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.
17 #ifndef FOLLY_DETAIL_CACHELOCALITY_H_
18 #define FOLLY_DETAIL_CACHELOCALITY_H_
27 #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 /// Holds a function pointer to the VDSO implementation of getcpu(2),
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
142 static Func vdsoFunc();
146 template <template <typename> class Atom>
147 struct SequentialThreadId {
149 /// Returns the thread id assigned to the current thread
150 static size_t get() {
152 if (UNLIKELY(rv == 0)) {
153 rv = currentId = ++prevId;
159 static Atom<size_t> prevId;
161 static FOLLY_TLS size_t currentId;
165 struct HashingThreadId {
166 static size_t get() {
167 pthread_t pid = pthread_self();
169 memcpy(&id, &pid, std::min(sizeof(pid), sizeof(id)));
170 return hash::twang_32from64(id);
174 /// A class that lazily binds a unique (for each implementation of Atom)
175 /// identifier to a thread. This is a fallback mechanism for the access
176 /// spreader if __vdso_getcpu can't be loaded
177 template <typename ThreadId>
178 struct FallbackGetcpu {
179 /// Fills the thread id into the cpu and node out params (if they
180 /// are non-null). This method is intended to act like getcpu when a
181 /// fast-enough form of getcpu isn't available or isn't desired
182 static int getcpu(unsigned* cpu, unsigned* node, void* /* unused */) {
183 auto id = ThreadId::get();
195 typedef FallbackGetcpu<SequentialThreadId<std::atomic>> FallbackGetcpuType;
197 typedef FallbackGetcpu<HashingThreadId> FallbackGetcpuType;
200 template <template <typename> class Atom, size_t kMaxCpus>
201 struct AccessSpreaderArray;
203 /// AccessSpreader arranges access to a striped data structure in such a
204 /// way that concurrently executing threads are likely to be accessing
205 /// different stripes. It does NOT guarantee uncontended access.
206 /// Your underlying algorithm must be thread-safe without spreading, this
207 /// is merely an optimization. AccessSpreader::current(n) is typically
208 /// much faster than a cache miss (22 nanos on my dev box, tested fast
209 /// in both 2.6 and 3.2 kernels).
211 /// You are free to create your own AccessSpreader-s or to cache the
212 /// results of AccessSpreader<>::shared(n), but you will probably want
213 /// to use one of the system-wide shared ones. Calling .current() on
214 /// a particular AccessSpreader instance only saves about 1 nanosecond
215 /// over calling AccessSpreader<>::shared(n).
217 /// If available (and not using the deterministic testing implementation)
218 /// AccessSpreader uses the getcpu system call via VDSO and the
219 /// precise locality information retrieved from sysfs by CacheLocality.
220 /// This provides optimal anti-sharing at a fraction of the cost of a
223 /// When there are not as many stripes as processors, we try to optimally
224 /// place the cache sharing boundaries. This means that if you have 2
225 /// stripes and run on a dual-socket system, your 2 stripes will each get
226 /// all of the cores from a single socket. If you have 16 stripes on a
227 /// 16 core system plus hyperthreading (32 cpus), each core will get its
228 /// own stripe and there will be no cache sharing at all.
230 /// AccessSpreader has a fallback mechanism for when __vdso_getcpu can't be
231 /// loaded, or for use during deterministic testing. Using sched_getcpu or
232 /// the getcpu syscall would negate the performance advantages of access
233 /// spreading, so we use a thread-local value and a shared atomic counter
234 /// to spread access out.
236 /// AccessSpreader is templated on the template type that is used
237 /// to implement atomics, as a way to instantiate the underlying
238 /// heuristics differently for production use and deterministic unit
239 /// testing. See DeterministicScheduler for more. If you aren't using
240 /// DeterministicScheduler, you can just use the default template parameter
242 template <template <typename> class Atom = std::atomic>
243 struct AccessSpreader {
245 /// Returns a never-destructed shared AccessSpreader instance.
246 /// numStripes should be > 0.
247 static const AccessSpreader& shared(size_t numStripes) {
248 // sharedInstances[0] actually has numStripes == 1
249 assert(numStripes > 0);
251 // the last shared element handles all large sizes
252 return AccessSpreaderArray<Atom, kMaxCpus>::sharedInstance[std::min(
253 size_t(kMaxCpus), numStripes)];
256 /// Returns the stripe associated with the current CPU, assuming
257 /// that there are numStripes (non-zero) stripes. Equivalent to
258 /// AccessSpreader::shared(numStripes)->current.
259 static size_t current(size_t numStripes) {
260 return shared(numStripes).current();
263 /// stripeByCore uses 1 stripe per L1 cache, according to
264 /// CacheLocality::system<>(). Use stripeByCore.numStripes() to see
265 /// its width, or stripeByCore.current() to get the current stripe
266 static const AccessSpreader stripeByCore;
268 /// stripeByChip uses 1 stripe per last-level cache, which is the fewest
269 /// number of stripes for which off-chip communication can be avoided
270 /// (assuming all caches are on-chip). Use stripeByChip.numStripes()
271 /// to see its width, or stripeByChip.current() to get the current stripe
272 static const AccessSpreader stripeByChip;
274 /// Constructs an AccessSpreader that will return values from
275 /// 0 to numStripes-1 (inclusive), precomputing the mapping
276 /// from CPU to stripe. There is no use in having more than
277 /// CacheLocality::system<Atom>().localityIndexByCpu.size() stripes or
279 explicit AccessSpreader(
280 size_t spreaderNumStripes,
281 const CacheLocality& cacheLocality = CacheLocality::system<Atom>(),
282 Getcpu::Func getcpuFunc = nullptr)
283 : getcpuFunc_(getcpuFunc ? getcpuFunc
284 : pickGetcpuFunc(spreaderNumStripes)),
285 numStripes_(spreaderNumStripes) {
286 auto n = cacheLocality.numCpus;
287 for (size_t cpu = 0; cpu < kMaxCpus && cpu < n; ++cpu) {
288 auto index = cacheLocality.localityIndexByCpu[cpu];
290 // as index goes from 0..n, post-transform value goes from
292 stripeByCpu[cpu] = (index * numStripes_) / n;
293 assert(stripeByCpu[cpu] < numStripes_);
295 for (size_t cpu = n; cpu < kMaxCpus; ++cpu) {
296 stripeByCpu[cpu] = stripeByCpu[cpu - n];
300 /// Returns 1 more than the maximum value that can be returned from
302 size_t numStripes() const { return numStripes_; }
304 /// Returns the stripe associated with the current CPU
305 size_t current() const {
307 getcpuFunc_(&cpu, nullptr, nullptr);
308 return stripeByCpu[cpu % kMaxCpus];
312 /// If there are more cpus than this nothing will crash, but there
313 /// might be unnecessary sharing
314 enum { kMaxCpus = 128 };
316 typedef uint8_t CompactStripe;
318 static_assert((kMaxCpus & (kMaxCpus - 1)) == 0,
319 "kMaxCpus should be a power of two so modulo is fast");
320 static_assert(kMaxCpus - 1 <= std::numeric_limits<CompactStripe>::max(),
321 "stripeByCpu element type isn't wide enough");
323 /// Points to the getcpu-like function we are using to obtain the
324 /// current cpu. It should not be assumed that the returned cpu value
325 /// is in range. We use a member for this instead of a static so that
326 /// this fetch preloads a prefix the stripeByCpu array
327 Getcpu::Func getcpuFunc_;
329 /// A precomputed map from cpu to stripe. Rather than add a layer of
330 /// indirection requiring a dynamic bounds check and another cache miss,
331 /// we always precompute the whole array
332 CompactStripe stripeByCpu[kMaxCpus];
336 /// Returns the best getcpu implementation for this type and width
337 /// of AccessSpreader
338 static Getcpu::Func pickGetcpuFunc(size_t numStripes);
342 Getcpu::Func AccessSpreader<std::atomic>::pickGetcpuFunc(size_t);
344 /// An array of kMaxCpus+1 AccessSpreader<Atom> instances constructed
345 /// with default params, with the zero-th element having 1 stripe
346 template <template <typename> class Atom, size_t kMaxStripe>
347 struct AccessSpreaderArray {
349 AccessSpreaderArray() {
350 for (size_t i = 0; i <= kMaxStripe; ++i) {
351 new (raw + i) AccessSpreader<Atom>(std::max(size_t(1), i));
355 ~AccessSpreaderArray() {
356 for (size_t i = 0; i <= kMaxStripe; ++i) {
357 auto p = static_cast<AccessSpreader<Atom>*>(static_cast<void*>(raw + i));
358 p->~AccessSpreader();
362 AccessSpreader<Atom> const& operator[](size_t index) const {
363 return *static_cast<AccessSpreader<Atom> const*>(
364 static_cast<void const*>(raw + index));
368 // AccessSpreader uses sharedInstance
369 friend AccessSpreader<Atom>;
371 static AccessSpreaderArray<Atom, kMaxStripe> sharedInstance;
373 /// aligned_storage is uninitialized, we use placement new since there
374 /// is no AccessSpreader default constructor
375 typename std::aligned_storage<sizeof(AccessSpreader<Atom>),
376 CacheLocality::kFalseSharingRange>::type
382 #endif /* FOLLY_DETAIL_CacheLocality_H_ */