You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/cache/fast_lru_cache.h

455 lines
16 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#pragma once
#include <array>
#include <memory>
#include <string>
#include "cache/cache_key.h"
#include "cache/sharded_cache.h"
#include "port/lang.h"
#include "port/malloc.h"
#include "port/port.h"
#include "rocksdb/secondary_cache.h"
#include "util/autovector.h"
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
#include "util/distributed_mutex.h"
namespace ROCKSDB_NAMESPACE {
namespace fast_lru_cache {
// LRU cache implementation using an open-address hash table.
// Every slot in the hash table is an LRUHandle. Because handles can be
// referenced externally, we can't discard them immediately once they are
// deleted (via a delete or an LRU eviction) or replaced by a new version
// (via an insert of the same key). The state of an element is defined by
// the following two properties:
// (R) Referenced: An element can be referenced externally (refs > 0), or not.
// Importantly, an element can be evicted if and only if it's not
// referenced. In particular, when an element becomes referenced, it's
// temporarily taken out of the LRU list until all references to it
// are dropped.
// (V) Visible: An element can visible for lookups (IS_VISIBLE set), or not.
// Initially, every element is visible. An element that is not visible is
// called a ghost.
// These properties induce 4 different states, with transitions defined as
// follows:
// - V --> not V: When a visible element is deleted or replaced by a new
// version.
// - Not V --> V: This cannot happen. A ghost remains in that state until it's
// not referenced any more, at which point it's ready to be removed from the
// hash table. (A ghost simply waits to transition to the afterlife---it will
// never be visible again.)
// - R --> not R: When all references to an element are dropped.
// - Not R --> R: When an unreferenced element becomes referenced. This can only
// happen if the element is V, since references to an element can only be
// created when it's visible.
// Internally, the cache uses an open-addressed hash table to index the handles.
// We use tombstone counters to keep track of displacements.
// Because of the tombstones and the two possible visibility states of an
// element, the table slots can be in 4 different states:
// 1. Visible element (IS_ELEMENT set and IS_VISIBLE set): The slot contains a
// key-value element.
// 2. Ghost element (IS_ELEMENT set and IS_VISIBLE unset): The slot contains an
// element that has been removed, but it's still referenced. It's invisible
// to lookups.
// 3. Tombstone (IS_ELEMENT unset and displacements > 0): The slot contains a
// tombstone.
// 4. Empty (IS_ELEMENT unset and displacements == 0): The slot is unused.
// A slot that is an element can further have IS_VISIBLE set or not.
// When a ghost is removed from the table, it can either transition to being a
// tombstone or an empty slot, depending on the number of displacements of the
// slot. In any case, the slot becomes available. When a handle is inserted
// into that slot, it becomes a visible element again.
constexpr uint8_t kCacheKeySize =
static_cast<uint8_t>(sizeof(ROCKSDB_NAMESPACE::CacheKey));
// The load factor p is a real number in (0, 1) such that at all
// times at most a fraction p of all slots, without counting tombstones,
// are occupied by elements. This means that the probability that a
// random probe hits an empty slot is at most p, and thus at most 1/p probes
// are required on average. We use p = 70%, so between 1 and 2 probes are
// needed on average.
// Because the size of the hash table is always rounded up to the next
// power of 2, p is really an upper bound on the actual load factor---the
// actual load factor is anywhere between p/2 and p. This is a bit wasteful,
// but bear in mind that slots only hold metadata, not actual values.
// Since space cost is dominated by the values (the LSM blocks),
// overprovisioning the table with metadata only increases the total cache space
// usage by a tiny fraction.
constexpr double kLoadFactor = 0.7;
// Arbitrary seeds.
constexpr uint32_t kProbingSeed1 = 0xbc9f1d34;
constexpr uint32_t kProbingSeed2 = 0x7a2bb9d5;
// An experimental (under development!) alternative to LRUCache
struct LRUHandle {
void* value;
Cache::DeleterFn deleter;
LRUHandle* next;
LRUHandle* prev;
size_t total_charge; // TODO(opt): Only allow uint32_t?
// The hash of key(). Used for fast sharding and comparisons.
uint32_t hash;
// The number of external refs to this entry. The cache itself is not counted.
uint32_t refs;
enum Flags : uint8_t {
// Whether the handle is visible to Lookups.
IS_VISIBLE = (1 << 0),
// Whether the slot is in use by an element.
IS_ELEMENT = (1 << 1),
};
uint8_t flags;
// The number of elements that hash to this slot or a lower one,
// but wind up in a higher slot.
uint32_t displacements;
std::array<char, kCacheKeySize> key_data;
LRUHandle() {
value = nullptr;
deleter = nullptr;
next = nullptr;
prev = nullptr;
total_charge = 0;
hash = 0;
refs = 0;
flags = 0;
displacements = 0;
key_data.fill(0);
}
Slice key() const { return Slice(key_data.data(), kCacheKeySize); }
// Increase the reference count by 1.
void Ref() { refs++; }
// Just reduce the reference count by 1. Return true if it was last reference.
bool Unref() {
assert(refs > 0);
refs--;
return refs == 0;
}
// Return true if there are external refs, false otherwise.
bool HasRefs() const { return refs > 0; }
bool IsVisible() const { return flags & IS_VISIBLE; }
void SetIsVisible(bool is_visible) {
if (is_visible) {
flags |= IS_VISIBLE;
} else {
flags &= ~IS_VISIBLE;
}
}
bool IsElement() const { return flags & IS_ELEMENT; }
void SetIsElement(bool is_element) {
if (is_element) {
flags |= IS_ELEMENT;
} else {
flags &= ~IS_ELEMENT;
}
}
void FreeData() {
assert(refs == 0);
if (deleter) {
(*deleter)(key(), value);
}
}
// Calculate the memory usage by metadata.
inline size_t CalcMetaCharge(
CacheMetadataChargePolicy metadata_charge_policy) const {
if (metadata_charge_policy != kFullChargeCacheMetadata) {
return 0;
} else {
// #ifdef ROCKSDB_MALLOC_USABLE_SIZE
// return malloc_usable_size(
// const_cast<void*>(static_cast<const void*>(this)));
// #else
// TODO(Guido) malloc_usable_size only works when we call it on
// a pointer allocated with malloc. Because our handles are all
// allocated in a single shot as an array, the user can't call
// CalcMetaCharge (or CalcTotalCharge or GetCharge) on a handle
// pointer returned by the cache. Moreover, malloc_usable_size
// expects a heap-allocated handle, but sometimes in our code we
// wish to pass a stack-allocated handle (this is only a performance
// concern).
// What is the right way to compute metadata charges with pre-allocated
// handles?
return sizeof(LRUHandle);
// #endif
}
}
inline void CalcTotalCharge(
size_t charge, CacheMetadataChargePolicy metadata_charge_policy) {
total_charge = charge + CalcMetaCharge(metadata_charge_policy);
}
inline size_t GetCharge(
CacheMetadataChargePolicy metadata_charge_policy) const {
size_t meta_charge = CalcMetaCharge(metadata_charge_policy);
assert(total_charge >= meta_charge);
return total_charge - meta_charge;
}
inline bool IsEmpty() {
return !this->IsElement() && this->displacements == 0;
}
inline bool IsTombstone() {
return !this->IsElement() && this->displacements > 0;
}
inline bool Matches(const Slice& some_key, uint32_t some_hash) {
return this->IsElement() && this->hash == some_hash &&
this->key() == some_key;
}
};
// TODO(Guido) Update the following comment.
// We provide our own simple hash table since it removes a whole bunch
// of porting hacks and is also faster than some of the built-in hash
// table implementations in some of the compiler/runtime combinations
// we have tested. E.g., readrandom speeds up by ~5% over the g++
// 4.4.3's builtin hashtable.
class LRUHandleTable {
public:
explicit LRUHandleTable(uint8_t hash_bits);
~LRUHandleTable();
// Returns a pointer to a visible element matching the key/hash, or
// nullptr if not present.
LRUHandle* Lookup(const Slice& key, uint32_t hash);
// Inserts a copy of h into the hash table.
// Returns a pointer to the inserted handle, or nullptr if no slot
// available was found. If an existing visible element matching the
// key/hash is already present in the hash table, the argument old
// is set to pointe to it; otherwise, it's set to nullptr.
LRUHandle* Insert(LRUHandle* h, LRUHandle** old);
// Removes h from the hash table. The handle must already be off
// the LRU list.
void Remove(LRUHandle* h);
// Turns a visible element h into a ghost (i.e., not visible).
void Exclude(LRUHandle* h);
// Assigns a copy of h to the given slot.
void Assign(int slot, LRUHandle* h);
template <typename T>
void ApplyToEntriesRange(T func, uint32_t index_begin, uint32_t index_end) {
for (uint32_t i = index_begin; i < index_end; i++) {
LRUHandle* h = &array_[i];
if (h->IsVisible()) {
func(h);
}
}
}
uint8_t GetLengthBits() const { return length_bits_; }
uint32_t GetOccupancy() const { return occupancy_; }
private:
int FindVisibleElement(const Slice& key, uint32_t hash, int& probe,
int displacement);
int FindAvailableSlot(const Slice& key, int& probe, int displacement);
int FindVisibleElementOrAvailableSlot(const Slice& key, uint32_t hash,
int& probe, int displacement);
// Returns the index of the first slot probed (hashing with
// the given key) with a handle e such that cond(e) is true.
// Otherwise, if no match is found, returns -1.
// For every handle e probed except the final slot, updates
// e->displacements += displacement.
// The argument probe is modified such that consecutive calls
// to FindSlot continue probing right after where the previous
// call left.
int FindSlot(const Slice& key, std::function<bool(LRUHandle*)> cond,
int& probe, int displacement);
// Number of hash bits used for table index.
// The size of the table is 1 << length_bits_.
uint8_t length_bits_;
// Number of elements in the table.
uint32_t occupancy_;
std::unique_ptr<LRUHandle[]> array_;
};
// A single shard of sharded cache.
class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
public:
LRUCacheShard(size_t capacity, size_t estimated_value_size,
bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy);
~LRUCacheShard() override = default;
// Separate from constructor so caller can easily make an array of LRUCache
// if current usage is more than new capacity, the function will attempt to
// free the needed space.
void SetCapacity(size_t capacity) override;
// Set the flag to reject insertion if cache if full.
void SetStrictCapacityLimit(bool strict_capacity_limit) override;
// Like Cache methods, but with an extra "hash" parameter.
// Insert an item into the hash table and, if handle is null, insert into
// the LRU list. Older items are evicted as necessary. If the cache is full
// and free_handle_on_fail is true, the item is deleted and handle is set to
// nullptr.
Status Insert(const Slice& key, uint32_t hash, void* value, size_t charge,
Cache::DeleterFn deleter, Cache::Handle** handle,
Cache::Priority priority) override;
Status Insert(const Slice& key, uint32_t hash, void* value,
const Cache::CacheItemHelper* helper, size_t charge,
Cache::Handle** handle, Cache::Priority priority) override {
return Insert(key, hash, value, charge, helper->del_cb, handle, priority);
}
Cache::Handle* Lookup(const Slice& key, uint32_t hash,
const Cache::CacheItemHelper* /*helper*/,
const Cache::CreateCallback& /*create_cb*/,
Cache::Priority /*priority*/, bool /*wait*/,
Statistics* /*stats*/) override {
return Lookup(key, hash);
}
Cache::Handle* Lookup(const Slice& key, uint32_t hash) override;
bool Release(Cache::Handle* handle, bool /*useful*/,
bool erase_if_last_ref) override {
return Release(handle, erase_if_last_ref);
}
bool IsReady(Cache::Handle* /*handle*/) override { return true; }
void Wait(Cache::Handle* /*handle*/) override {}
bool Ref(Cache::Handle* handle) override;
bool Release(Cache::Handle* handle, bool erase_if_last_ref = false) override;
void Erase(const Slice& key, uint32_t hash) override;
size_t GetUsage() const override;
size_t GetPinnedUsage() const override;
void ApplyToSomeEntries(
const std::function<void(const Slice& key, void* value, size_t charge,
DeleterFn deleter)>& callback,
uint32_t average_entries_per_lock, uint32_t* state) override;
void EraseUnRefEntries() override;
std::string GetPrintableOptions() const override;
private:
friend class LRUCache;
void LRU_Remove(LRUHandle* e);
void LRU_Insert(LRUHandle* e);
// Free some space following strict LRU policy until enough space
// to hold (usage_ + charge) is freed or the lru list is empty
// This function is not thread safe - it needs to be executed while
// holding the mutex_.
void EvictFromLRU(size_t charge, autovector<LRUHandle>* deleted);
// Returns the number of bits used to hash an element in the hash
// table.
static uint8_t CalcHashBits(size_t capacity, size_t estimated_value_size,
CacheMetadataChargePolicy metadata_charge_policy);
// Initialized before use.
size_t capacity_;
// Whether to reject insertion if cache reaches its full capacity.
bool strict_capacity_limit_;
// Dummy head of LRU list.
// lru.prev is newest entry, lru.next is oldest entry.
// LRU contains items which can be evicted, ie reference only by cache
LRUHandle lru_;
// Pointer to head of low-pri pool in LRU list.
LRUHandle* lru_low_pri_;
// ------------^^^^^^^^^^^^^-----------
// Not frequently modified data members
// ------------------------------------
//
// We separate data members that are updated frequently from the ones that
// are not frequently updated so that they don't share the same cache line
// which will lead into false cache sharing
//
// ------------------------------------
// Frequently modified data members
// ------------vvvvvvvvvvvvv-----------
LRUHandleTable table_;
// Memory size for entries residing in the cache.
size_t usage_;
// Memory size for entries residing only in the LRU list.
size_t lru_usage_;
// mutex_ protects the following state.
// We don't count mutex_ as the cache's internal state so semantically we
// don't mind mutex_ invoking the non-const actions.
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
mutable DMutex mutex_;
};
class LRUCache
#ifdef NDEBUG
final
#endif
: public ShardedCache {
public:
LRUCache(size_t capacity, size_t estimated_value_size, int num_shard_bits,
bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy =
kDontChargeCacheMetadata);
~LRUCache() override;
const char* Name() const override { return "LRUCache"; }
CacheShard* GetShard(uint32_t shard) override;
const CacheShard* GetShard(uint32_t shard) const override;
void* Value(Handle* handle) override;
size_t GetCharge(Handle* handle) const override;
uint32_t GetHash(Handle* handle) const override;
DeleterFn GetDeleter(Handle* handle) const override;
void DisownData() override;
private:
LRUCacheShard* shards_ = nullptr;
int num_shards_ = 0;
};
} // namespace fast_lru_cache
std::shared_ptr<Cache> NewFastLRUCache(
size_t capacity, size_t estimated_value_size, int num_shard_bits,
bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy);
} // namespace ROCKSDB_NAMESPACE