topology
91 13. Switch Behavior Issues
92 13.1 Link Establishment and Failover Delays
93 13.2 Duplicated Incoming Packets
95 14. Hardware Specific Considerations
98 15. Frequently Asked Questions
100 16. Resources and Links
103 1. Bonding Driver Installation
104 ==============================
106 Most popular distro kernels ship with the bonding driver
107 already available as a module. If your distro does not, or you
108 have need to compile bonding from source (e.g., configuring and
109 installing a mainline kernel from kernel.org), you'll need to perform
112 1.1 Configure and build the kernel with bonding
113 -----------------------------------------------
115 The current version of the bonding driver is available in the
116 drivers/net/bonding subdirectory of the most recent kernel source
117 (which is available on http://kernel.org). Most users "rolling their
118 own" will want to use the most recent kernel from kernel.org.
120 Configure kernel with "make menuconfig" (or "make xconfig" or
121 "make config"), then select "Bonding driver support" in the "Network
122 device support" section. It is recommended that you configure the
123 driver as module since it is currently the only way to pass parameters
124 to the driver or configure more than one bonding device.
126 Build and install the new kernel and modules.
128 1.2 Bonding Control Utility
129 -------------------------------------
131 It is recommended to configure bonding via iproute2 (netlink)
132 or sysfs, the old ifenslave control utility is obsolete.
134 2. Bonding Driver Options
135 =========================
137 Options for the bonding driver are supplied as parameters to the
138 bonding module at load time, or are specified via sysfs.
140 Module options may be given as command line arguments to the
141 insmod or modprobe command, but are usually specified in either the
142 /etc/modrobe.d/*.conf configuration files, or in a distro-specific
143 configuration file (some of which are detailed in the next section).
145 Details on bonding support for sysfs is provided in the
146 "Configuring Bonding Manually via Sysfs" section, below.
148 The available bonding driver parameters are listed below. If a
149 parameter is not specified the default value is used. When initially
150 configuring a bond, it is recommended "tail -f /var/log/messages" be
151 run in a separate window to watch for bonding driver error messages.
153 It is critical that either the miimon or arp_interval and
154 arp_ip_target parameters be specified, otherwise serious network
155 degradation will occur during link failures. Very few devices do not
156 support at least miimon, so there is really no reason not to use it.
158 Options with textual values will accept either the text name
159 or, for backwards compatibility, the option value. E.g.,
160 "mode=802.3ad" and "mode=4" set the same mode.
162 The parameters are as follows:
166 Specifies the new active slave for modes that support it
167 (active-backup, balance-alb and balance-tlb). Possible values
168 are the name of any currently enslaved interface, or an empty
169 string. If a name is given, the slave and its link must be up in order
170 to be selected as the new active slave. If an empty string is
171 specified, the current active slave is cleared, and a new active
172 slave is selected automatically.
174 Note that this is only available through the sysfs interface. No module
175 parameter by this name exists.
177 The normal value of this option is the name of the currently
178 active slave, or the empty string if there is no active slave or
179 the current mode does not use an active slave.
183 Specifies the 802.3ad aggregation selection logic to use. The
184 possible values and their effects are:
188 The active aggregator is chosen by largest aggregate
191 Reselection of the active aggregator occurs only when all
192 slaves of the active aggregator are down or the active
193 aggregator has no slaves.
195 This is the default value.
199 The active aggregator is chosen by largest aggregate
200 bandwidth. Reselection occurs if:
202 - A slave is added to or removed from the bond
204 - Any slave's link state changes
206 - Any slave's 802.3ad association state changes
208 - The bond's administrative state changes to up
212 The active aggregator is chosen by the largest number of
213 ports (slaves). Reselection occurs as described under the
214 "bandwidth" setting, above.
216 The bandwidth and count selection policies permit failover of
217 802.3ad aggregations when partial failure of the active aggregator
218 occurs. This keeps the aggregator with the highest availability
219 (either in bandwidth or in number of ports) active at all times.
221 This option was added in bonding version 3.4.0.
225 Specifies that duplicate frames (received on inactive ports) should be
226 dropped (0) or delivered (1).
228 Normally, bonding will drop duplicate frames (received on inactive
229 ports), which is desirable for most users. But there are some times
230 it is nice to allow duplicate frames to be delivered.
232 The default value is 0 (drop duplicate frames received on inactive
237 Specifies the ARP link monitoring frequency in milliseconds.
239 The ARP monitor works by periodically checking the slave
240 devices to determine whether they have sent or received
241 traffic recently (the precise criteria depends upon the
242 bonding mode, and the state of the slave). Regular traffic is
243 generated via ARP probes issued for the addresses specified by
244 the arp_ip_target option.
246 This behavior can be modified by the arp_validate option,
249 If ARP monitoring is used in an etherchannel compatible mode
250 (modes 0 and 2), the switch should be configured in a mode
251 that evenly distributes packets across all links. If the
252 switch is configured to distribute the packets in an XOR
253 fashion, all replies from the ARP targets will be received on
254 the same link which could cause the other team members to
255 fail. ARP monitoring should not be used in conjunction with
256 miimon. A value of 0 disables ARP monitoring. The default
261 Specifies the IP addresses to use as ARP monitoring peers when
262 arp_interval is > 0. These are the targets of the ARP request
263 sent to determine the health of the link to the targets.
264 Specify these values in ddd.ddd.ddd.ddd format. Multiple IP
265 addresses must be separated by a comma. At least one IP
266 address must be given for ARP monitoring to function. The
267 maximum number of targets that can be specified is 16. The
268 default value is no IP addresses.
272 Specifies whether or not ARP probes and replies should be
273 validated in any mode that supports arp monitoring, or whether
274 non-ARP traffic should be filtered (disregarded) for link
281 No validation or filtering is performed.
285 Validation is performed only for the active slave.
289 Validation is performed only for backup slaves.
293 Validation is performed for all slaves.
297 Filtering is applied to all slaves. No validation is
302 Filtering is applied to all slaves, validation is performed
303 only for the active slave.
307 Filtering is applied to all slaves, validation is performed
308 only for backup slaves.
312 Enabling validation causes the ARP monitor to examine the incoming
313 ARP requests and replies, and only consider a slave to be up if it
314 is receiving the appropriate ARP traffic.
316 For an active slave, the validation checks ARP replies to confirm
317 that they were generated by an arp_ip_target. Since backup slaves
318 do not typically receive these replies, the validation performed
319 for backup slaves is on the broadcast ARP request sent out via the
320 active slave. It is possible that some switch or network
321 configurations may result in situations wherein the backup slaves
322 do not receive the ARP requests; in such a situation, validation
323 of backup slaves must be disabled.
325 The validation of ARP requests on backup slaves is mainly helping
326 bonding to decide which slaves are more likely to work in case of
327 the active slave failure, it doesn't really guarantee that the
328 backup slave will work if it's selected as the next active slave.
330 Validation is useful in network configurations in which multiple
331 bonding hosts are concurrently issuing ARPs to one or more targets
332 beyond a common switch. Should the link between the switch and
333 target fail (but not the switch itself), the probe traffic
334 generated by the multiple bonding instances will fool the standard
335 ARP monitor into considering the links as still up. Use of
336 validation can resolve this, as the ARP monitor will only consider
337 ARP requests and replies associated with its own instance of
342 Enabling filtering causes the ARP monitor to only use incoming ARP
343 packets for link availability purposes. Arriving packets that are
344 not ARPs are delivered normally, but do not count when determining
345 if a slave is available.
347 Filtering operates by only considering the reception of ARP
348 packets (any ARP packet, regardless of source or destination) when
349 determining if a slave has received traffic for link availability
352 Filtering is useful in network configurations in which significant
353 levels of third party broadcast traffic would fool the standard
354 ARP monitor into considering the links as still up. Use of
355 filtering can resolve this, as only ARP traffic is considered for
356 link availability purposes.
358 This option was added in bonding version 3.1.0.
362 Specifies the quantity of arp_ip_targets that must be reachable
363 in order for the ARP monitor to consider a slave as being up.
364 This option affects only active-backup mode for slaves with
365 arp_validation enabled.
371 consider the slave up only when any of the arp_ip_targets
376 consider the slave up only when all of the arp_ip_targets
381 Specifies the time, in milliseconds, to wait before disabling
382 a slave after a link failure has been detected. This option
383 is only valid for the miimon link monitor. The downdelay
384 value should be a multiple of the miimon value; if not, it
385 will be rounded down to the nearest multiple. The default
390 Specifies whether active-backup mode should set all slaves to
391 the same MAC address at enslavement (the traditional
392 behavior), or, when enabled, perform special handling of the
393 bond's MAC address in accordance with the selected policy.
399 This setting disables fail_over_mac, and causes
400 bonding to set all slaves of an active-backup bond to
401 the same MAC address at enslavement time. This is the
406 The "active" fail_over_mac policy indicates that the
407 MAC address of the bond should always be the MAC
408 address of the currently active slave. The MAC
409 address of the slaves is not changed; instead, the MAC
410 address of the bond changes during a failover.
412 This policy is useful for devices that cannot ever
413 alter their MAC address, or for devices that refuse
414 incoming broadcasts with their own source MAC (which
415 interferes with the ARP monitor).
417 The down side of this policy is that every device on
418 the network must be updated via gratuitous ARP,
419 vs. just updating a switch or set of switches (which
420 often takes place for any traffic, not just ARP
421 traffic, if the switch snoops incoming traffic to
422 update its tables) for the traditional method. If the
423 gratuitous ARP is lost, communication may be
426 When this policy is used in conjunction with the mii
427 monitor, devices which assert link up prior to being
428 able to actually transmit and receive are particularly
429 susceptible to loss of the gratuitous ARP, and an
430 appropriate updelay setting may be required.
434 The "follow" fail_over_mac policy causes the MAC
435 address of the bond to be selected normally (normally
436 the MAC address of the first slave added to the bond).
437 However, the second and subsequent slaves are not set
438 to this MAC address while they are in a backup role; a
439 slave is programmed with the bond's MAC address at
440 failover time (and the formerly active slave receives
441 the newly active slave's MAC address).
443 This policy is useful for multiport devices that
444 either become confused or incur a performance penalty
445 when multiple ports are programmed with the same MAC
449 The default policy is none, unless the first slave cannot
450 change its MAC address, in which case the active policy is
453 This option may be modified via sysfs only when no slaves are
456 This option was added in bonding version 3.2.0. The "follow"
457 policy was added in bonding version 3.3.0.
461 Option specifying the rate in which we'll ask our link partner
462 to transmit LACPDU packets in 802.3ad mode. Possible values
466 Request partner to transmit LACPDUs every 30 seconds
469 Request partner to transmit LACPDUs every 1 second
475 Specifies the number of bonding devices to create for this
476 instance of the bonding driver. E.g., if max_bonds is 3, and
477 the bonding driver is not already loaded, then bond0, bond1
478 and bond2 will be created. The default value is 1. Specifying
479 a value of 0 will load bonding, but will not create any devices.
483 Specifies the MII link monitoring frequency in milliseconds.
484 This determines how often the link state of each slave is
485 inspected for link failures. A value of zero disables MII
486 link monitoring. A value of 100 is a good starting point.
487 The use_carrier option, below, affects how the link state is
488 determined. See the High Availability section for additional
489 information. The default value is 0.
493 Specifies the minimum number of links that must be active before
494 asserting carrier. It is similar to the Cisco EtherChannel min-links
495 feature. This allows setting the minimum number of member ports that
496 must be up (link-up state) before marking the bond device as up
497 (carrier on). This is useful for situations where higher level services
498 such as clustering want to ensure a minimum number of low bandwidth
499 links are active before switchover. This option only affect 802.3ad
502 The default value is 0. This will cause carrier to be asserted (for
503 802.3ad mode) whenever there is an active aggregator, regardless of the
504 number of available links in that aggregator. Note that, because an
505 aggregator cannot be active without at least one available link,
506 setting this option to 0 or to 1 has the exact same effect.
510 Specifies one of the bonding policies. The default is
511 balance-rr (round robin). Possible values are:
515 Round-robin policy: Transmit packets in sequential
516 order from the first available slave through the
517 last. This mode provides load balancing and fault
522 Active-backup policy: Only one slave in the bond is
523 active. A different slave becomes active if, and only
524 if, the active slave fails. The bond's MAC address is
525 externally visible on only one port (network adapter)
526 to avoid confusing the switch.
528 In bonding version 2.6.2 or later, when a failover
529 occurs in active-backup mode, bonding will issue one
530 or more gratuitous ARPs on the newly active slave.
531 One gratuitous ARP is issued for the bonding master
532 interface and each VLAN interfaces configured above
533 it, provided that the interface has at least one IP
534 address configured. Gratuitous ARPs issued for VLAN
535 interfaces are tagged with the appropriate VLAN id.
537 This mode provides fault tolerance. The primary
538 option, documented below, affects the behavior of this
543 XOR policy: Transmit based on the selected transmit
544 hash policy. The default policy is a simple [(source
545 MAC address XOR'd with destination MAC address) modulo
546 slave count]. Alternate transmit policies may be
547 selected via the xmit_hash_policy option, described
550 This mode provides load balancing and fault tolerance.
554 Broadcast policy: transmits everything on all slave
555 interfaces. This mode provides fault tolerance.
559 IEEE 802.3ad Dynamic link aggregation. Creates
560 aggregation groups that share the same speed and
561 duplex settings. Utilizes all slaves in the active
562 aggregator according to the 802.3ad specification.
564 Slave selection for outgoing traffic is done according
565 to the transmit hash policy, which may be changed from
566 the default simple XOR policy via the xmit_hash_policy
567 option, documented below. Note that not all transmit
568 policies may be 802.3ad compliant, particularly in
569 regards to the packet mis-ordering requirements of
570 section 43.2.4 of the 802.3ad standard. Differing
571 peer implementations will have varying tolerances for
576 1. Ethtool support in the base drivers for retrieving
577 the speed and duplex of each slave.
579 2. A switch that supports IEEE 802.3ad Dynamic link
582 Most switches will require some type of configuration
583 to enable 802.3ad mode.
587 Adaptive transmit load balancing: channel bonding that
588 does not require any special switch support.
590 In tlb_dynamic_lb=1 mode; the outgoing traffic is
591 distributed according to the current load (computed
592 relative to the speed) on each slave.
594 In tlb_dynamic_lb=0 mode; the load balancing based on
595 current load is disabled and the load is distributed
596 only using the hash distribution.
598 Incoming traffic is received by the current slave.
599 If the receiving slave fails, another slave takes over
600 the MAC address of the failed receiving slave.
604 Ethtool support in the base drivers for retrieving the
609 Adaptive load balancing: includes balance-tlb plus
610 receive load balancing (rlb) for IPV4 traffic, and
611 does not require any special switch support. The
612 receive load balancing is achieved by ARP negotiation.
613 The bonding driver intercepts the ARP Replies sent by
614 the local system on their way out and overwrites the
615 source hardware address with the unique hardware
616 address of one of the slaves in the bond such that
617 different peers use different hardware addresses for
620 Receive traffic from connections created by the server
621 is also balanced. When the local system sends an ARP
622 Request the bonding driver copies and saves the peer's
623 IP information from the ARP packet. When the ARP
624 Reply arrives from the peer, its hardware address is
625 retrieved and the bonding driver initiates an ARP
626 reply to this peer assigning it to one of the slaves
627 in the bond. A problematic outcome of using ARP
628 negotiation for balancing is that each time that an
629 ARP request is broadcast it uses the hardware address
630 of the bond. Hence, peers learn the hardware address
631 of the bond and the balancing of receive traffic
632 collapses to the current slave. This is handled by
633 sending updates (ARP Replies) to all the peers with
634 their individually assigned hardware address such that
635 the traffic is redistributed. Receive traffic is also
636 redistributed when a new slave is added to the bond
637 and when an inactive slave is re-activated. The
638 receive load is distributed sequentially (round robin)
639 among the group of highest speed slaves in the bond.
641 When a link is reconnected or a new slave joins the
642 bond the receive traffic is redistributed among all
643 active slaves in the bond by initiating ARP Replies
644 with the selected MAC address to each of the
645 clients. The updelay parameter (detailed below) must
646 be set to a value equal or greater than the switch's
647 forwarding delay so that the ARP Replies sent to the
648 peers will not be blocked by the switch.
652 1. Ethtool support in the base drivers for retrieving
653 the speed of each slave.
655 2. Base driver support for setting the hardware
656 address of a device while it is open. This is
657 required so that there will always be one slave in the
658 team using the bond hardware address (the
659 curr_active_slave) while having a unique hardware
660 address for each slave in the bond. If the
661 curr_active_slave fails its hardware address is
662 swapped with the new curr_active_slave that was
668 Specify the number of peer notifications (gratuitous ARPs and
669 unsolicited IPv6 Neighbor Advertisements) to be issued after a
670 failover event. As soon as the link is up on the new slave
671 (possibly immediately) a peer notification is sent on the
672 bonding device and each VLAN sub-device. This is repeated at
673 each link monitor interval (arp_interval or miimon, whichever
674 is active) if the number is greater than 1.
676 The valid range is 0 - 255; the default value is 1. These options
677 affect only the active-backup mode. These options were added for
678 bonding versions 3.3.0 and 3.4.0 respectively.
680 From Linux 3.0 and bonding version 3.7.1, these notifications
681 are generated by the ipv4 and ipv6 code and the numbers of
682 repetitions cannot be set independently.
686 Specify the number of packets to transmit through a slave before
687 moving to the next one. When set to 0 then a slave is chosen at
690 The valid range is 0 - 65535; the default value is 1. This option
691 has effect only in balance-rr mode.
695 A string (eth0, eth2, etc) specifying which slave is the
696 primary device. The specified device will always be the
697 active slave while it is available. Only when the primary is
698 off-line will alternate devices be used. This is useful when
699 one slave is preferred over another, e.g., when one slave has
700 higher throughput than another.
702 The primary option is only valid for active-backup(1),
703 balance-tlb (5) and balance-alb (6) mode.
707 Specifies the reselection policy for the primary slave. This
708 affects how the primary slave is chosen to become the active slave
709 when failure of the active slave or recovery of the primary slave
710 occurs. This option is designed to prevent flip-flopping between
711 the primary slave and other slaves. Possible values are:
713 always or 0 (default)
715 The primary slave becomes the active slave whenever it
720 The primary slave becomes the active slave when it comes
721 back up, if the speed and duplex of the primary slave is
722 better than the speed and duplex of the current active
727 The primary slave becomes the active slave only if the
728 current active slave fails and the primary slave is up.
730 The primary_reselect setting is ignored in two cases:
732 If no slaves are active, the first slave to recover is
733 made the active slave.
735 When initially enslaved, the primary slave is always made
738 Changing the primary_reselect policy via sysfs will cause an
739 immediate selection of the best active slave according to the new
740 policy. This may or may not result in a change of the active
741 slave, depending upon the circumstances.
743 This option was added for bonding version 3.6.0.
747 Specifies if dynamic shuffling of flows is enabled in tlb
748 mode. The value has no effect on any other modes.
750 The default behavior of tlb mode is to shuffle active flows across
751 slaves based on the load in that interval. This gives nice lb
752 characteristics but can cause packet reordering. If re-ordering is
753 a concern use this variable to disable flow shuffling and rely on
754 load balancing provided solely by the hash distribution.
755 xmit-hash-policy can be used to select the appropriate hashing for
758 The sysfs entry can be used to change the setting per bond device
759 and the initial value is derived from the module parameter. The
760 sysfs entry is allowed to be changed only if the bond device is
763 The default value is "1" that enables flow shuffling while value "0"
764 disables it. This option was added in bonding driver 3.7.1
769 Specifies the time, in milliseconds, to wait before enabling a
770 slave after a link recovery has been detected. This option is
771 only valid for the miimon link monitor. The updelay value
772 should be a multiple of the miimon value; if not, it will be
773 rounded down to the nearest multiple. The default value is 0.
777 Specifies whether or not miimon should use MII or ETHTOOL
778 ioctls vs. netif_carrier_ok() to determine the link
779 status. The MII or ETHTOOL ioctls are less efficient and
780 utilize a deprecated calling sequence within the kernel. The
781 netif_carrier_ok() relies on the device driver to maintain its
782 state with netif_carrier_on/off; at this writing, most, but
783 not all, device drivers support this facility.
785 If bonding insists that the link is up when it should not be,
786 it may be that your network device driver does not support
787 netif_carrier_on/off. The default state for netif_carrier is
788 "carrier on," so if a driver does not support netif_carrier,
789 it will appear as if the link is always up. In this case,
790 setting use_carrier to 0 will cause bonding to revert to the
791 MII / ETHTOOL ioctl method to determine the link state.
793 A value of 1 enables the use of netif_carrier_ok(), a value of
794 0 will use the deprecated MII / ETHTOOL ioctls. The default
799 Selects the transmit hash policy to use for slave selection in
800 balance-xor, 802.3ad, and tlb modes. Possible values are:
804 Uses XOR of hardware MAC addresses to generate the
807 (source MAC XOR destination MAC) modulo slave count
809 This algorithm will place all traffic to a particular
810 network peer on the same slave.
812 This algorithm is 802.3ad compliant.
816 This policy uses a combination of layer2 and layer3
817 protocol information to generate the hash.
819 Uses XOR of hardware MAC addresses and IP addresses to
820 generate the hash. The formula is
822 hash = source MAC XOR destination MAC
823 hash = hash XOR source IP XOR destination IP
824 hash = hash XOR (hash RSHIFT 16)
825 hash = hash XOR (hash RSHIFT 8)
826 And then hash is reduced modulo slave count.
828 If the protocol is IPv6 then the source and destination
829 addresses are first hashed using ipv6_addr_hash.
831 This algorithm will place all traffic to a particular
832 network peer on the same slave. For non-IP traffic,
833 the formula is the same as for the layer2 transmit
836 This policy is intended to provide a more balanced
837 distribution of traffic than layer2 alone, especially
838 in environments where a layer3 gateway device is
839 required to reach most destinations.
841 This algorithm is 802.3ad compliant.
845 This policy uses upper layer protocol information,
846 when available, to generate the hash. This allows for
847 traffic to a particular network peer to span multiple
848 slaves, although a single connection will not span
851 The formula for unfragmented TCP and UDP packets is
853 hash = source port, destination port (as in the header)
854 hash = hash XOR source IP XOR destination IP
855 hash = hash XOR (hash RSHIFT 16)
856 hash = hash XOR (hash RSHIFT 8)
857 And then hash is reduced modulo slave count.
859 If the protocol is IPv6 then the source and destination
860 addresses are first hashed using ipv6_addr_hash.
862 For fragmented TCP or UDP packets and all other IPv4 and
863 IPv6 protocol traffic, the source and destination port
864 information is omitted. For non-IP traffic, the
865 formula is the same as for the layer2 transmit hash
868 This algorithm is not fully 802.3ad compliant. A
869 single TCP or UDP conversation containing both
870 fragmented and unfragmented packets will see packets
871 striped across two interfaces. This may result in out
872 of order delivery. Most traffic types will not meet
873 this criteria, as TCP rarely fragments traffic, and
874 most UDP traffic is not involved in extended
875 conversations. Other implementations of 802.3ad may
876 or may not tolerate this noncompliance.
880 This policy uses the same formula as layer2+3 but it
881 relies on skb_flow_dissect to obtain the header fields
882 which might result in the use of inner headers if an
883 encapsulation protocol is used. For example this will
884 improve the performance for tunnel users because the
885 packets will be distributed according to the encapsulated
890 This policy uses the same formula as layer3+4 but it
891 relies on skb_flow_dissect to obtain the header fields
892 which might result in the use of inner headers if an
893 encapsulation protocol is used. For example this will
894 improve the performance for tunnel users because the
895 packets will be distributed according to the encapsulated
898 The default value is layer2. This option was added in bonding
899 version 2.6.3. In earlier versions of bonding, this parameter
900 does not exist, and the layer2 policy is the only policy. The
901 layer2+3 value was added for bonding version 3.2.2.
905 Specifies the number of IGMP membership reports to be issued after
906 a failover event. One membership report is issued immediately after
907 the failover, subsequent packets are sent in each 200ms interval.
909 The valid range is 0 - 255; the default value is 1. A value of 0
910 prevents the IGMP membership report from being issued in response
911 to the failover event.
913 This option is useful for bonding modes balance-rr (0), active-backup
914 (1), balance-tlb (5) and balance-alb (6), in which a failover can
915 switch the IGMP traffic from one slave to another. Therefore a fresh
916 IGMP report must be issued to cause the switch to forward the incoming
917 IGMP traffic over the newly selected slave.
919 This option was added for bonding version 3.7.0.
923 Specifies the number of seconds between instances where the bonding
924 driver sends learning packets to each slaves peer switch.
926 The valid range is 1 - 0x7fffffff; the default value is 1. This Option
927 has effect only in balance-tlb and balance-alb modes.
929 3. Configuring Bonding Devices
930 ==============================
932 You can configure bonding using either your distro's network
933 initialization scripts, or manually using either iproute2 or the
934 sysfs interface. Distros generally use one of three packages for the
935 network initialization scripts: initscripts, sysconfig or interfaces.
936 Recent versions of these packages have support for bonding, while older
939 We will first describe the options for configuring bonding for
940 distros using versions of initscripts, sysconfig and interfaces with full
941 or partial support for bonding, then provide information on enabling
942 bonding without support from the network initialization scripts (i.e.,
943 older versions of initscripts or sysconfig).
945 If you're unsure whether your distro uses sysconfig,
946 initscripts or interfaces, or don't know if it's new enough, have no fear.
947 Determining this is fairly straightforward.
949 First, look for a file called interfaces in /etc/network directory.
950 If this file is present in your system, then your system use interfaces. See
951 Configuration with Interfaces Support.
953 Else, issue the command:
957 It will respond with a line of text starting with either
958 "initscripts" or "sysconfig," followed by some numbers. This is the
959 package that provides your network initialization scripts.
961 Next, to determine if your installation supports bonding,
964 $ grep ifenslave /sbin/ifup
966 If this returns any matches, then your initscripts or
967 sysconfig has support for bonding.
969 3.1 Configuration with Sysconfig Support
970 ----------------------------------------
972 This section applies to distros using a version of sysconfig
973 with bonding support, for example, SuSE Linux Enterprise Server 9.
975 SuSE SLES 9's networking configuration system does support
976 bonding, however, at this writing, the YaST system configuration
977 front end does not provide any means to work with bonding devices.
978 Bonding devices can be managed by hand, however, as follows.
980 First, if they have not already been configured, configure the
981 slave devices. On SLES 9, this is most easily done by running the
982 yast2 sysconfig configuration utility. The goal is for to create an
983 ifcfg-id file for each slave device. The simplest way to accomplish
984 this is to configure the devices for DHCP (this is only to get the
985 file ifcfg-id file created; see below for some issues with DHCP). The
986 name of the configuration file for each device will be of the form:
988 ifcfg-id-xx:xx:xx:xx:xx:xx
990 Where the "xx" portion will be replaced with the digits from
991 the device's permanent MAC address.
993 Once the set of ifcfg-id-xx:xx:xx:xx:xx:xx files has been
994 created, it is necessary to edit the configuration files for the slave
995 devices (the MAC addresses correspond to those of the slave devices).
996 Before editing, the file will contain multiple lines, and will look
1002 UNIQUE='XNzu.WeZGOGF+4wE'
1003 _nm_name='bus-pci-0001:61:01.0'
1005 Change the BOOTPROTO and STARTMODE lines to the following:
1010 Do not alter the UNIQUE or _nm_name lines. Remove any other
1011 lines (USERCTL, etc).
1013 Once the ifcfg-id-xx:xx:xx:xx:xx:xx files have been modified,
1014 it's time to create the configuration file for the bonding device
1015 itself. This file is named ifcfg-bondX, where X is the number of the
1016 bonding device to create, starting at 0. The first such file is
1017 ifcfg-bond0, the second is ifcfg-bond1, and so on. The sysconfig
1018 network configuration system will correctly start multiple instances
1021 The contents of the ifcfg-bondX file is as follows:
1024 BROADCAST="10.0.2.255"
1026 NETMASK="255.255.0.0"
1030 BONDING_MASTER="yes"
1031 BONDING_MODULE_OPTS="mode=active-backup miimon=100"
1032 BONDING_SLAVE0="eth0"
1033 BONDING_SLAVE1="bus-pci-0000:06:08.1"
1035 Replace the sample BROADCAST, IPADDR, NETMASK and NETWORK
1036 values with the appropriate values for your network.
1038 The STARTMODE specifies when the device is brought online.
1039 The possible values are:
1041 onboot: The device is started at boot time. If you're not
1042 sure, this is probably what you want.
1044 manual: The device is started only when ifup is called
1045 manually. Bonding devices may be configured this
1046 way if you do not wish them to start automatically
1047 at boot for some reason.
1049 hotplug: The device is started by a hotplug event. This is not
1050 a valid choice for a bonding device.
1052 off or ignore: The device configuration is ignored.
1054 The line BONDING_MASTER='yes' indicates that the device is a
1055 bonding master device. The only useful value is "yes."
1057 The contents of BONDING_MODULE_OPTS are supplied to the
1058 instance of the bonding module for this device. Specify the options
1059 for the bonding mode, link monitoring, and so on here. Do not include
1060 the max_bonds bonding parameter; this will confuse the configuration
1061 system if you have multiple bonding devices.
1063 Finally, supply one BONDING_SLAVEn="slave device" for each
1064 slave. where "n" is an increasing value, one for each slave. The
1065 "slave device" is either an interface name, e.g., "eth0", or a device
1066 specifier for the network device. The interface name is easier to
1067 find, but the ethN names are subject to change at boot time if, e.g.,
1068 a device early in the sequence has failed. The device specifiers
1069 (bus-pci-0000:06:08.1 in the example above) specify the physical
1070 network device, and will not change unless the device's bus location
1071 changes (for example, it is moved from one PCI slot to another). The
1072 example above uses one of each type for demonstration purposes; most
1073 configurations will choose one or the other for all slave devices.
1075 When all configuration files have been modified or created,
1076 networking must be restarted for the configuration changes to take
1077 effect. This can be accomplished via the following:
1079 # /etc/init.d/network restart
1081 Note that the network control script (/sbin/ifdown) will
1082 remove the bonding module as part of the network shutdown processing,
1083 so it is not necessary to remove the module by hand if, e.g., the
1084 module parameters have changed.
1086 Also, at this writing, YaST/YaST2 will not manage bonding
1087 devices (they do not show bonding interfaces on its list of network
1088 devices). It is necessary to edit the configuration file by hand to
1089 change the bonding configuration.
1091 Additional general options and details of the ifcfg file
1092 format can be found in an example ifcfg template file:
1094 /etc/sysconfig/network/ifcfg.template
1096 Note that the template does not document the various BONDING_
1097 settings described above, but does describe many of the other options.
1099 3.1.1 Using DHCP with Sysconfig
1100 -------------------------------
1102 Under sysconfig, configuring a device with BOOTPROTO='dhcp'
1103 will cause it to query DHCP for its IP address information. At this
1104 writing, this does not function for bonding devices; the scripts
1105 attempt to obtain the device address from DHCP prior to adding any of
1106 the slave devices. Without active slaves, the DHCP requests are not
1107 sent to the network.
1109 3.1.2 Configuring Multiple Bonds with Sysconfig
1110 -----------------------------------------------
1112 The sysconfig network initialization system is capable of
1113 handling multiple bonding devices. All that is necessary is for each
1114 bonding instance to have an appropriately configured ifcfg-bondX file
1115 (as described above). Do not specify the "max_bonds" parameter to any
1116 instance of bonding, as this will confuse sysconfig. If you require
1117 multiple bonding devices with identical parameters, create multiple
1120 Because the sysconfig scripts supply the bonding module
1121 options in the ifcfg-bondX file, it is not necessary to add them to
1122 the system /etc/modules.d/*.conf configuration files.
1124 3.2 Configuration with Initscripts Support
1125 ------------------------------------------
1127 This section applies to distros using a recent version of
1128 initscripts with bonding support, for example, Red Hat Enterprise Linux
1129 version 3 or later, Fedora, etc. On these systems, the network
1130 initialization scripts have knowledge of bonding, and can be configured to
1131 control bonding devices. Note that older versions of the initscripts
1132 package have lower levels of support for bonding; this will be noted where
1135 These distros will not automatically load the network adapter
1136 driver unless the ethX device is configured with an IP address.
1137 Because of this constraint, users must manually configure a
1138 network-script file for all physical adapters that will be members of
1139 a bondX link. Network script files are located in the directory:
1141 /etc/sysconfig/network-scripts
1143 The file name must be prefixed with "ifcfg-eth" and suffixed
1144 with the adapter's physical adapter number. For example, the script
1145 for eth0 would be named /etc/sysconfig/network-scripts/ifcfg-eth0.
1146 Place the following text in the file:
1155 The DEVICE= line will be different for every ethX device and
1156 must correspond with the name of the file, i.e., ifcfg-eth1 must have
1157 a device line of DEVICE=eth1. The setting of the MASTER= line will
1158 also depend on the final bonding interface name chosen for your bond.
1159 As with other network devices, these typically start at 0, and go up
1160 one for each device, i.e., the first bonding instance is bond0, the
1161 second is bond1, and so on.
1163 Next, create a bond network script. The file name for this
1164 script will be /etc/sysconfig/network-scripts/ifcfg-bondX where X is
1165 the number of the bond. For bond0 the file is named "ifcfg-bond0",
1166 for bond1 it is named "ifcfg-bond1", and so on. Within that file,
1167 place the following text:
1171 NETMASK=255.255.255.0
1173 BROADCAST=192.168.1.255
1178 Be sure to change the networking specific lines (IPADDR,
1179 NETMASK, NETWORK and BROADCAST) to match your network configuration.
1181 For later versions of initscripts, such as that found with Fedora
1182 7 (or later) and Red Hat Enterprise Linux version 5 (or later), it is possible,
1183 and, indeed, preferable, to specify the bonding options in the ifcfg-bond0
1184 file, e.g. a line of the format:
1186 BONDING_OPTS="mode=active-backup arp_interval=60 arp_ip_target=192.168.1.254"
1188 will configure the bond with the specified options. The options
1189 specified in BONDING_OPTS are identical to the bonding module parameters
1190 except for the arp_ip_target field when using versions of initscripts older
1191 than and 8.57 (Fedora 8) and 8.45.19 (Red Hat Enterprise Linux 5.2). When
1192 using older versions each target should be included as a separate option and
1193 should be preceded by a '+' to indicate it should be added to the list of
1194 queried targets, e.g.,
1196 arp_ip_target=+192.168.1.1 arp_ip_target=+192.168.1.2
1198 is the proper syntax to specify multiple targets. When specifying
1199 options via BONDING_OPTS, it is not necessary to edit /etc/modprobe.d/*.conf.
1201 For even older versions of initscripts that do not support
1202 BONDING_OPTS, it is necessary to edit /etc/modprobe.d/*.conf, depending upon
1203 your distro) to load the bonding module with your desired options when the
1204 bond0 interface is brought up. The following lines in /etc/modprobe.d/*.conf
1205 will load the bonding module, and select its options:
1208 options bond0 mode=balance-alb miimon=100
1210 Replace the sample parameters with the appropriate set of
1211 options for your configuration.
1213 Finally run "/etc/rc.d/init.d/network restart" as root. This
1214 will restart the networking subsystem and your bond link should be now
1217 3.2.1 Using DHCP with Initscripts
1218 ---------------------------------
1220 Recent versions of initscripts (the versions supplied with Fedora
1221 Core 3 and Red Hat Enterprise Linux 4, or later versions, are reported to
1222 work) have support for assigning IP information to bonding devices via
1225 To configure bonding for DHCP, configure it as described
1226 above, except replace the line "BOOTPROTO=none" with "BOOTPROTO=dhcp"
1227 and add a line consisting of "TYPE=Bonding". Note that the TYPE value
1230 3.2.2 Configuring Multiple Bonds with Initscripts
1231 -------------------------------------------------
1233 Initscripts packages that are included with Fedora 7 and Red Hat
1234 Enterprise Linux 5 support multiple bonding interfaces by simply
1235 specifying the appropriate BONDING_OPTS= in ifcfg-bondX where X is the
1236 number of the bond. This support requires sysfs support in the kernel,
1237 and a bonding driver of version 3.0.0 or later. Other configurations may
1238 not support this method for specifying multiple bonding interfaces; for
1239 those instances, see the "Configuring Multiple Bonds Manually" section,
1242 3.3 Configuring Bonding Manually with iproute2
1243 -----------------------------------------------
1245 This section applies to distros whose network initialization
1246 scripts (the sysconfig or initscripts package) do not have specific
1247 knowledge of bonding. One such distro is SuSE Linux Enterprise Server
1250 The general method for these systems is to place the bonding
1251 module parameters into a config file in /etc/modprobe.d/ (as
1252 appropriate for the installed distro), then add modprobe and/or
1253 `ip link` commands to the system's global init script. The name of
1254 the global init script differs; for sysconfig, it is
1255 /etc/init.d/boot.local and for initscripts it is /etc/rc.d/rc.local.
1257 For example, if you wanted to make a simple bond of two e100
1258 devices (presumed to be eth0 and eth1), and have it persist across
1259 reboots, edit the appropriate file (/etc/init.d/boot.local or
1260 /etc/rc.d/rc.local), and add the following:
1262 modprobe bonding mode=balance-alb miimon=100
1264 ifconfig bond0 192.168.1.1 netmask 255.255.255.0 up
1265 ip link set eth0 master bond0
1266 ip link set eth1 master bond0
1268 Replace the example bonding module parameters and bond0
1269 network configuration (IP address, netmask, etc) with the appropriate
1270 values for your configuration.
1272 Unfortunately, this method will not provide support for the
1273 ifup and ifdown scripts on the bond devices. To reload the bonding
1274 configuration, it is necessary to run the initialization script, e.g.,
1276 # /etc/init.d/boot.local
1280 # /etc/rc.d/rc.local
1282 It may be desirable in such a case to create a separate script
1283 which only initializes the bonding configuration, then call that
1284 separate script from within boot.local. This allows for bonding to be
1285 enabled without re-running the entire global init script.
1287 To shut down the bonding devices, it is necessary to first
1288 mark the bonding device itself as being down, then remove the
1289 appropriate device driver modules. For our example above, you can do
1292 # ifconfig bond0 down
1296 Again, for convenience, it may be desirable to create a script
1297 with these commands.
1300 3.3.1 Configuring Multiple Bonds Manually
1301 -----------------------------------------
1303 This section contains information on configuring multiple
1304 bonding devices with differing options for those systems whose network
1305 initialization scripts lack support for configuring multiple bonds.
1307 If you require multiple bonding devices, but all with the same
1308 options, you may wish to use the "max_bonds" module parameter,
1311 To create multiple bonding devices with differing options, it is
1312 preferable to use bonding parameters exported by sysfs, documented in the
1315 For versions of bonding without sysfs support, the only means to
1316 provide multiple instances of bonding with differing options is to load
1317 the bonding driver multiple times. Note that current versions of the
1318 sysconfig network initialization scripts handle this automatically; if
1319 your distro uses these scripts, no special action is needed. See the
1320 section Configuring Bonding Devices, above, if you're not sure about your
1321 network initialization scripts.
1323 To load multiple instances of the module, it is necessary to
1324 specify a different name for each instance (the module loading system
1325 requires that every loaded module, even multiple instances of the same
1326 module, have a unique name). This is accomplished by supplying multiple
1327 sets of bonding options in /etc/modprobe.d/*.conf, for example:
1330 options bond0 -o bond0 mode=balance-rr miimon=100
1333 options bond1 -o bond1 mode=balance-alb miimon=50
1335 will load the bonding module two times. The first instance is
1336 named "bond0" and creates the bond0 device in balance-rr mode with an
1337 miimon of 100. The second instance is named "bond1" and creates the
1338 bond1 device in balance-alb mode with an miimon of 50.
1340 In some circumstances (typically with older distributions),
1341 the above does not work, and the second bonding instance never sees
1342 its options. In that case, the second options line can be substituted
1345 install bond1 /sbin/modprobe --ignore-install bonding -o bond1 \
1346 mode=balance-alb miimon=50
1348 This may be repeated any number of times, specifying a new and
1349 unique name in place of bond1 for each subsequent instance.
1351 It has been observed that some Red Hat supplied kernels are unable
1352 to rename modules at load time (the "-o bond1" part). Attempts to pass
1353 that option to modprobe will produce an "Operation not permitted" error.
1354 This has been reported on some Fedora Core kernels, and has been seen on
1355 RHEL 4 as well. On kernels exhibiting this problem, it will be impossible
1356 to configure multiple bonds with differing parameters (as they are older
1357 kernels, and also lack sysfs support).
1359 3.4 Configuring Bonding Manually via Sysfs
1360 ------------------------------------------
1362 Starting with version 3.0.0, Channel Bonding may be configured
1363 via the sysfs interface. This interface allows dynamic configuration
1364 of all bonds in the system without unloading the module. It also
1365 allows for adding and removing bonds at runtime. Ifenslave is no
1366 longer required, though it is still supported.
1368 Use of the sysfs interface allows you to use multiple bonds
1369 with different configurations without having to reload the module.
1370 It also allows you to use multiple, differently configured bonds when
1371 bonding is compiled into the kernel.
1373 You must have the sysfs filesystem mounted to configure
1374 bonding this way. The examples in this document assume that you
1375 are using the standard mount point for sysfs, e.g. /sys. If your
1376 sysfs filesystem is mounted elsewhere, you will need to adjust the
1377 example paths accordingly.
1379 Creating and Destroying Bonds
1380 -----------------------------
1381 To add a new bond foo:
1382 # echo +foo > /sys/class/net/bonding_masters
1384 To remove an existing bond bar:
1385 # echo -bar > /sys/class/net/bonding_masters
1387 To show all existing bonds:
1388 # cat /sys/class/net/bonding_masters
1390 NOTE: due to 4K size limitation of sysfs files, this list may be
1391 truncated if you have more than a few hundred bonds. This is unlikely
1392 to occur under normal operating conditions.
1394 Adding and Removing Slaves
1395 --------------------------
1396 Interfaces may be enslaved to a bond using the file
1397 /sys/class/net/<bond>/bonding/slaves. The semantics for this file
1398 are the same as for the bonding_masters file.
1400 To enslave interface eth0 to bond bond0:
1402 # echo +eth0 > /sys/class/net/bond0/bonding/slaves
1404 To free slave eth0 from bond bond0:
1405 # echo -eth0 > /sys/class/net/bond0/bonding/slaves
1407 When an interface is enslaved to a bond, symlinks between the
1408 two are created in the sysfs filesystem. In this case, you would get
1409 /sys/class/net/bond0/slave_eth0 pointing to /sys/class/net/eth0, and
1410 /sys/class/net/eth0/master pointing to /sys/class/net/bond0.
1412 This means that you can tell quickly whether or not an
1413 interface is enslaved by looking for the master symlink. Thus:
1414 # echo -eth0 > /sys/class/net/eth0/master/bonding/slaves
1415 will free eth0 from whatever bond it is enslaved to, regardless of
1416 the name of the bond interface.
1418 Changing a Bond's Configuration
1419 -------------------------------
1420 Each bond may be configured individually by manipulating the
1421 files located in /sys/class/net/<bond name>/bonding
1423 The names of these files correspond directly with the command-
1424 line parameters described elsewhere in this file, and, with the
1425 exception of arp_ip_target, they accept the same values. To see the
1426 current setting, simply cat the appropriate file.
1428 A few examples will be given here; for specific usage
1429 guidelines for each parameter, see the appropriate section in this
1432 To configure bond0 for balance-alb mode:
1433 # ifconfig bond0 down
1434 # echo 6 > /sys/class/net/bond0/bonding/mode
1436 # echo balance-alb > /sys/class/net/bond0/bonding/mode
1437 NOTE: The bond interface must be down before the mode can be
1440 To enable MII monitoring on bond0 with a 1 second interval:
1441 # echo 1000 > /sys/class/net/bond0/bonding/miimon
1442 NOTE: If ARP monitoring is enabled, it will disabled when MII
1443 monitoring is enabled, and vice-versa.
1446 # echo +192.168.0.100 > /sys/class/net/bond0/bonding/arp_ip_target
1447 # echo +192.168.0.101 > /sys/class/net/bond0/bonding/arp_ip_target
1448 NOTE: up to 16 target addresses may be specified.
1450 To remove an ARP target:
1451 # echo -192.168.0.100 > /sys/class/net/bond0/bonding/arp_ip_target
1453 To configure the interval between learning packet transmits:
1454 # echo 12 > /sys/class/net/bond0/bonding/lp_interval
1455 NOTE: the lp_inteval is the number of seconds between instances where
1456 the bonding driver sends learning packets to each slaves peer switch. The
1457 default interval is 1 second.
1459 Example Configuration
1460 ---------------------
1461 We begin with the same example that is shown in section 3.3,
1462 executed with sysfs, and without using ifenslave.
1464 To make a simple bond of two e100 devices (presumed to be eth0
1465 and eth1), and have it persist across reboots, edit the appropriate
1466 file (/etc/init.d/boot.local or /etc/rc.d/rc.local), and add the
1471 echo balance-alb > /sys/class/net/bond0/bonding/mode
1472 ifconfig bond0 192.168.1.1 netmask 255.255.255.0 up
1473 echo 100 > /sys/class/net/bond0/bonding/miimon
1474 echo +eth0 > /sys/class/net/bond0/bonding/slaves
1475 echo +eth1 > /sys/class/net/bond0/bonding/slaves
1477 To add a second bond, with two e1000 interfaces in
1478 active-backup mode, using ARP monitoring, add the following lines to
1482 echo +bond1 > /sys/class/net/bonding_masters
1483 echo active-backup > /sys/class/net/bond1/bonding/mode
1484 ifconfig bond1 192.168.2.1 netmask 255.255.255.0 up
1485 echo +192.168.2.100 /sys/class/net/bond1/bonding/arp_ip_target
1486 echo 2000 > /sys/class/net/bond1/bonding/arp_interval
1487 echo +eth2 > /sys/class/net/bond1/bonding/slaves
1488 echo +eth3 > /sys/class/net/bond1/bonding/slaves
1490 3.5 Configuration with Interfaces Support
1491 -----------------------------------------
1493 This section applies to distros which use /etc/network/interfaces file
1494 to describe network interface configuration, most notably Debian and it's
1497 The ifup and ifdown commands on Debian don't support bonding out of
1498 the box. The ifenslave-2.6 package should be installed to provide bonding
1499 support. Once installed, this package will provide bond-* options to be used
1500 into /etc/network/interfaces.
1502 Note that ifenslave-2.6 package will load the bonding module and use
1503 the ifenslave command when appropriate.
1505 Example Configurations
1506 ----------------------
1508 In /etc/network/interfaces, the following stanza will configure bond0, in
1509 active-backup mode, with eth0 and eth1 as slaves.
1512 iface bond0 inet dhcp
1513 bond-slaves eth0 eth1
1514 bond-mode active-backup
1516 bond-primary eth0 eth1
1518 If the above configuration doesn't work, you might have a system using
1519 upstart for system startup. This is most notably true for recent
1520 Ubuntu versions. The following stanza in /etc/network/interfaces will
1521 produce the same result on those systems.
1524 iface bond0 inet dhcp
1526 bond-mode active-backup
1530 iface eth0 inet manual
1532 bond-primary eth0 eth1
1535 iface eth1 inet manual
1537 bond-primary eth0 eth1
1539 For a full list of bond-* supported options in /etc/network/interfaces and some
1540 more advanced examples tailored to you particular distros, see the files in
1541 /usr/share/doc/ifenslave-2.6.
1543 3.6 Overriding Configuration for Special Cases
1544 ----------------------------------------------
1546 When using the bonding driver, the physical port which transmits a frame is
1547 typically selected by the bonding driver, and is not relevant to the user or
1548 system administrator. The output port is simply selected using the policies of
1549 the selected bonding mode. On occasion however, it is helpful to direct certain
1550 classes of traffic to certain physical interfaces on output to implement
1551 slightly more complex policies. For example, to reach a web server over a
1552 bonded interface in which eth0 connects to a private network, while eth1
1553 connects via a public network, it may be desirous to bias the bond to send said
1554 traffic over eth0 first, using eth1 only as a fall back, while all other traffic
1555 can safely be sent over either interface. Such configurations may be achieved
1556 using the traffic control utilities inherent in linux.
1558 By default the bonding driver is multiqueue aware and 16 queues are created
1559 when the driver initializes (see Documentation/networking/multiqueue.txt
1560 for details). If more or less queues are desired the module parameter
1561 tx_queues can be used to change this value. There is no sysfs parameter
1562 available as the allocation is done at module init time.
1564 The output of the file /proc/net/bonding/bondX has changed so the output Queue
1565 ID is now printed for each slave:
1567 Bonding Mode: fault-tolerance (active-backup)
1569 Currently Active Slave: eth0
1571 MII Polling Interval (ms): 0
1575 Slave Interface: eth0
1577 Link Failure Count: 0
1578 Permanent HW addr: 00:1a:a0:12:8f:cb
1581 Slave Interface: eth1
1583 Link Failure Count: 0
1584 Permanent HW addr: 00:1a:a0:12:8f:cc
1587 The queue_id for a slave can be set using the command:
1589 # echo "eth1:2" > /sys/class/net/bond0/bonding/queue_id
1591 Any interface that needs a queue_id set should set it with multiple calls
1592 like the one above until proper priorities are set for all interfaces. On
1593 distributions that allow configuration via initscripts, multiple 'queue_id'
1594 arguments can be added to BONDING_OPTS to set all needed slave queues.
1596 These queue id's can be used in conjunction with the tc utility to configure
1597 a multiqueue qdisc and filters to bias certain traffic to transmit on certain
1598 slave devices. For instance, say we wanted, in the above configuration to
1599 force all traffic bound to 192.168.1.100 to use eth1 in the bond as its output
1600 device. The following commands would accomplish this:
1602 # tc qdisc add dev bond0 handle 1 root multiq
1604 # tc filter add dev bond0 protocol ip parent 1: prio 1 u32 match ip dst \
1605 192.168.1.100 action skbedit queue_mapping 2
1607 These commands tell the kernel to attach a multiqueue queue discipline to the
1608 bond0 interface and filter traffic enqueued to it, such that packets with a dst
1609 ip of 192.168.1.100 have their output queue mapping value overwritten to 2.
1610 This value is then passed into the driver, causing the normal output path
1611 selection policy to be overridden, selecting instead qid 2, which maps to eth1.
1613 Note that qid values begin at 1. Qid 0 is reserved to initiate to the driver
1614 that normal output policy selection should take place. One benefit to simply
1615 leaving the qid for a slave to 0 is the multiqueue awareness in the bonding
1616 driver that is now present. This awareness allows tc filters to be placed on
1617 slave devices as well as bond devices and the bonding driver will simply act as
1618 a pass-through for selecting output queues on the slave device rather than
1619 output port selection.
1621 This feature first appeared in bonding driver version 3.7.0 and support for
1622 output slave selection was limited to round-robin and active-backup modes.
1624 4 Querying Bonding Configuration
1625 =================================
1627 4.1 Bonding Configuration
1628 -------------------------
1630 Each bonding device has a read-only file residing in the
1631 /proc/net/bonding directory. The file contents include information
1632 about the bonding configuration, options and state of each slave.
1634 For example, the contents of /proc/net/bonding/bond0 after the
1635 driver is loaded with parameters of mode=0 and miimon=1000 is
1636 generally as follows:
1638 Ethernet Channel Bonding Driver: 2.6.1 (October 29, 2004)
1639 Bonding Mode: load balancing (round-robin)
1640 Currently Active Slave: eth0
1642 MII Polling Interval (ms): 1000
1646 Slave Interface: eth1
1648 Link Failure Count: 1
1650 Slave Interface: eth0
1652 Link Failure Count: 1
1654 The precise format and contents will change depending upon the
1655 bonding configuration, state, and version of the bonding driver.
1657 4.2 Network configuration
1658 -------------------------
1660 The network configuration can be inspected using the ifconfig
1661 command. Bonding devices will have the MASTER flag set; Bonding slave
1662 devices will have the SLAVE flag set. The ifconfig output does not
1663 contain information on which slaves are associated with which masters.
1665 In the example below, the bond0 interface is the master
1666 (MASTER) while eth0 and eth1 are slaves (SLAVE). Notice all slaves of
1667 bond0 have the same MAC address (HWaddr) as bond0 for all modes except
1668 TLB and ALB that require a unique MAC address for each slave.
1671 bond0 Link encap:Ethernet HWaddr 00:C0:F0:1F:37:B4
1672 inet addr:XXX.XXX.XXX.YYY Bcast:XXX.XXX.XXX.255 Mask:255.255.252.0
1673 UP BROADCAST RUNNING MASTER MULTICAST MTU:1500 Metric:1
1674 RX packets:7224794 errors:0 dropped:0 overruns:0 frame:0
1675 TX packets:3286647 errors:1 dropped:0 overruns:1 carrier:0
1676 collisions:0 txqueuelen:0
1678 eth0 Link encap:Ethernet HWaddr 00:C0:F0:1F:37:B4
1679 UP BROADCAST RUNNING SLAVE MULTICAST MTU:1500 Metric:1
1680 RX packets:3573025 errors:0 dropped:0 overruns:0 frame:0
1681 TX packets:1643167 errors:1 dropped:0 overruns:1 carrier:0
1682 collisions:0 txqueuelen:100
1683 Interrupt:10 Base address:0x1080
1685 eth1 Link encap:Ethernet HWaddr 00:C0:F0:1F:37:B4
1686 UP BROADCAST RUNNING SLAVE MULTICAST MTU:1500 Metric:1
1687 RX packets:3651769 errors:0 dropped:0 overruns:0 frame:0
1688 TX packets:1643480 errors:0 dropped:0 overruns:0 carrier:0
1689 collisions:0 txqueuelen:100
1690 Interrupt:9 Base address:0x1400
1692 5. Switch Configuration
1693 =======================
1695 For this section, "switch" refers to whatever system the
1696 bonded devices are directly connected to (i.e., where the other end of
1697 the cable plugs into). This may be an actual dedicated switch device,
1698 or it may be another regular system (e.g., another computer running
1701 The active-backup, balance-tlb and balance-alb modes do not
1702 require any specific configuration of the switch.
1704 The 802.3ad mode requires that the switch have the appropriate
1705 ports configured as an 802.3ad aggregation. The precise method used
1706 to configure this varies from switch to switch, but, for example, a
1707 Cisco 3550 series switch requires that the appropriate ports first be
1708 grouped together in a single etherchannel instance, then that
1709 etherchannel is set to mode "lacp" to enable 802.3ad (instead of
1710 standard EtherChannel).
1712 The balance-rr, balance-xor and broadcast modes generally
1713 require that the switch have the appropriate ports grouped together.
1714 The nomenclature for such a group differs between switches, it may be
1715 called an "etherchannel" (as in the Cisco example, above), a "trunk
1716 group" or some other similar variation. For these modes, each switch
1717 will also have its own configuration options for the switch's transmit
1718 policy to the bond. Typical choices include XOR of either the MAC or
1719 IP addresses. The transmit policy of the two peers does not need to
1720 match. For these three modes, the bonding mode really selects a
1721 transmit policy for an EtherChannel group; all three will interoperate
1722 with another EtherChannel group.
1725 6. 802.1q VLAN Support
1726 ======================
1728 It is possible to configure VLAN devices over a bond interface
1729 using the 8021q driver. However, only packets coming from the 8021q
1730 driver and passing through bonding will be tagged by default. Self
1731 generated packets, for example, bonding's learning packets or ARP
1732 packets generated by either ALB mode or the ARP monitor mechanism, are
1733 tagged internally by bonding itself. As a result, bonding must
1734 "learn" the VLAN IDs configured above it, and use those IDs to tag
1735 self generated packets.
1737 For reasons of simplicity, and to support the use of adapters
1738 that can do VLAN hardware acceleration offloading, the bonding
1739 interface declares itself as fully hardware offloading capable, it gets
1740 the add_vid/kill_vid notifications to gather the necessary
1741 information, and it propagates those actions to the slaves. In case
1742 of mixed adapter types, hardware accelerated tagged packets that
1743 should go through an adapter that is not offloading capable are
1744 "un-accelerated" by the bonding driver so the VLAN tag sits in the
1747 VLAN interfaces *must* be added on top of a bonding interface
1748 only after enslaving at least one slave. The bonding interface has a
1749 hardware address of 00:00:00:00:00:00 until the first slave is added.
1750 If the VLAN interface is created prior to the first enslavement, it
1751 would pick up the all-zeroes hardware address. Once the first slave
1752 is attached to the bond, the bond device itself will pick up the
1753 slave's hardware address, which is then available for the VLAN device.
1755 Also, be aware that a similar problem can occur if all slaves
1756 are released from a bond that still has one or more VLAN interfaces on
1757 top of it. When a new slave is added, the bonding interface will
1758 obtain its hardware address from the first slave, which might not
1759 match the hardware address of the VLAN interfaces (which was
1760 ultimately copied from an earlier slave).
1762 There are two methods to insure that the VLAN device operates
1763 with the correct hardware address if all slaves are removed from a
1766 1. Remove all VLAN interfaces then recreate them
1768 2. Set the bonding interface's hardware address so that it
1769 matches the hardware address of the VLAN interfaces.
1771 Note that changing a VLAN interface's HW address would set the
1772 underlying device -- i.e. the bonding interface -- to promiscuous
1773 mode, which might not be what you want.
1779 The bonding driver at present supports two schemes for
1780 monitoring a slave device's link state: the ARP monitor and the MII
1783 At the present time, due to implementation restrictions in the
1784 bonding driver itself, it is not possible to enable both ARP and MII
1785 monitoring simultaneously.
1787 7.1 ARP Monitor Operation
1788 -------------------------
1790 The ARP monitor operates as its name suggests: it sends ARP
1791 queries to one or more designated peer systems on the network, and
1792 uses the response as an indication that the link is operating. This
1793 gives some assurance that traffic is actually flowing to and from one
1794 or more peers on the local network.
1796 The ARP monitor relies on the device driver itself to verify
1797 that traffic is flowing. In particular, the driver must keep up to
1798 date the last receive time, dev->last_rx, and transmit start time,
1799 dev->trans_start. If these are not updated by the driver, then the
1800 ARP monitor will immediately fail any slaves using that driver, and
1801 those slaves will stay down. If networking monitoring (tcpdump, etc)
1802 shows the ARP requests and replies on the network, then it may be that
1803 your device driver is not updating last_rx and trans_start.
1805 7.2 Configuring Multiple ARP Targets
1806 ------------------------------------
1808 While ARP monitoring can be done with just one target, it can
1809 be useful in a High Availability setup to have several targets to
1810 monitor. In the case of just one target, the target itself may go
1811 down or have a problem making it unresponsive to ARP requests. Having
1812 an additional target (or several) increases the reliability of the ARP
1815 Multiple ARP targets must be separated by commas as follows:
1817 # example options for ARP monitoring with three targets
1819 options bond0 arp_interval=60 arp_ip_target=192.168.0.1,192.168.0.3,192.168.0.9
1821 For just a single target the options would resemble:
1823 # example options for ARP monitoring with one target
1825 options bond0 arp_interval=60 arp_ip_target=192.168.0.100
1828 7.3 MII Monitor Operation
1829 -------------------------
1831 The MII monitor monitors only the carrier state of the local
1832 network interface. It accomplishes this in one of three ways: by
1833 depending upon the device driver to maintain its carrier state, by
1834 querying the device's MII registers, or by making an ethtool query to
1837 If the use_carrier module parameter is 1 (the default value),
1838 then the MII monitor will rely on the driver for carrier state
1839 information (via the netif_carrier subsystem). As explained in the
1840 use_carrier parameter information, above, if the MII monitor fails to
1841 detect carrier loss on the device (e.g., when the cable is physically
1842 disconnected), it may be that the driver does not support
1845 If use_carrier is 0, then the MII monitor will first query the
1846 device's (via ioctl) MII registers and check the link state. If that
1847 request fails (not just that it returns carrier down), then the MII
1848 monitor will make an ethtool ETHOOL_GLINK request to attempt to obtain
1849 the same information. If both methods fail (i.e., the driver either
1850 does not support or had some error in processing both the MII register
1851 and ethtool requests), then the MII monitor will assume the link is
1854 8. Potential Sources of Trouble
1855 ===============================
1857 8.1 Adventures in Routing
1858 -------------------------
1860 When bonding is configured, it is important that the slave
1861 devices not have routes that supersede routes of the master (or,
1862 generally, not have routes at all). For example, suppose the bonding
1863 device bond0 has two slaves, eth0 and eth1, and the routing table is
1866 Kernel IP routing table
1867 Destination Gateway Genmask Flags MSS Window irtt Iface
1868 10.0.0.0 0.0.0.0 255.255.0.0 U 40 0 0 eth0
1869 10.0.0.0 0.0.0.0 255.255.0.0 U 40 0 0 eth1
1870 10.0.0.0 0.0.0.0 255.255.0.0 U 40 0 0 bond0
1871 127.0.0.0 0.0.0.0 255.0.0.0 U 40 0 0 lo
1873 This routing configuration will likely still update the
1874 receive/transmit times in the driver (needed by the ARP monitor), but
1875 may bypass the bonding driver (because outgoing traffic to, in this
1876 case, another host on network 10 would use eth0 or eth1 before bond0).
1878 The ARP monitor (and ARP itself) may become confused by this
1879 configuration, because ARP requests (generated by the ARP monitor)
1880 will be sent on one interface (bond0), but the corresponding reply
1881 will arrive on a different interface (eth0). This reply looks to ARP
1882 as an unsolicited ARP reply (because ARP matches replies on an
1883 interface basis), and is discarded. The MII monitor is not affected
1884 by the state of the routing table.
1886 The solution here is simply to insure that slaves do not have
1887 routes of their own, and if for some reason they must, those routes do
1888 not supersede routes of their master. This should generally be the
1889 case, but unusual configurations or errant manual or automatic static
1890 route additions may cause trouble.
1892 8.2 Ethernet Device Renaming
1893 ----------------------------
1895 On systems with network configuration scripts that do not
1896 associate physical devices directly with network interface names (so
1897 that the same physical device always has the same "ethX" name), it may
1898 be necessary to add some special logic to config files in
1901 For example, given a modules.conf containing the following:
1904 options bond0 mode=some-mode miimon=50
1910 If neither eth0 and eth1 are slaves to bond0, then when the
1911 bond0 interface comes up, the devices may end up reordered. This
1912 happens because bonding is loaded first, then its slave device's
1913 drivers are loaded next. Since no other drivers have been loaded,
1914 when the e1000 driver loads, it will receive eth0 and eth1 for its
1915 devices, but the bonding configuration tries to enslave eth2 and eth3
1916 (which may later be assigned to the tg3 devices).
1918 Adding the following:
1920 add above bonding e1000 tg3
1922 causes modprobe to load e1000 then tg3, in that order, when
1923 bonding is loaded. This command is fully documented in the
1924 modules.conf manual page.
1926 On systems utilizing modprobe an equivalent problem can occur.
1927 In this case, the following can be added to config files in
1928 /etc/modprobe.d/ as:
1930 softdep bonding pre: tg3 e1000
1932 This will load tg3 and e1000 modules before loading the bonding one.
1933 Full documentation on this can be found in the modprobe.d and modprobe
1936 8.3. Painfully Slow Or No Failed Link Detection By Miimon
1937 ---------------------------------------------------------
1939 By default, bonding enables the use_carrier option, which
1940 instructs bonding to trust the driver to maintain carrier state.
1942 As discussed in the options section, above, some drivers do
1943 not support the netif_carrier_on/_off link state tracking system.
1944 With use_carrier enabled, bonding will always see these links as up,
1945 regardless of their actual state.
1947 Additionally, other drivers do support netif_carrier, but do
1948 not maintain it in real time, e.g., only polling the link state at
1949 some fixed interval. In this case, miimon will detect failures, but
1950 only after some long period of time has expired. If it appears that
1951 miimon is very slow in detecting link failures, try specifying
1952 use_carrier=0 to see if that improves the failure detection time. If
1953 it does, then it may be that the driver checks the carrier state at a
1954 fixed interval, but does not cache the MII register values (so the
1955 use_carrier=0 method of querying the registers directly works). If
1956 use_carrier=0 does not improve the failover, then the driver may cache
1957 the registers, or the problem may be elsewhere.
1959 Also, remember that miimon only checks for the device's
1960 carrier state. It has no way to determine the state of devices on or
1961 beyond other ports of a switch, or if a switch is refusing to pass
1962 traffic while still maintaining carrier on.
1967 If running SNMP agents, the bonding driver should be loaded
1968 before any network drivers participating in a bond. This requirement
1969 is due to the interface index (ipAdEntIfIndex) being associated to
1970 the first interface found with a given IP address. That is, there is
1971 only one ipAdEntIfIndex for each IP address. For example, if eth0 and
1972 eth1 are slaves of bond0 and the driver for eth0 is loaded before the
1973 bonding driver, the interface for the IP address will be associated
1974 with the eth0 interface. This configuration is shown below, the IP
1975 address 192.168.1.1 has an interface index of 2 which indexes to eth0
1976 in the ifDescr table (ifDescr.2).
1978 interfaces.ifTable.ifEntry.ifDescr.1 = lo
1979 interfaces.ifTable.ifEntry.ifDescr.2 = eth0
1980 interfaces.ifTable.ifEntry.ifDescr.3 = eth1
1981 interfaces.ifTable.ifEntry.ifDescr.4 = eth2
1982 interfaces.ifTable.ifEntry.ifDescr.5 = eth3
1983 interfaces.ifTable.ifEntry.ifDescr.6 = bond0
1984 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.10.10.10.10 = 5
1985 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.192.168.1.1 = 2
1986 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.10.74.20.94 = 4
1987 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.127.0.0.1 = 1
1989 This problem is avoided by loading the bonding driver before
1990 any network drivers participating in a bond. Below is an example of
1991 loading the bonding driver first, the IP address 192.168.1.1 is
1992 correctly associated with ifDescr.2.
1994 interfaces.ifTable.ifEntry.ifDescr.1 = lo
1995 interfaces.ifTable.ifEntry.ifDescr.2 = bond0
1996 interfaces.ifTable.ifEntry.ifDescr.3 = eth0
1997 interfaces.ifTable.ifEntry.ifDescr.4 = eth1
1998 interfaces.ifTable.ifEntry.ifDescr.5 = eth2
1999 interfaces.ifTable.ifEntry.ifDescr.6 = eth3
2000 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.10.10.10.10 = 6
2001 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.192.168.1.1 = 2
2002 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.10.74.20.94 = 5
2003 ip.ipAddrTable.ipAddrEntry.ipAdEntIfIndex.127.0.0.1 = 1
2005 While some distributions may not report the interface name in
2006 ifDescr, the association between the IP address and IfIndex remains
2007 and SNMP functions such as Interface_Scan_Next will report that
2010 10. Promiscuous mode
2011 ====================
2013 When running network monitoring tools, e.g., tcpdump, it is
2014 common to enable promiscuous mode on the device, so that all traffic
2015 is seen (instead of seeing only traffic destined for the local host).
2016 The bonding driver handles promiscuous mode changes to the bonding
2017 master device (e.g., bond0), and propagates the setting to the slave
2020 For the balance-rr, balance-xor, broadcast, and 802.3ad modes,
2021 the promiscuous mode setting is propagated to all slaves.
2023 For the active-backup, balance-tlb and balance-alb modes, the
2024 promiscuous mode setting is propagated only to the active slave.
2026 For balance-tlb mode, the active slave is the slave currently
2027 receiving inbound traffic.
2029 For balance-alb mode, the active slave is the slave used as a
2030 "primary." This slave is used for mode-specific control traffic, for
2031 sending to peers that are unassigned or if the load is unbalanced.
2033 For the active-backup, balance-tlb and balance-alb modes, when
2034 the active slave changes (e.g., due to a link failure), the
2035 promiscuous setting will be propagated to the new active slave.
2037 11. Configuring Bonding for High Availability
2038 =============================================
2040 High Availability refers to configurations that provide
2041 maximum network availability by having redundant or backup devices,
2042 links or switches between the host and the rest of the world. The
2043 goal is to provide the maximum availability of network connectivity
2044 (i.e., the network always works), even though other configurations
2045 could provide higher throughput.
2047 11.1 High Availability in a Single Switch Topology
2048 --------------------------------------------------
2050 If two hosts (or a host and a single switch) are directly
2051 connected via multiple physical links, then there is no availability
2052 penalty to optimizing for maximum bandwidth. In this case, there is
2053 only one switch (or peer), so if it fails, there is no alternative
2054 access to fail over to. Additionally, the bonding load balance modes
2055 support link monitoring of their members, so if individual links fail,
2056 the load will be rebalanced across the remaining devices.
2058 See Section 12, "Configuring Bonding for Maximum Throughput"
2059 for information on configuring bonding with one peer device.
2061 11.2 High Availability in a Multiple Switch Topology
2062 ----------------------------------------------------
2064 With multiple switches, the configuration of bonding and the
2065 network changes dramatically. In multiple switch topologies, there is
2066 a trade off between network availability and usable bandwidth.
2068 Below is a sample network, configured to maximize the
2069 availability of the network:
2073 +-----+----+ +-----+----+
2074 | |port2 ISL port2| |
2075 | switch A +--------------------------+ switch B |
2077 +-----+----+ +-----++---+
2080 +-------------+ host1 +---------------+
2083 In this configuration, there is a link between the two
2084 switches (ISL, or inter switch link), and multiple ports connecting to
2085 the outside world ("port3" on each switch). There is no technical
2086 reason that this could not be extended to a third switch.
2088 11.2.1 HA Bonding Mode Selection for Multiple Switch Topology
2089 -------------------------------------------------------------
2091 In a topology such as the example above, the active-backup and
2092 broadcast modes are the only useful bonding modes when optimizing for
2093 availability; the other modes require all links to terminate on the
2094 same peer for them to behave rationally.
2096 active-backup: This is generally the preferred mode, particularly if
2097 the switches have an ISL and play together well. If the
2098 network configuration is such that one switch is specifically
2099 a backup switch (e.g., has lower capacity, higher cost, etc),
2100 then the primary option can be used to insure that the
2101 preferred link is always used when it is available.
2103 broadcast: This mode is really a special purpose mode, and is suitable
2104 only for very specific needs. For example, if the two
2105 switches are not connected (no ISL), and the networks beyond
2106 them are totally independent. In this case, if it is
2107 necessary for some specific one-way traffic to reach both
2108 independent networks, then the broadcast mode may be suitable.
2110 11.2.2 HA Link Monitoring Selection for Multiple Switch Topology
2111 ----------------------------------------------------------------
2113 The choice of link monitoring ultimately depends upon your
2114 switch. If the switch can reliably fail ports in response to other
2115 failures, then either the MII or ARP monitors should work. For
2116 example, in the above example, if the "port3" link fails at the remote
2117 end, the MII monitor has no direct means to detect this. The ARP
2118 monitor could be configured with a target at the remote end of port3,
2119 thus detecting that failure without switch support.
2121 In general, however, in a multiple switch topology, the ARP
2122 monitor can provide a higher level of reliability in detecting end to
2123 end connectivity failures (which may be caused by the failure of any
2124 individual component to pass traffic for any reason). Additionally,
2125 the ARP monitor should be configured with multiple targets (at least
2126 one for each switch in the network). This will insure that,
2127 regardless of which switch is active, the ARP monitor has a suitable
2130 Note, also, that of late many switches now support a functionality
2131 generally referred to as "trunk failover." This is a feature of the
2132 switch that causes the link state of a particular switch port to be set
2133 down (or up) when the state of another switch port goes down (or up).
2134 Its purpose is to propagate link failures from logically "exterior" ports
2135 to the logically "interior" ports that bonding is able to monitor via
2136 miimon. Availability and configuration for trunk failover varies by
2137 switch, but this can be a viable alternative to the ARP monitor when using
2140 12. Configuring Bonding for Maximum Throughput
2141 ==============================================
2143 12.1 Maximizing Throughput in a Single Switch Topology
2144 ------------------------------------------------------
2146 In a single switch configuration, the best method to maximize
2147 throughput depends upon the application and network environment. The
2148 various load balancing modes each have strengths and weaknesses in
2149 different environments, as detailed below.
2151 For this discussion, we will break down the topologies into
2152 two categories. Depending upon the destination of most traffic, we
2153 categorize them into either "gatewayed" or "local" configurations.
2155 In a gatewayed configuration, the "switch" is acting primarily
2156 as a router, and the majority of traffic passes through this router to
2157 other networks. An example would be the following:
2160 +----------+ +----------+
2161 | |eth0 port1| | to other networks
2162 | Host A +---------------------+ router +------------------->
2163 | +---------------------+ | Hosts B and C are out
2164 | |eth1 port2| | here somewhere
2165 +----------+ +----------+
2167 The router may be a dedicated router device, or another host
2168 acting as a gateway. For our discussion, the important point is that
2169 the majority of traffic from Host A will pass through the router to
2170 some other network before reaching its final destination.
2172 In a gatewayed network configuration, although Host A may
2173 communicate with many other systems, all of its traffic will be sent
2174 and received via one other peer on the local network, the router.
2176 Note that the case of two systems connected directly via
2177 multiple physical links is, for purposes of configuring bonding, the
2178 same as a gatewayed configuration. In that case, it happens that all
2179 traffic is destined for the "gateway" itself, not some other network
2182 In a local configuration, the "switch" is acting primarily as
2183 a switch, and the majority of traffic passes through this switch to
2184 reach other stations on the same network. An example would be the
2187 +----------+ +----------+ +--------+
2188 | |eth0 port1| +-------+ Host B |
2189 | Host A +------------+ switch |port3 +--------+
2190 | +------------+ | +--------+
2191 | |eth1 port2| +------------------+ Host C |
2192 +----------+ +----------+port4 +--------+
2195 Again, the switch may be a dedicated switch device, or another
2196 host acting as a gateway. For our discussion, the important point is
2197 that the majority of traffic from Host A is destined for other hosts
2198 on the same local network (Hosts B and C in the above example).
2200 In summary, in a gatewayed configuration, traffic to and from
2201 the bonded device will be to the same MAC level peer on the network
2202 (the gateway itself, i.e., the router), regardless of its final
2203 destination. In a local configuration, traffic flows directly to and
2204 from the final destinations, thus, each destination (Host B, Host C)
2205 will be addressed directly by their individual MAC addresses.
2207 This distinction between a gatewayed and a local network
2208 configuration is important because many of the load balancing modes
2209 available use the MAC addresses of the local network source and
2210 destination to make load balancing decisions. The behavior of each
2211 mode is described below.
2214 12.1.1 MT Bonding Mode Selection for Single Switch Topology
2215 -----------------------------------------------------------
2217 This configuration is the easiest to set up and to understand,
2218 although you will have to decide which bonding mode best suits your
2219 needs. The trade offs for each mode are detailed below:
2221 balance-rr: This mode is the only mode that will permit a single
2222 TCP/IP connection to stripe traffic across multiple
2223 interfaces. It is therefore the only mode that will allow a
2224 single TCP/IP stream to utilize more than one interface's
2225 worth of throughput. This comes at a cost, however: the
2226 striping generally results in peer systems receiving packets out
2227 of order, causing TCP/IP's congestion control system to kick
2228 in, often by retransmitting segments.
2230 It is possible to adjust TCP/IP's congestion limits by
2231 altering the net.ipv4.tcp_reordering sysctl parameter. The
2232 usual default value is 3, and the maximum useful value is 127.
2233 For a four interface balance-rr bond, expect that a single
2234 TCP/IP stream will utilize no more than approximately 2.3
2235 interface's worth of throughput, even after adjusting
2238 Note that the fraction of packets that will be delivered out of
2239 order is highly variable, and is unlikely to be zero. The level
2240 of reordering depends upon a variety of factors, including the
2241 networking interfaces, the switch, and the topology of the
2242 configuration. Speaking in general terms, higher speed network
2243 cards produce more reordering (due to factors such as packet
2244 coalescing), and a "many to many" topology will reorder at a
2245 higher rate than a "many slow to one fast" configuration.
2247 Many switches do not support any modes that stripe traffic
2248 (instead choosing a port based upon IP or MAC level addresses);
2249 for those devices, traffic for a particular connection flowing
2250 through the switch to a balance-rr bond will not utilize greater
2251 than one interface's worth of bandwidth.
2253 If you are utilizing protocols other than TCP/IP, UDP for
2254 example, and your application can tolerate out of order
2255 delivery, then this mode can allow for single stream datagram
2256 performance that scales near linearly as interfaces are added
2259 This mode requires the switch to have the appropriate ports
2260 configured for "etherchannel" or "trunking."
2262 active-backup: There is not much advantage in this network topology to
2263 the active-backup mode, as the inactive backup devices are all
2264 connected to the same peer as the primary. In this case, a
2265 load balancing mode (with link monitoring) will provide the
2266 same level of network availability, but with increased
2267 available bandwidth. On the plus side, active-backup mode
2268 does not require any configuration of the switch, so it may
2269 have value if the hardware available does not support any of
2270 the load balance modes.
2272 balance-xor: This mode will limit traffic such that packets destined
2273 for specific peers will always be sent over the same
2274 interface. Since the destination is determined by the MAC
2275 addresses involved, this mode works best in a "local" network
2276 configuration (as described above), with destinations all on
2277 the same local network. This mode is likely to be suboptimal
2278 if all your traffic is passed through a single router (i.e., a
2279 "gatewayed" network configuration, as described above).
2281 As with balance-rr, the switch ports need to be configured for
2282 "etherchannel" or "trunking."
2284 broadcast: Like active-backup, there is not much advantage to this
2285 mode in this type of network topology.
2287 802.3ad: This mode can be a good choice for this type of network
2288 topology. The 802.3ad mode is an IEEE standard, so all peers
2289 that implement 802.3ad should interoperate well. The 802.3ad
2290 protocol includes automatic configuration of the aggregates,
2291 so minimal manual configuration of the switch is needed
2292 (typically only to designate that some set of devices is
2293 available for 802.3ad). The 802.3ad standard also mandates
2294 that frames be delivered in order (within certain limits), so
2295 in general single connections will not see misordering of
2296 packets. The 802.3ad mode does have some drawbacks: the
2297 standard mandates that all devices in the aggregate operate at
2298 the same speed and duplex. Also, as with all bonding load
2299 balance modes other than balance-rr, no single connection will
2300 be able to utilize more than a single interface's worth of
2303 Additionally, the linux bonding 802.3ad implementation
2304 distributes traffic by peer (using an XOR of MAC addresses),
2305 so in a "gatewayed" configuration, all outgoing traffic will
2306 generally use the same device. Incoming traffic may also end
2307 up on a single device, but that is dependent upon the
2308 balancing policy of the peer's 8023.ad implementation. In a
2309 "local" configuration, traffic will be distributed across the
2310 devices in the bond.
2312 Finally, the 802.3ad mode mandates the use of the MII monitor,
2313 therefore, the ARP monitor is not available in this mode.
2315 balance-tlb: The balance-tlb mode balances outgoing traffic by peer.
2316 Since the balancing is done according to MAC address, in a
2317 "gatewayed" configuration (as described above), this mode will
2318 send all traffic across a single device. However, in a
2319 "local" network configuration, this mode balances multiple
2320 local network peers across devices in a vaguely intelligent
2321 manner (not a simple XOR as in balance-xor or 802.3ad mode),
2322 so that mathematically unlucky MAC addresses (i.e., ones that
2323 XOR to the same value) will not all "bunch up" on a single
2326 Unlike 802.3ad, interfaces may be of differing speeds, and no
2327 special switch configuration is required. On the down side,
2328 in this mode all incoming traffic arrives over a single
2329 interface, this mode requires certain ethtool support in the
2330 network device driver of the slave interfaces, and the ARP
2331 monitor is not available.
2333 balance-alb: This mode is everything that balance-tlb is, and more.
2334 It has all of the features (and restrictions) of balance-tlb,
2335 and will also balance incoming traffic from local network
2336 peers (as described in the Bonding Module Options section,
2339 The only additional down side to this mode is that the network
2340 device driver must support changing the hardware address while
2343 12.1.2 MT Link Monitoring for Single Switch Topology
2344 ----------------------------------------------------
2346 The choice of link monitoring may largely depend upon which
2347 mode you choose to use. The more advanced load balancing modes do not
2348 support the use of the ARP monitor, and are thus restricted to using
2349 the MII monitor (which does not provide as high a level of end to end
2350 assurance as the ARP monitor).
2352 12.2 Maximum Throughput in a Multiple Switch Topology
2353 -----------------------------------------------------
2355 Multiple switches may be utilized to optimize for throughput
2356 when they are configured in parallel as part of an isolated network
2357 between two or more systems, for example:
2363 +--------+ | +---------+
2365 +------+---+ +-----+----+ +-----+----+
2366 | Switch A | | Switch B | | Switch C |
2367 +------+---+ +-----+----+ +-----+----+
2369 +--------+ | +---------+
2375 In this configuration, the switches are isolated from one
2376 another. One reason to employ a topology such as this is for an
2377 isolated network with many hosts (a cluster configured for high
2378 performance, for example), using multiple smaller switches can be more
2379 cost effective than a single larger switch, e.g., on a network with 24
2380 hosts, three 24 port switches can be significantly less expensive than
2381 a single 72 port switch.
2383 If access beyond the network is required, an individual host
2384 can be equipped with an additional network device connected to an
2385 external network; this host then additionally acts as a gateway.
2387 12.2.1 MT Bonding Mode Selection for Multiple Switch Topology
2388 -------------------------------------------------------------
2390 In actual practice, the bonding mode typically employed in
2391 configurations of this type is balance-rr. Historically, in this
2392 network configuration, the usual caveats about out of order packet
2393 delivery are mitigated by the use of network adapters that do not do
2394 any kind of packet coalescing (via the use of NAPI, or because the
2395 device itself does not generate interrupts until some number of
2396 packets has arrived). When employed in this fashion, the balance-rr
2397 mode allows individual connections between two hosts to effectively
2398 utilize greater than one interface's bandwidth.
2400 12.2.2 MT Link Monitoring for Multiple Switch Topology
2401 ------------------------------------------------------
2403 Again, in actual practice, the MII monitor is most often used
2404 in this configuration, as performance is given preference over
2405 availability. The ARP monitor will function in this topology, but its
2406 advantages over the MII monitor are mitigated by the volume of probes
2407 needed as the number of systems involved grows (remember that each
2408 host in the network is configured with bonding).
2410 13. Switch Behavior Issues
2411 ==========================
2413 13.1 Link Establishment and Failover Delays
2414 -------------------------------------------
2416 Some switches exhibit undesirable behavior with regard to the
2417 timing of link up and down reporting by the switch.
2419 First, when a link comes up, some switches may indicate that
2420 the link is up (carrier available), but not pass traffic over the
2421 interface for some period of time. This delay is typically due to
2422 some type of autonegotiation or routing protocol, but may also occur
2423 during switch initialization (e.g., during recovery after a switch
2424 failure). If you find this to be a problem, specify an appropriate
2425 value to the updelay bonding module option to delay the use of the
2426 relevant interface(s).
2428 Second, some switches may "bounce" the link state one or more
2429 times while a link is changing state. This occurs most commonly while
2430 the switch is initializing. Again, an appropriate updelay value may
2433 Note that when a bonding interface has no active links, the
2434 driver will immediately reuse the first link that goes up, even if the
2435 updelay parameter has been specified (the updelay is ignored in this
2436 case). If there are slave interfaces waiting for the updelay timeout
2437 to expire, the interface that first went into that state will be
2438 immediately reused. This reduces down time of the network if the
2439 value of updelay has been overestimated, and since this occurs only in
2440 cases with no connectivity, there is no additional penalty for
2441 ignoring the updelay.
2443 In addition to the concerns about switch timings, if your
2444 switches take a long time to go into backup mode, it may be desirable
2445 to not activate a backup interface immediately after a link goes down.
2446 Failover may be delayed via the downdelay bonding module option.
2448 13.2 Duplicated Incoming Packets
2449 --------------------------------
2451 NOTE: Starting with version 3.0.2, the bonding driver has logic to
2452 suppress duplicate packets, which should largely eliminate this problem.
2453 The following description is kept for reference.
2455 It is not uncommon to observe a short burst of duplicated
2456 traffic when the bonding device is first used, or after it has been
2457 idle for some period of time. This is most easily observed by issuing
2458 a "ping" to some other host on the network, and noticing that the
2459 output from ping flags duplicates (typically one per slave).
2461 For example, on a bond in active-backup mode with five slaves
2462 all connected to one switch, the output may appear as follows:
2465 PING 10.0.4.2 (10.0.4.2) from 10.0.3.10 : 56(84) bytes of data.
2466 64 bytes from 10.0.4.2: icmp_seq=1 ttl=64 time=13.7 ms
2467 64 bytes from 10.0.4.2: icmp_seq=1 ttl=64 time=13.8 ms (DUP!)
2468 64 bytes from 10.0.4.2: icmp_seq=1 ttl=64 time=13.8 ms (DUP!)
2469 64 bytes from 10.0.4.2: icmp_seq=1 ttl=64 time=13.8 ms (DUP!)
2470 64 bytes from 10.0.4.2: icmp_seq=1 ttl=64 time=13.8 ms (DUP!)
2471 64 bytes from 10.0.4.2: icmp_seq=2 ttl=64 time=0.216 ms
2472 64 bytes from 10.0.4.2: icmp_seq=3 ttl=64 time=0.267 ms
2473 64 bytes from 10.0.4.2: icmp_seq=4 ttl=64 time=0.222 ms
2475 This is not due to an error in the bonding driver, rather, it
2476 is a side effect of how many switches update their MAC forwarding
2477 tables. Initially, the switch does not associate the MAC address in
2478 the packet with a particular switch port, and so it may send the
2479 traffic to all ports until its MAC forwarding table is updated. Since
2480 the interfaces attached to the bond may occupy multiple ports on a
2481 single switch, when the switch (temporarily) floods the traffic to all
2482 ports, the bond device receives multiple copies of the same packet
2483 (one per slave device).
2485 The duplicated packet behavior is switch dependent, some
2486 switches exhibit this, and some do not. On switches that display this
2487 behavior, it can be induced by clearing the MAC forwarding table (on
2488 most Cisco switches, the privileged command "clear mac address-table
2489 dynamic" will accomplish this).
2491 14. Hardware Specific Considerations
2492 ====================================
2494 This section contains additional information for configuring
2495 bonding on specific hardware platforms, or for interfacing bonding
2496 with particular switches or other devices.
2498 14.1 IBM BladeCenter
2499 --------------------
2501 This applies to the JS20 and similar systems.
2503 On the JS20 blades, the bonding driver supports only
2504 balance-rr, active-backup, balance-tlb and balance-alb modes. This is
2505 largely due to the network topology inside the BladeCenter, detailed
2508 JS20 network adapter information
2509 --------------------------------
2511 All JS20s come with two Broadcom Gigabit Ethernet ports
2512 integrated on the planar (that's "motherboard" in IBM-speak). In the
2513 BladeCenter chassis, the eth0 port of all JS20 blades is hard wired to
2514 I/O Module #1; similarly, all eth1 ports are wired to I/O Module #2.
2515 An add-on Broadcom daughter card can be installed on a JS20 to provide
2516 two more Gigabit Ethernet ports. These ports, eth2 and eth3, are
2517 wired to I/O Modules 3 and 4, respectively.
2519 Each I/O Module may contain either a switch or a passthrough
2520 module (which allows ports to be directly connected to an external
2521 switch). Some bonding modes require a specific BladeCenter internal
2522 network topology in order to function; these are detailed below.
2524 Additional BladeCenter-specific networking information can be
2525 found in two IBM Redbooks (www.ibm.com/redbooks):
2527 "IBM eServer BladeCenter Networking Options"
2528 "IBM eServer BladeCenter Layer 2-7 Network Switching"
2530 BladeCenter networking configuration
2531 ------------------------------------
2533 Because a BladeCenter can be configured in a very large number
2534 of ways, this discussion will be confined to describing basic
2537 Normally, Ethernet Switch Modules (ESMs) are used in I/O
2538 modules 1 and 2. In this configuration, the eth0 and eth1 ports of a
2539 JS20 will be connected to different internal switches (in the
2540 respective I/O modules).
2542 A passthrough module (OPM or CPM, optical or copper,
2543 passthrough module) connects the I/O module directly to an external
2544 switch. By using PMs in I/O module #1 and #2, the eth0 and eth1
2545 interfaces of a JS20 can be redirected to the outside world and
2546 connected to a common external switch.
2548 Depending upon the mix of ESMs and PMs, the network will
2549 appear to bonding as either a single switch topology (all PMs) or as a
2550 multiple switch topology (one or more ESMs, zero or more PMs). It is
2551 also possible to connect ESMs together, resulting in a configuration
2552 much like the example in "High Availability in a Multiple Switch
2555 Requirements for specific modes
2556 -------------------------------
2558 The balance-rr mode requires the use of passthrough modules
2559 for devices in the bond, all connected to an common external switch.
2560 That switch must be configured for "etherchannel" or "trunking" on the
2561 appropriate ports, as is usual for balance-rr.
2563 The balance-alb and balance-tlb modes will function with
2564 either switch modules or passthrough modules (or a mix). The only
2565 specific requirement for these modes is that all network interfaces
2566 must be able to reach all destinations for traffic sent over the
2567 bonding device (i.e., the network must converge at some point outside
2570 The active-backup mode has no additional requirements.
2572 Link monitoring issues
2573 ----------------------
2575 When an Ethernet Switch Module is in place, only the ARP
2576 monitor will reliably detect link loss to an external switch. This is
2577 nothing unusual, but examination of the BladeCenter cabinet would
2578 suggest that the "external" network ports are the ethernet ports for
2579 the system, when it fact there is a switch between these "external"
2580 ports and the devices on the JS20 system itself. The MII monitor is
2581 only able to detect link failures between the ESM and the JS20 system.
2583 When a passthrough module is in place, the MII monitor does
2584 detect failures to the "external" port, which is then directly
2585 connected to the JS20 system.
2590 The Serial Over LAN (SoL) link is established over the primary
2591 ethernet (eth0) only, therefore, any loss of link to eth0 will result
2592 in losing your SoL connection. It will not fail over with other
2593 network traffic, as the SoL system is beyond the control of the
2596 It may be desirable to disable spanning tree on the switch
2597 (either the internal Ethernet Switch Module, or an external switch) to
2598 avoid fail-over delay issues when using bonding.
2601 15. Frequently Asked Questions
2602 ==============================
2606 Yes. The old 2.0.xx channel bonding patch was not SMP safe.
2607 The new driver was designed to be SMP safe from the start.
2609 2. What type of cards will work with it?
2611 Any Ethernet type cards (you can even mix cards - a Intel
2612 EtherExpress PRO/100 and a 3com 3c905b, for example). For most modes,
2613 devices need not be of the same speed.
2615 Starting with version 3.2.1, bonding also supports Infiniband
2616 slaves in active-backup mode.
2618 3. How many bonding devices can I have?
2622 4. How many slaves can a bonding device have?
2624 This is limited only by the number of network interfaces Linux
2625 supports and/or the number of network cards you can place in your
2628 5. What happens when a slave link dies?
2630 If link monitoring is enabled, then the failing device will be
2631 disabled. The active-backup mode will fail over to a backup link, and
2632 other modes will ignore the failed link. The link will continue to be
2633 monitored, and should it recover, it will rejoin the bond (in whatever
2634 manner is appropriate for the mode). See the sections on High
2635 Availability and the documentation for each mode for additional
2638 Link monitoring can be enabled via either the miimon or
2639 arp_interval parameters (described in the module parameters section,
2640 above). In general, miimon monitors the carrier state as sensed by
2641 the underlying network device, and the arp monitor (arp_interval)
2642 monitors connectivity to another host on the local network.
2644 If no link monitoring is configured, the bonding driver will
2645 be unable to detect link failures, and will assume that all links are
2646 always available. This will likely result in lost packets, and a
2647 resulting degradation of performance. The precise performance loss
2648 depends upon the bonding mode and network configuration.
2650 6. Can bonding be used for High Availability?
2652 Yes. See the section on High Availability for details.
2654 7. Which switches/systems does it work with?
2656 The full answer to this depends upon the desired mode.
2658 In the basic balance modes (balance-rr and balance-xor), it
2659 works with any system that supports etherchannel (also called
2660 trunking). Most managed switches currently available have such
2661 support, and many unmanaged switches as well.
2663 The advanced balance modes (balance-tlb and balance-alb) do
2664 not have special switch requirements, but do need device drivers that
2665 support specific features (described in the appropriate section under
2666 module parameters, above).
2668 In 802.3ad mode, it works with systems that support IEEE
2669 802.3ad Dynamic Link Aggregation. Most managed and many unmanaged
2670 switches currently available support 802.3ad.
2672 The active-backup mode should work with any Layer-II switch.
2674 8. Where does a bonding device get its MAC address from?
2676 When using slave devices that have fixed MAC addresses, or when
2677 the fail_over_mac option is enabled, the bonding device's MAC address is
2678 the MAC address of the active slave.
2680 For other configurations, if not explicitly configured (with
2681 ifconfig or ip link), the MAC address of the bonding device is taken from
2682 its first slave device. This MAC address is then passed to all following
2683 slaves and remains persistent (even if the first slave is removed) until
2684 the bonding device is brought down or reconfigured.
2686 If you wish to change the MAC address, you can set it with
2687 ifconfig or ip link:
2689 # ifconfig bond0 hw ether 00:11:22:33:44:55
2691 # ip link set bond0 address 66:77:88:99:aa:bb
2693 The MAC address can be also changed by bringing down/up the
2694 device and then changing its slaves (or their order):
2696 # ifconfig bond0 down ; modprobe -r bonding
2697 # ifconfig bond0 .... up
2698 # ifenslave bond0 eth...
2700 This method will automatically take the address from the next
2701 slave that is added.
2703 To restore your slaves' MAC addresses, you need to detach them
2704 from the bond (`ifenslave -d bond0 eth0'). The bonding driver will
2705 then restore the MAC addresses that the slaves had before they were
2708 16. Resources and Links
2709 =======================
2711 The latest version of the bonding driver can be found in the latest
2712 version of the linux kernel, found on http://kernel.org
2714 The latest version of this document can be found in the latest kernel
2715 source (named Documentation/networking/bonding.txt).
2717 Discussions regarding the usage of the bonding driver take place on the
2718 bonding-devel mailing list, hosted at sourceforge.net. If you have questions or
2719 problems, post them to the list. The list address is:
2721 bonding-devel@lists.sourceforge.net
2723 The administrative interface (to subscribe or unsubscribe) can
2726 https://lists.sourceforge.net/lists/listinfo/bonding-devel
2728 Discussions regarding the development of the bonding driver take place
2729 on the main Linux network mailing list, hosted at vger.kernel.org. The list
2732 netdev@vger.kernel.org
2734 The administrative interface (to subscribe or unsubscribe) can
2737 http://vger.kernel.org/vger-lists.html#netdev
2739 Donald Becker's Ethernet Drivers and diag programs may be found at :
2740 - http://web.archive.org/web/*/http://www.scyld.com/network/
2742 You will also find a lot of information regarding Ethernet, NWay, MII,
2743 etc. at www.scyld.com.