2 * Copyright 2016 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
26 #include <type_traits>
29 #include <folly/Hash.h>
30 #include <folly/Likely.h>
31 #include <folly/Portability.h>
36 // This file contains several classes that might be useful if you are
37 // trying to dynamically optimize cache locality: CacheLocality reads
38 // cache sharing information from sysfs to determine how CPUs should be
39 // grouped to minimize contention, Getcpu provides fast access to the
40 // current CPU via __vdso_getcpu, and AccessSpreader uses these two to
41 // optimally spread accesses among a predetermined number of stripes.
43 // AccessSpreader<>::current(n) microbenchmarks at 22 nanos, which is
44 // substantially less than the cost of a cache miss. This means that we
45 // can effectively use it to reduce cache line ping-pong on striped data
46 // structures such as IndexedMemPool or statistics counters.
48 // Because CacheLocality looks at all of the cache levels, it can be
49 // used for different levels of optimization. AccessSpreader(2) does
50 // per-chip spreading on a dual socket system. AccessSpreader(numCpus)
51 // does perfect per-cpu spreading. AccessSpreader(numCpus / 2) does
52 // perfect L1 spreading in a system with hyperthreading enabled.
54 struct CacheLocality {
56 /// 1 more than the maximum value that can be returned from sched_getcpu
57 /// or getcpu. This is the number of hardware thread contexts provided
61 /// Holds the number of caches present at each cache level (0 is
62 /// the closest to the cpu). This is the number of AccessSpreader
63 /// stripes needed to avoid cross-cache communication at the specified
64 /// layer. numCachesByLevel.front() is the number of L1 caches and
65 /// numCachesByLevel.back() is the number of last-level caches.
66 std::vector<size_t> numCachesByLevel;
68 /// A map from cpu (from sched_getcpu or getcpu) to an index in the
69 /// range 0..numCpus-1, where neighboring locality indices are more
70 /// likely to share caches then indices far away. All of the members
71 /// of a particular cache level be contiguous in their locality index.
72 /// For example, if numCpus is 32 and numCachesByLevel.back() is 2,
73 /// then cpus with a locality index < 16 will share one last-level
74 /// cache and cpus with a locality index >= 16 will share the other.
75 std::vector<size_t> localityIndexByCpu;
77 /// Returns the best CacheLocality information available for the current
78 /// system, cached for fast access. This will be loaded from sysfs if
79 /// possible, otherwise it will be correct in the number of CPUs but
80 /// not in their sharing structure.
82 /// If you are into yo dawgs, this is a shared cache of the local
83 /// locality of the shared caches.
85 /// The template parameter here is used to allow injection of a
86 /// repeatable CacheLocality structure during testing. Rather than
87 /// inject the type of the CacheLocality provider into every data type
88 /// that transitively uses it, all components select between the default
89 /// sysfs implementation and a deterministic implementation by keying
90 /// off the type of the underlying atomic. See DeterministicScheduler.
91 template <template <typename> class Atom = std::atomic>
92 static const CacheLocality& system();
94 /// Reads CacheLocality information from a tree structured like
95 /// the sysfs filesystem. The provided function will be evaluated
96 /// for each sysfs file that needs to be queried. The function
97 /// should return a string containing the first line of the file
98 /// (not including the newline), or an empty string if the file does
99 /// not exist. The function will be called with paths of the form
100 /// /sys/devices/system/cpu/cpu*/cache/index*/{type,shared_cpu_list} .
101 /// Throws an exception if no caches can be parsed at all.
102 static CacheLocality readFromSysfsTree(
103 const std::function<std::string(std::string)>& mapping);
105 /// Reads CacheLocality information from the real sysfs filesystem.
106 /// Throws an exception if no cache information can be loaded.
107 static CacheLocality readFromSysfs();
109 /// Returns a usable (but probably not reflective of reality)
110 /// CacheLocality structure with the specified number of cpus and a
111 /// single cache level that associates one cpu per cache.
112 static CacheLocality uniform(size_t numCpus);
115 /// Memory locations on the same cache line are subject to false
116 /// sharing, which is very bad for performance. Microbenchmarks
117 /// indicate that pairs of cache lines also see interference under
118 /// heavy use of atomic operations (observed for atomic increment on
119 /// Sandy Bridge). See FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
120 kFalseSharingRange = 128
124 kFalseSharingRange == 128,
125 "FOLLY_ALIGN_TO_AVOID_FALSE_SHARING should track kFalseSharingRange");
128 // TODO replace __attribute__ with alignas and 128 with kFalseSharingRange
130 /// An attribute that will cause a variable or field to be aligned so that
131 /// it doesn't have false sharing with anything at a smaller memory address.
132 #define FOLLY_ALIGN_TO_AVOID_FALSE_SHARING FOLLY_ALIGNED(128)
134 /// Knows how to derive a function pointer to the VDSO implementation of
135 /// getcpu(2), if available
137 /// Function pointer to a function with the same signature as getcpu(2).
138 typedef int (*Func)(unsigned* cpu, unsigned* node, void* unused);
140 /// Returns a pointer to the VDSO implementation of getcpu(2), if
141 /// available, or nullptr otherwise. This function may be quite
142 /// expensive, be sure to cache the result.
143 static Func resolveVdsoFunc();
147 template <template <typename> class Atom>
148 struct SequentialThreadId {
150 /// Returns the thread id assigned to the current thread
151 static size_t get() {
153 if (UNLIKELY(rv == 0)) {
154 rv = currentId = ++prevId;
160 static Atom<size_t> prevId;
162 static FOLLY_TLS size_t currentId;
166 struct HashingThreadId {
167 static size_t get() {
168 pthread_t pid = pthread_self();
170 memcpy(&id, &pid, std::min(sizeof(pid), sizeof(id)));
171 return hash::twang_32from64(id);
175 /// A class that lazily binds a unique (for each implementation of Atom)
176 /// identifier to a thread. This is a fallback mechanism for the access
177 /// spreader if __vdso_getcpu can't be loaded
178 template <typename ThreadId>
179 struct FallbackGetcpu {
180 /// Fills the thread id into the cpu and node out params (if they
181 /// are non-null). This method is intended to act like getcpu when a
182 /// fast-enough form of getcpu isn't available or isn't desired
183 static int getcpu(unsigned* cpu, unsigned* node, void* /* unused */) {
184 auto id = ThreadId::get();
196 typedef FallbackGetcpu<SequentialThreadId<std::atomic>> FallbackGetcpuType;
198 typedef FallbackGetcpu<HashingThreadId> FallbackGetcpuType;
201 /// AccessSpreader arranges access to a striped data structure in such a
202 /// way that concurrently executing threads are likely to be accessing
203 /// different stripes. It does NOT guarantee uncontended access.
204 /// Your underlying algorithm must be thread-safe without spreading, this
205 /// is merely an optimization. AccessSpreader::current(n) is typically
206 /// much faster than a cache miss (12 nanos on my dev box, tested fast
207 /// in both 2.6 and 3.2 kernels).
209 /// If available (and not using the deterministic testing implementation)
210 /// AccessSpreader uses the getcpu system call via VDSO and the
211 /// precise locality information retrieved from sysfs by CacheLocality.
212 /// This provides optimal anti-sharing at a fraction of the cost of a
215 /// When there are not as many stripes as processors, we try to optimally
216 /// place the cache sharing boundaries. This means that if you have 2
217 /// stripes and run on a dual-socket system, your 2 stripes will each get
218 /// all of the cores from a single socket. If you have 16 stripes on a
219 /// 16 core system plus hyperthreading (32 cpus), each core will get its
220 /// own stripe and there will be no cache sharing at all.
222 /// AccessSpreader has a fallback mechanism for when __vdso_getcpu can't be
223 /// loaded, or for use during deterministic testing. Using sched_getcpu
224 /// or the getcpu syscall would negate the performance advantages of
225 /// access spreading, so we use a thread-local value and a shared atomic
226 /// counter to spread access out. On systems lacking both a fast getcpu()
227 /// and TLS, we hash the thread id to spread accesses.
229 /// AccessSpreader is templated on the template type that is used
230 /// to implement atomics, as a way to instantiate the underlying
231 /// heuristics differently for production use and deterministic unit
232 /// testing. See DeterministicScheduler for more. If you aren't using
233 /// DeterministicScheduler, you can just use the default template parameter
235 template <template <typename> class Atom = std::atomic>
236 struct AccessSpreader {
238 /// Returns the stripe associated with the current CPU. The returned
239 /// value will be < numStripes.
240 static size_t current(size_t numStripes) {
241 // widthAndCpuToStripe[0] will actually work okay (all zeros), but
242 // something's wrong with the caller
243 assert(numStripes > 0);
246 getcpuFunc(&cpu, nullptr, nullptr);
247 return widthAndCpuToStripe[std::min(size_t(kMaxCpus),
248 numStripes)][cpu % kMaxCpus];
252 /// If there are more cpus than this nothing will crash, but there
253 /// might be unnecessary sharing
254 enum { kMaxCpus = 128 };
256 typedef uint8_t CompactStripe;
258 static_assert((kMaxCpus & (kMaxCpus - 1)) == 0,
259 "kMaxCpus should be a power of two so modulo is fast");
260 static_assert(kMaxCpus - 1 <= std::numeric_limits<CompactStripe>::max(),
261 "stripeByCpu element type isn't wide enough");
263 /// Points to the getcpu-like function we are using to obtain the
264 /// current cpu. It should not be assumed that the returned cpu value
265 /// is in range. We use a static for this so that we can prearrange a
266 /// valid value in the pre-constructed state and avoid the need for a
267 /// conditional on every subsequent invocation (not normally a big win,
268 /// but 20% on some inner loops here).
269 static Getcpu::Func getcpuFunc;
271 /// For each level of splitting up to kMaxCpus, maps the cpu (mod
272 /// kMaxCpus) to the stripe. Rather than performing any inequalities
273 /// or modulo on the actual number of cpus, we just fill in the entire
275 static CompactStripe widthAndCpuToStripe[kMaxCpus + 1][kMaxCpus];
277 static bool initialized;
279 /// Returns the best getcpu implementation for Atom
280 static Getcpu::Func pickGetcpuFunc();
282 /// Always claims to be on CPU zero, node zero
283 static int degenerateGetcpu(unsigned* cpu, unsigned* node, void*) {
284 if (cpu != nullptr) {
287 if (node != nullptr) {
293 // The function to call for fast lookup of getcpu is a singleton, as
294 // is the precomputed table of locality information. AccessSpreader
295 // is used in very tight loops, however (we're trying to race an L1
296 // cache miss!), so the normal singleton mechanisms are noticeably
297 // expensive. Even a not-taken branch guarding access to getcpuFunc
298 // slows AccessSpreader::current from 12 nanos to 14. As a result, we
299 // populate the static members with simple (but valid) values that can
300 // be filled in by the linker, and then follow up with a normal static
301 // initializer call that puts in the proper version. This means that
302 // when there are initialization order issues we will just observe a
303 // zero stripe. Once a sanitizer gets smart enough to detect this as
304 // a race or undefined behavior, we can annotate it.
306 static bool initialize() {
307 getcpuFunc = pickGetcpuFunc();
309 auto& cacheLocality = CacheLocality::system<Atom>();
310 auto n = cacheLocality.numCpus;
311 for (size_t width = 0; width <= kMaxCpus; ++width) {
312 auto numStripes = std::max(size_t{1}, width);
313 for (size_t cpu = 0; cpu < kMaxCpus && cpu < n; ++cpu) {
314 auto index = cacheLocality.localityIndexByCpu[cpu];
316 // as index goes from 0..n, post-transform value goes from
318 widthAndCpuToStripe[width][cpu] = (index * numStripes) / n;
319 assert(widthAndCpuToStripe[width][cpu] < numStripes);
321 for (size_t cpu = n; cpu < kMaxCpus; ++cpu) {
322 widthAndCpuToStripe[width][cpu] = widthAndCpuToStripe[width][cpu - n];
330 Getcpu::Func AccessSpreader<std::atomic>::pickGetcpuFunc();
332 #define DECLARE_ACCESS_SPREADER_TYPE(Atom) \
336 Getcpu::Func AccessSpreader<Atom>::getcpuFunc = \
337 AccessSpreader<Atom>::degenerateGetcpu; \
339 typename AccessSpreader<Atom>::CompactStripe \
340 AccessSpreader<Atom>::widthAndCpuToStripe[129][128] = {}; \
342 bool AccessSpreader<Atom>::initialized = AccessSpreader<Atom>::initialize(); \
346 } // namespace detail