2 * Copyright 2017 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.
25 #include <type_traits>
28 #include <folly/Hash.h>
29 #include <folly/Likely.h>
30 #include <folly/Portability.h>
35 // This file contains several classes that might be useful if you are
36 // trying to dynamically optimize cache locality: CacheLocality reads
37 // cache sharing information from sysfs to determine how CPUs should be
38 // grouped to minimize contention, Getcpu provides fast access to the
39 // current CPU via __vdso_getcpu, and AccessSpreader uses these two to
40 // optimally spread accesses among a predetermined number of stripes.
42 // AccessSpreader<>::current(n) microbenchmarks at 22 nanos, which is
43 // substantially less than the cost of a cache miss. This means that we
44 // can effectively use it to reduce cache line ping-pong on striped data
45 // structures such as IndexedMemPool or statistics counters.
47 // Because CacheLocality looks at all of the cache levels, it can be
48 // used for different levels of optimization. AccessSpreader(2) does
49 // per-chip spreading on a dual socket system. AccessSpreader(numCpus)
50 // does perfect per-cpu spreading. AccessSpreader(numCpus / 2) does
51 // perfect L1 spreading in a system with hyperthreading enabled.
53 struct CacheLocality {
55 /// 1 more than the maximum value that can be returned from sched_getcpu
56 /// or getcpu. This is the number of hardware thread contexts provided
60 /// Holds the number of caches present at each cache level (0 is
61 /// the closest to the cpu). This is the number of AccessSpreader
62 /// stripes needed to avoid cross-cache communication at the specified
63 /// layer. numCachesByLevel.front() is the number of L1 caches and
64 /// numCachesByLevel.back() is the number of last-level caches.
65 std::vector<size_t> numCachesByLevel;
67 /// A map from cpu (from sched_getcpu or getcpu) to an index in the
68 /// range 0..numCpus-1, where neighboring locality indices are more
69 /// likely to share caches then indices far away. All of the members
70 /// of a particular cache level be contiguous in their locality index.
71 /// For example, if numCpus is 32 and numCachesByLevel.back() is 2,
72 /// then cpus with a locality index < 16 will share one last-level
73 /// cache and cpus with a locality index >= 16 will share the other.
74 std::vector<size_t> localityIndexByCpu;
76 /// Returns the best CacheLocality information available for the current
77 /// system, cached for fast access. This will be loaded from sysfs if
78 /// possible, otherwise it will be correct in the number of CPUs but
79 /// not in their sharing structure.
81 /// If you are into yo dawgs, this is a shared cache of the local
82 /// locality of the shared caches.
84 /// The template parameter here is used to allow injection of a
85 /// repeatable CacheLocality structure during testing. Rather than
86 /// inject the type of the CacheLocality provider into every data type
87 /// that transitively uses it, all components select between the default
88 /// sysfs implementation and a deterministic implementation by keying
89 /// off the type of the underlying atomic. See DeterministicScheduler.
90 template <template <typename> class Atom = std::atomic>
91 static const CacheLocality& system();
93 /// Reads CacheLocality information from a tree structured like
94 /// the sysfs filesystem. The provided function will be evaluated
95 /// for each sysfs file that needs to be queried. The function
96 /// should return a string containing the first line of the file
97 /// (not including the newline), or an empty string if the file does
98 /// not exist. The function will be called with paths of the form
99 /// /sys/devices/system/cpu/cpu*/cache/index*/{type,shared_cpu_list} .
100 /// Throws an exception if no caches can be parsed at all.
101 static CacheLocality readFromSysfsTree(
102 const std::function<std::string(std::string)>& mapping);
104 /// Reads CacheLocality information from the real sysfs filesystem.
105 /// Throws an exception if no cache information can be loaded.
106 static CacheLocality readFromSysfs();
108 /// Returns a usable (but probably not reflective of reality)
109 /// CacheLocality structure with the specified number of cpus and a
110 /// single cache level that associates one cpu per cache.
111 static CacheLocality uniform(size_t numCpus);
114 /// Memory locations on the same cache line are subject to false
115 /// sharing, which is very bad for performance. Microbenchmarks
116 /// indicate that pairs of cache lines also see interference under
117 /// heavy use of atomic operations (observed for atomic increment on
118 /// Sandy Bridge). See FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
119 kFalseSharingRange = 128
123 kFalseSharingRange == 128,
124 "FOLLY_ALIGN_TO_AVOID_FALSE_SHARING should track kFalseSharingRange");
127 // TODO replace __attribute__ with alignas and 128 with kFalseSharingRange
129 /// An attribute that will cause a variable or field to be aligned so that
130 /// it doesn't have false sharing with anything at a smaller memory address.
131 #define FOLLY_ALIGN_TO_AVOID_FALSE_SHARING FOLLY_ALIGNED(128)
133 /// Knows how to derive a function pointer to the VDSO implementation of
134 /// getcpu(2), if available
136 /// Function pointer to a function with the same signature as getcpu(2).
137 typedef int (*Func)(unsigned* cpu, unsigned* node, void* unused);
139 /// Returns a pointer to the VDSO implementation of getcpu(2), if
140 /// available, or nullptr otherwise. This function may be quite
141 /// expensive, be sure to cache the result.
142 static Func resolveVdsoFunc();
146 template <template <typename> class Atom>
147 struct SequentialThreadId {
149 /// Returns the thread id assigned to the current thread
150 static unsigned get() {
152 if (UNLIKELY(rv == 0)) {
153 rv = currentId = ++prevId;
159 static Atom<unsigned> prevId;
161 static FOLLY_TLS unsigned currentId;
164 template <template <typename> class Atom>
165 Atom<unsigned> SequentialThreadId<Atom>::prevId(0);
167 template <template <typename> class Atom>
168 FOLLY_TLS unsigned SequentialThreadId<Atom>::currentId(0);
170 // Suppress this instantiation in other translation units. It is
171 // instantiated in CacheLocality.cpp
172 extern template struct SequentialThreadId<std::atomic>;
175 struct HashingThreadId {
176 static unsigned get() {
177 pthread_t pid = pthread_self();
179 memcpy(&id, &pid, std::min(sizeof(pid), sizeof(id)));
180 return hash::twang_32from64(id);
184 /// A class that lazily binds a unique (for each implementation of Atom)
185 /// identifier to a thread. This is a fallback mechanism for the access
186 /// spreader if __vdso_getcpu can't be loaded
187 template <typename ThreadId>
188 struct FallbackGetcpu {
189 /// Fills the thread id into the cpu and node out params (if they
190 /// are non-null). This method is intended to act like getcpu when a
191 /// fast-enough form of getcpu isn't available or isn't desired
192 static int getcpu(unsigned* cpu, unsigned* node, void* /* unused */) {
193 auto id = ThreadId::get();
205 typedef FallbackGetcpu<SequentialThreadId<std::atomic>> FallbackGetcpuType;
207 typedef FallbackGetcpu<HashingThreadId> FallbackGetcpuType;
210 /// AccessSpreader arranges access to a striped data structure in such a
211 /// way that concurrently executing threads are likely to be accessing
212 /// different stripes. It does NOT guarantee uncontended access.
213 /// Your underlying algorithm must be thread-safe without spreading, this
214 /// is merely an optimization. AccessSpreader::current(n) is typically
215 /// much faster than a cache miss (12 nanos on my dev box, tested fast
216 /// in both 2.6 and 3.2 kernels).
218 /// If available (and not using the deterministic testing implementation)
219 /// AccessSpreader uses the getcpu system call via VDSO and the
220 /// precise locality information retrieved from sysfs by CacheLocality.
221 /// This provides optimal anti-sharing at a fraction of the cost of a
224 /// When there are not as many stripes as processors, we try to optimally
225 /// place the cache sharing boundaries. This means that if you have 2
226 /// stripes and run on a dual-socket system, your 2 stripes will each get
227 /// all of the cores from a single socket. If you have 16 stripes on a
228 /// 16 core system plus hyperthreading (32 cpus), each core will get its
229 /// own stripe and there will be no cache sharing at all.
231 /// AccessSpreader has a fallback mechanism for when __vdso_getcpu can't be
232 /// loaded, or for use during deterministic testing. Using sched_getcpu
233 /// or the getcpu syscall would negate the performance advantages of
234 /// access spreading, so we use a thread-local value and a shared atomic
235 /// counter to spread access out. On systems lacking both a fast getcpu()
236 /// and TLS, we hash the thread id to spread accesses.
238 /// AccessSpreader is templated on the template type that is used
239 /// to implement atomics, as a way to instantiate the underlying
240 /// heuristics differently for production use and deterministic unit
241 /// testing. See DeterministicScheduler for more. If you aren't using
242 /// DeterministicScheduler, you can just use the default template parameter
244 template <template <typename> class Atom = std::atomic>
245 struct AccessSpreader {
247 /// Returns the stripe associated with the current CPU. The returned
248 /// value will be < numStripes.
249 static size_t current(size_t numStripes) {
250 // widthAndCpuToStripe[0] will actually work okay (all zeros), but
251 // something's wrong with the caller
252 assert(numStripes > 0);
255 getcpuFunc(&cpu, nullptr, nullptr);
256 return widthAndCpuToStripe[std::min(size_t(kMaxCpus),
257 numStripes)][cpu % kMaxCpus];
261 /// If there are more cpus than this nothing will crash, but there
262 /// might be unnecessary sharing
263 enum { kMaxCpus = 128 };
265 typedef uint8_t CompactStripe;
267 static_assert((kMaxCpus & (kMaxCpus - 1)) == 0,
268 "kMaxCpus should be a power of two so modulo is fast");
269 static_assert(kMaxCpus - 1 <= std::numeric_limits<CompactStripe>::max(),
270 "stripeByCpu element type isn't wide enough");
272 /// Points to the getcpu-like function we are using to obtain the
273 /// current cpu. It should not be assumed that the returned cpu value
274 /// is in range. We use a static for this so that we can prearrange a
275 /// valid value in the pre-constructed state and avoid the need for a
276 /// conditional on every subsequent invocation (not normally a big win,
277 /// but 20% on some inner loops here).
278 static Getcpu::Func getcpuFunc;
280 /// For each level of splitting up to kMaxCpus, maps the cpu (mod
281 /// kMaxCpus) to the stripe. Rather than performing any inequalities
282 /// or modulo on the actual number of cpus, we just fill in the entire
284 static CompactStripe widthAndCpuToStripe[kMaxCpus + 1][kMaxCpus];
286 static bool initialized;
288 /// Returns the best getcpu implementation for Atom
289 static Getcpu::Func pickGetcpuFunc() {
290 auto best = Getcpu::resolveVdsoFunc();
291 return best ? best : &FallbackGetcpuType::getcpu;
294 /// Always claims to be on CPU zero, node zero
295 static int degenerateGetcpu(unsigned* cpu, unsigned* node, void*) {
296 if (cpu != nullptr) {
299 if (node != nullptr) {
305 // The function to call for fast lookup of getcpu is a singleton, as
306 // is the precomputed table of locality information. AccessSpreader
307 // is used in very tight loops, however (we're trying to race an L1
308 // cache miss!), so the normal singleton mechanisms are noticeably
309 // expensive. Even a not-taken branch guarding access to getcpuFunc
310 // slows AccessSpreader::current from 12 nanos to 14. As a result, we
311 // populate the static members with simple (but valid) values that can
312 // be filled in by the linker, and then follow up with a normal static
313 // initializer call that puts in the proper version. This means that
314 // when there are initialization order issues we will just observe a
315 // zero stripe. Once a sanitizer gets smart enough to detect this as
316 // a race or undefined behavior, we can annotate it.
318 static bool initialize() {
319 getcpuFunc = pickGetcpuFunc();
321 auto& cacheLocality = CacheLocality::system<Atom>();
322 auto n = cacheLocality.numCpus;
323 for (size_t width = 0; width <= kMaxCpus; ++width) {
324 auto numStripes = std::max(size_t{1}, width);
325 for (size_t cpu = 0; cpu < kMaxCpus && cpu < n; ++cpu) {
326 auto index = cacheLocality.localityIndexByCpu[cpu];
328 // as index goes from 0..n, post-transform value goes from
330 widthAndCpuToStripe[width][cpu] =
331 CompactStripe((index * numStripes) / n);
332 assert(widthAndCpuToStripe[width][cpu] < numStripes);
334 for (size_t cpu = n; cpu < kMaxCpus; ++cpu) {
335 widthAndCpuToStripe[width][cpu] = widthAndCpuToStripe[width][cpu - n];
342 template <template <typename> class Atom>
343 Getcpu::Func AccessSpreader<Atom>::getcpuFunc =
344 AccessSpreader<Atom>::degenerateGetcpu;
346 template <template <typename> class Atom>
347 typename AccessSpreader<Atom>::CompactStripe
348 AccessSpreader<Atom>::widthAndCpuToStripe[kMaxCpus + 1][kMaxCpus] = {};
350 template <template <typename> class Atom>
351 bool AccessSpreader<Atom>::initialized = AccessSpreader<Atom>::initialize();
353 // Suppress this instantiation in other translation units. It is
354 // instantiated in CacheLocality.cpp
355 extern template struct AccessSpreader<std::atomic>;
357 } // namespace detail