2015-11-24 22:01:09 +01:00
|
|
|
// Copyright (c) 2013, Facebook, Inc. All rights reserved.
|
|
|
|
// This source code is licensed under the BSD-style license found in the
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// LICENSE file in the root directory of this source tree. An additional
|
|
|
|
// grant of patent rights can be found in the PATENTS file in the same
|
|
|
|
// directory.
|
2015-11-24 22:01:09 +01:00
|
|
|
//
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// 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.
|
|
|
|
//
|
|
|
|
// InlineSkipList is derived from SkipList (skiplist.h), but it optimizes
|
|
|
|
// the memory layout by requiring that the key storage be allocated through
|
|
|
|
// the skip list instance. For the common case of SkipList<const char*,
|
|
|
|
// Cmp> this saves 1 pointer per skip list node and gives better cache
|
|
|
|
// locality, at the expense of wasted padding from using AllocateAligned
|
|
|
|
// instead of Allocate for the keys. The unused padding will be from
|
|
|
|
// 0 to sizeof(void*)-1 bytes, and the space savings are sizeof(void*)
|
|
|
|
// bytes, so despite the padding the space used is always less than
|
|
|
|
// SkipList<const char*, ..>.
|
2015-11-24 22:01:09 +01:00
|
|
|
//
|
|
|
|
// Thread safety -------------
|
|
|
|
//
|
|
|
|
// Writes require external synchronization, most likely a mutex. Reads
|
|
|
|
// require a guarantee that the InlineSkipList will not be destroyed while
|
|
|
|
// the read is in progress. Apart from that, reads progress without any
|
|
|
|
// internal locking or synchronization.
|
|
|
|
//
|
|
|
|
// Invariants:
|
|
|
|
//
|
|
|
|
// (1) Allocated nodes are never deleted until the InlineSkipList is
|
|
|
|
// destroyed. This is trivially guaranteed by the code since we never
|
|
|
|
// delete any skip list nodes.
|
|
|
|
//
|
|
|
|
// (2) The contents of a Node except for the next/prev pointers are
|
|
|
|
// immutable after the Node has been linked into the InlineSkipList.
|
|
|
|
// Only Insert() modifies the list, and it is careful to initialize a
|
|
|
|
// node and use release-stores to publish the nodes in one or more lists.
|
|
|
|
//
|
|
|
|
// ... prev vs. next pointer ordering ...
|
|
|
|
//
|
|
|
|
|
|
|
|
#pragma once
|
|
|
|
#include <assert.h>
|
|
|
|
#include <stdlib.h>
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
#include <atomic>
|
2015-11-24 22:01:09 +01:00
|
|
|
#include "port/port.h"
|
|
|
|
#include "util/allocator.h"
|
|
|
|
#include "util/random.h"
|
|
|
|
|
|
|
|
namespace rocksdb {
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
2015-11-24 22:01:09 +01:00
|
|
|
class InlineSkipList {
|
|
|
|
private:
|
|
|
|
struct Node;
|
|
|
|
|
|
|
|
public:
|
|
|
|
// Create a new InlineSkipList object that will use "cmp" for comparing
|
|
|
|
// keys, and will allocate memory using "*allocator". Objects allocated
|
|
|
|
// in the allocator must remain allocated for the lifetime of the
|
|
|
|
// skiplist object.
|
|
|
|
explicit InlineSkipList(Comparator cmp, Allocator* allocator,
|
|
|
|
int32_t max_height = 12,
|
|
|
|
int32_t branching_factor = 4);
|
|
|
|
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// Allocates a key and a skip-list node, returning a pointer to the
|
|
|
|
// key portion of the node.
|
2015-11-24 22:29:50 +01:00
|
|
|
char* AllocateKey(size_t key_size);
|
|
|
|
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// Inserts a key allocated by AllocateKey, after the actual key value
|
|
|
|
// has been filled in.
|
|
|
|
//
|
2015-11-24 22:01:09 +01:00
|
|
|
// REQUIRES: nothing that compares equal to key is currently in the list.
|
2015-11-24 22:29:50 +01:00
|
|
|
void Insert(const char* key);
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Returns true iff an entry that compares equal to key is in the list.
|
2015-11-24 22:29:50 +01:00
|
|
|
bool Contains(const char* key) const;
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Return estimated number of entries smaller than `key`.
|
2015-11-24 22:29:50 +01:00
|
|
|
uint64_t EstimateCount(const char* key) const;
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Iteration over the contents of a skip list
|
|
|
|
class Iterator {
|
|
|
|
public:
|
|
|
|
// Initialize an iterator over the specified list.
|
|
|
|
// The returned iterator is not valid.
|
|
|
|
explicit Iterator(const InlineSkipList* list);
|
|
|
|
|
|
|
|
// Change the underlying skiplist used for this iterator
|
|
|
|
// This enables us not changing the iterator without deallocating
|
|
|
|
// an old one and then allocating a new one
|
|
|
|
void SetList(const InlineSkipList* list);
|
|
|
|
|
|
|
|
// Returns true iff the iterator is positioned at a valid node.
|
|
|
|
bool Valid() const;
|
|
|
|
|
|
|
|
// Returns the key at the current position.
|
|
|
|
// REQUIRES: Valid()
|
2015-11-24 22:29:50 +01:00
|
|
|
const char* key() const;
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Advances to the next position.
|
|
|
|
// REQUIRES: Valid()
|
|
|
|
void Next();
|
|
|
|
|
|
|
|
// Advances to the previous position.
|
|
|
|
// REQUIRES: Valid()
|
|
|
|
void Prev();
|
|
|
|
|
|
|
|
// Advance to the first entry with a key >= target
|
2015-11-24 22:29:50 +01:00
|
|
|
void Seek(const char* target);
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Position at the first entry in list.
|
|
|
|
// Final state of iterator is Valid() iff list is not empty.
|
|
|
|
void SeekToFirst();
|
|
|
|
|
|
|
|
// Position at the last entry in list.
|
|
|
|
// Final state of iterator is Valid() iff list is not empty.
|
|
|
|
void SeekToLast();
|
|
|
|
|
|
|
|
private:
|
|
|
|
const InlineSkipList* list_;
|
|
|
|
Node* node_;
|
|
|
|
// Intentionally copyable
|
|
|
|
};
|
|
|
|
|
|
|
|
private:
|
|
|
|
const uint16_t kMaxHeight_;
|
|
|
|
const uint16_t kBranching_;
|
|
|
|
const uint32_t kScaledInverseBranching_;
|
|
|
|
|
|
|
|
// Immutable after construction
|
|
|
|
Comparator const compare_;
|
|
|
|
Allocator* const allocator_; // Allocator used for allocations of nodes
|
|
|
|
|
|
|
|
Node* const head_;
|
|
|
|
|
|
|
|
// Modified only by Insert(). Read racily by readers, but stale
|
|
|
|
// values are ok.
|
|
|
|
std::atomic<int> max_height_; // Height of the entire list
|
|
|
|
|
|
|
|
// Used for optimizing sequential insert patterns. Tricky. prev_[i] for
|
|
|
|
// i up to max_height_ is the predecessor of prev_[0] and prev_height_
|
|
|
|
// is the height of prev_[0]. prev_[0] can only be equal to head before
|
|
|
|
// insertion, in which case max_height_ and prev_height_ are 1.
|
|
|
|
Node** prev_;
|
|
|
|
int32_t prev_height_;
|
|
|
|
|
|
|
|
inline int GetMaxHeight() const {
|
|
|
|
return max_height_.load(std::memory_order_relaxed);
|
|
|
|
}
|
|
|
|
|
|
|
|
int RandomHeight();
|
2015-11-24 22:29:50 +01:00
|
|
|
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
Node* AllocateNode(size_t key_size, int height);
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
bool Equal(const char* a, const char* b) const {
|
|
|
|
return (compare_(a, b) == 0);
|
|
|
|
}
|
2015-11-24 22:01:09 +01:00
|
|
|
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// Return true if key is greater than the data stored in "n". Null n
|
|
|
|
// is considered infinite.
|
2015-11-24 22:29:50 +01:00
|
|
|
bool KeyIsAfterNode(const char* key, Node* n) const;
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Returns the earliest node with a key >= key.
|
|
|
|
// Return nullptr if there is no such node.
|
2015-11-24 22:29:50 +01:00
|
|
|
Node* FindGreaterOrEqual(const char* key) const;
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Return the latest node with a key < key.
|
|
|
|
// Return head_ if there is no such node.
|
|
|
|
// Fills prev[level] with pointer to previous node at "level" for every
|
|
|
|
// level in [0..max_height_-1], if prev is non-null.
|
2015-11-24 22:29:50 +01:00
|
|
|
Node* FindLessThan(const char* key, Node** prev = nullptr) const;
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
// Return the last node in the list.
|
|
|
|
// Return head_ if list is empty.
|
|
|
|
Node* FindLast() const;
|
|
|
|
|
|
|
|
// No copying allowed
|
|
|
|
InlineSkipList(const InlineSkipList&);
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
InlineSkipList& operator=(const InlineSkipList&);
|
2015-11-24 22:01:09 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
// Implementation details follow
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
|
|
|
|
// The Node data type is more of a pointer into custom-managed memory than
|
|
|
|
// a traditional C++ struct. The key is stored in the bytes immediately
|
|
|
|
// after the struct, and the next_ pointers for nodes with height > 1 are
|
|
|
|
// stored immediately _before_ the struct. This avoids the need to include
|
|
|
|
// any pointer or sizing data, which reduces per-node memory overheads.
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
struct InlineSkipList<Comparator>::Node {
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// Stores the height of the node in the memory location normally used for
|
|
|
|
// next_[0]. This is used for passing data from AllocateKey to Insert.
|
|
|
|
void StashHeight(const int height) {
|
|
|
|
assert(sizeof(int) <= sizeof(next_[0]));
|
|
|
|
memcpy(&next_[0], &height, sizeof(int));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Retrieves the value passed to StashHeight. Undefined after a call
|
|
|
|
// to SetNext or NoBarrier_SetNext.
|
|
|
|
int UnstashHeight() const {
|
|
|
|
int rv;
|
|
|
|
memcpy(&rv, &next_[0], sizeof(int));
|
|
|
|
return rv;
|
|
|
|
}
|
2015-11-24 22:01:09 +01:00
|
|
|
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
const char* Key() const { return reinterpret_cast<const char*>(&next_[1]); }
|
2015-11-24 22:01:09 +01:00
|
|
|
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// Accessors/mutators for links. Wrapped in methods so we can add
|
|
|
|
// the appropriate barriers as necessary, and perform the necessary
|
|
|
|
// addressing trickery for storing links below the Node in memory.
|
2015-11-24 22:01:09 +01:00
|
|
|
Node* Next(int n) {
|
|
|
|
assert(n >= 0);
|
|
|
|
// Use an 'acquire load' so that we observe a fully initialized
|
|
|
|
// version of the returned Node.
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
return (next_[-n].load(std::memory_order_acquire));
|
2015-11-24 22:01:09 +01:00
|
|
|
}
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
|
2015-11-24 22:01:09 +01:00
|
|
|
void SetNext(int n, Node* x) {
|
|
|
|
assert(n >= 0);
|
|
|
|
// Use a 'release store' so that anybody who reads through this
|
|
|
|
// pointer observes a fully initialized version of the inserted node.
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
next_[-n].store(x, std::memory_order_release);
|
2015-11-24 22:01:09 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// No-barrier variants that can be safely used in a few locations.
|
|
|
|
Node* NoBarrier_Next(int n) {
|
|
|
|
assert(n >= 0);
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
return next_[-n].load(std::memory_order_relaxed);
|
2015-11-24 22:01:09 +01:00
|
|
|
}
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
|
2015-11-24 22:01:09 +01:00
|
|
|
void NoBarrier_SetNext(int n, Node* x) {
|
|
|
|
assert(n >= 0);
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
next_[-n].store(x, std::memory_order_relaxed);
|
2015-11-24 22:01:09 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
// next_[0] is the lowest level link (level 0). Higher levels are
|
|
|
|
// stored _earlier_, so level 1 is at next_[-1].
|
2015-11-24 22:01:09 +01:00
|
|
|
std::atomic<Node*> next_[1];
|
|
|
|
};
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline InlineSkipList<Comparator>::Iterator::Iterator(
|
2015-11-24 22:01:09 +01:00
|
|
|
const InlineSkipList* list) {
|
|
|
|
SetList(list);
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline void InlineSkipList<Comparator>::Iterator::SetList(
|
2015-11-24 22:01:09 +01:00
|
|
|
const InlineSkipList* list) {
|
|
|
|
list_ = list;
|
|
|
|
node_ = nullptr;
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline bool InlineSkipList<Comparator>::Iterator::Valid() const {
|
2015-11-24 22:01:09 +01:00
|
|
|
return node_ != nullptr;
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline const char* InlineSkipList<Comparator>::Iterator::key() const {
|
2015-11-24 22:01:09 +01:00
|
|
|
assert(Valid());
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
return node_->Key();
|
2015-11-24 22:01:09 +01:00
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline void InlineSkipList<Comparator>::Iterator::Next() {
|
2015-11-24 22:01:09 +01:00
|
|
|
assert(Valid());
|
|
|
|
node_ = node_->Next(0);
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline void InlineSkipList<Comparator>::Iterator::Prev() {
|
2015-11-24 22:01:09 +01:00
|
|
|
// Instead of using explicit "prev" links, we just search for the
|
|
|
|
// last node that falls before key.
|
|
|
|
assert(Valid());
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
node_ = list_->FindLessThan(node_->Key());
|
2015-11-24 22:01:09 +01:00
|
|
|
if (node_ == list_->head_) {
|
|
|
|
node_ = nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline void InlineSkipList<Comparator>::Iterator::Seek(const char* target) {
|
2015-11-24 22:01:09 +01:00
|
|
|
node_ = list_->FindGreaterOrEqual(target);
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline void InlineSkipList<Comparator>::Iterator::SeekToFirst() {
|
2015-11-24 22:01:09 +01:00
|
|
|
node_ = list_->head_->Next(0);
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
inline void InlineSkipList<Comparator>::Iterator::SeekToLast() {
|
2015-11-24 22:01:09 +01:00
|
|
|
node_ = list_->FindLast();
|
|
|
|
if (node_ == list_->head_) {
|
|
|
|
node_ = nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
int InlineSkipList<Comparator>::RandomHeight() {
|
2015-11-24 22:01:09 +01:00
|
|
|
auto rnd = Random::GetTLSInstance();
|
|
|
|
|
|
|
|
// Increase height with probability 1 in kBranching
|
|
|
|
int height = 1;
|
|
|
|
while (height < kMaxHeight_ && rnd->Next() < kScaledInverseBranching_) {
|
|
|
|
height++;
|
|
|
|
}
|
|
|
|
assert(height > 0);
|
|
|
|
assert(height <= kMaxHeight_);
|
|
|
|
return height;
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
bool InlineSkipList<Comparator>::KeyIsAfterNode(const char* key,
|
|
|
|
Node* n) const {
|
2015-11-24 22:01:09 +01:00
|
|
|
// nullptr n is considered infinite
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
return (n != nullptr) && (compare_(n->Key(), key) < 0);
|
2015-11-24 22:01:09 +01:00
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
typename InlineSkipList<Comparator>::Node*
|
|
|
|
InlineSkipList<Comparator>::FindGreaterOrEqual(const char* key) const {
|
2015-11-24 22:01:09 +01:00
|
|
|
// Note: It looks like we could reduce duplication by implementing
|
|
|
|
// this function as FindLessThan(key)->Next(0), but we wouldn't be able
|
|
|
|
// to exit early on equality and the result wouldn't even be correct.
|
|
|
|
// A concurrent insert might occur after FindLessThan(key) but before
|
|
|
|
// we get a chance to call Next(0).
|
|
|
|
Node* x = head_;
|
|
|
|
int level = GetMaxHeight() - 1;
|
|
|
|
Node* last_bigger = nullptr;
|
|
|
|
while (true) {
|
|
|
|
Node* next = x->Next(level);
|
|
|
|
// Make sure the lists are sorted
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
assert(x == head_ || next == nullptr || KeyIsAfterNode(next->Key(), x));
|
2015-11-24 22:01:09 +01:00
|
|
|
// Make sure we haven't overshot during our search
|
|
|
|
assert(x == head_ || KeyIsAfterNode(key, x));
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
int cmp = (next == nullptr || next == last_bigger)
|
|
|
|
? 1
|
|
|
|
: compare_(next->Key(), key);
|
2015-11-24 22:01:09 +01:00
|
|
|
if (cmp == 0 || (cmp > 0 && level == 0)) {
|
|
|
|
return next;
|
|
|
|
} else if (cmp < 0) {
|
|
|
|
// Keep searching in this list
|
|
|
|
x = next;
|
|
|
|
} else {
|
|
|
|
// Switch to next list, reuse compare_() result
|
|
|
|
last_bigger = next;
|
|
|
|
level--;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
typename InlineSkipList<Comparator>::Node*
|
|
|
|
InlineSkipList<Comparator>::FindLessThan(const char* key, Node** prev) const {
|
2015-11-24 22:01:09 +01:00
|
|
|
Node* x = head_;
|
|
|
|
int level = GetMaxHeight() - 1;
|
|
|
|
// KeyIsAfter(key, last_not_after) is definitely false
|
|
|
|
Node* last_not_after = nullptr;
|
|
|
|
while (true) {
|
|
|
|
Node* next = x->Next(level);
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
assert(x == head_ || next == nullptr || KeyIsAfterNode(next->Key(), x));
|
2015-11-24 22:01:09 +01:00
|
|
|
assert(x == head_ || KeyIsAfterNode(key, x));
|
|
|
|
if (next != last_not_after && KeyIsAfterNode(key, next)) {
|
|
|
|
// Keep searching in this list
|
|
|
|
x = next;
|
|
|
|
} else {
|
|
|
|
if (prev != nullptr) {
|
|
|
|
prev[level] = x;
|
|
|
|
}
|
|
|
|
if (level == 0) {
|
|
|
|
return x;
|
|
|
|
} else {
|
|
|
|
// Switch to next list, reuse KeyIUsAfterNode() result
|
|
|
|
last_not_after = next;
|
|
|
|
level--;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
typename InlineSkipList<Comparator>::Node*
|
|
|
|
InlineSkipList<Comparator>::FindLast() const {
|
2015-11-24 22:01:09 +01:00
|
|
|
Node* x = head_;
|
|
|
|
int level = GetMaxHeight() - 1;
|
|
|
|
while (true) {
|
|
|
|
Node* next = x->Next(level);
|
|
|
|
if (next == nullptr) {
|
|
|
|
if (level == 0) {
|
|
|
|
return x;
|
|
|
|
} else {
|
|
|
|
// Switch to next list
|
|
|
|
level--;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
x = next;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
uint64_t InlineSkipList<Comparator>::EstimateCount(const char* key) const {
|
2015-11-24 22:01:09 +01:00
|
|
|
uint64_t count = 0;
|
|
|
|
|
|
|
|
Node* x = head_;
|
|
|
|
int level = GetMaxHeight() - 1;
|
|
|
|
while (true) {
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
assert(x == head_ || compare_(x->Key(), key) < 0);
|
2015-11-24 22:01:09 +01:00
|
|
|
Node* next = x->Next(level);
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
if (next == nullptr || compare_(next->Key(), key) >= 0) {
|
2015-11-24 22:01:09 +01:00
|
|
|
if (level == 0) {
|
|
|
|
return count;
|
|
|
|
} else {
|
|
|
|
// Switch to next list
|
|
|
|
count *= kBranching_;
|
|
|
|
level--;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
x = next;
|
|
|
|
count++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
InlineSkipList<Comparator>::InlineSkipList(const Comparator cmp,
|
|
|
|
Allocator* allocator,
|
|
|
|
int32_t max_height,
|
|
|
|
int32_t branching_factor)
|
2015-11-24 22:01:09 +01:00
|
|
|
: kMaxHeight_(max_height),
|
|
|
|
kBranching_(branching_factor),
|
|
|
|
kScaledInverseBranching_((Random::kMaxNext + 1) / kBranching_),
|
|
|
|
compare_(cmp),
|
|
|
|
allocator_(allocator),
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
head_(AllocateNode(0, max_height)),
|
2015-11-24 22:01:09 +01:00
|
|
|
max_height_(1),
|
|
|
|
prev_height_(1) {
|
|
|
|
assert(max_height > 0 && kMaxHeight_ == static_cast<uint32_t>(max_height));
|
|
|
|
assert(branching_factor > 0 &&
|
|
|
|
kBranching_ == static_cast<uint32_t>(branching_factor));
|
|
|
|
assert(kScaledInverseBranching_ > 0);
|
|
|
|
// Allocate the prev_ Node* array, directly from the passed-in allocator.
|
|
|
|
// prev_ does not need to be freed, as its life cycle is tied up with
|
|
|
|
// the allocator as a whole.
|
|
|
|
prev_ = reinterpret_cast<Node**>(
|
|
|
|
allocator_->AllocateAligned(sizeof(Node*) * kMaxHeight_));
|
|
|
|
for (int i = 0; i < kMaxHeight_; i++) {
|
|
|
|
head_->SetNext(i, nullptr);
|
|
|
|
prev_[i] = head_;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
char* InlineSkipList<Comparator>::AllocateKey(size_t key_size) {
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
return const_cast<char*>(AllocateNode(key_size, RandomHeight())->Key());
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class Comparator>
|
|
|
|
typename InlineSkipList<Comparator>::Node*
|
|
|
|
InlineSkipList<Comparator>::AllocateNode(size_t key_size, int height) {
|
|
|
|
auto prefix = sizeof(std::atomic<Node*>) * (height - 1);
|
|
|
|
|
|
|
|
// prefix is space for the height - 1 pointers that we store before
|
|
|
|
// the Node instance (next_[-(height - 1) .. -1]). Node starts at
|
|
|
|
// raw + prefix, and holds the bottom-mode (level 0) skip list pointer
|
|
|
|
// next_[0]. key_size is the bytes for the key, which comes just after
|
|
|
|
// the Node.
|
|
|
|
char* raw = allocator_->AllocateAligned(prefix + sizeof(Node) + key_size);
|
|
|
|
Node* x = reinterpret_cast<Node*>(raw + prefix);
|
|
|
|
|
|
|
|
// Once we've linked the node into the skip list we don't actually need
|
|
|
|
// to know its height, because we can implicitly use the fact that we
|
|
|
|
// traversed into a node at level h to known that h is a valid level
|
|
|
|
// for that node. We need to convey the height to the Insert step,
|
|
|
|
// however, so that it can perform the proper links. Since we're not
|
|
|
|
// using the pointers at the moment, StashHeight temporarily borrow
|
|
|
|
// storage from next_[0] for that purpose.
|
|
|
|
x->StashHeight(height);
|
|
|
|
return x;
|
2015-11-24 22:29:50 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
template <class Comparator>
|
|
|
|
void InlineSkipList<Comparator>::Insert(const char* key) {
|
2015-11-24 22:01:09 +01:00
|
|
|
// fast path for sequential insertion
|
|
|
|
if (!KeyIsAfterNode(key, prev_[0]->NoBarrier_Next(0)) &&
|
|
|
|
(prev_[0] == head_ || KeyIsAfterNode(key, prev_[0]))) {
|
|
|
|
assert(prev_[0] != head_ || (prev_height_ == 1 && GetMaxHeight() == 1));
|
|
|
|
|
|
|
|
// Outside of this method prev_[1..max_height_] is the predecessor
|
|
|
|
// of prev_[0], and prev_height_ refers to prev_[0]. Inside Insert
|
|
|
|
// prev_[0..max_height - 1] is the predecessor of key. Switch from
|
|
|
|
// the external state to the internal
|
|
|
|
for (int i = 1; i < prev_height_; i++) {
|
|
|
|
prev_[i] = prev_[0];
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// TODO(opt): we could use a NoBarrier predecessor search as an
|
|
|
|
// optimization for architectures where memory_order_acquire needs
|
|
|
|
// a synchronization instruction. Doesn't matter on x86
|
|
|
|
FindLessThan(key, prev_);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Our data structure does not allow duplicate insertion
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
assert(prev_[0]->Next(0) == nullptr || !Equal(key, prev_[0]->Next(0)->Key()));
|
|
|
|
|
|
|
|
// Find the Node that we placed before the key in AllocateKey
|
|
|
|
Node* x = reinterpret_cast<Node*>(const_cast<char*>(key)) - 1;
|
|
|
|
int height = x->UnstashHeight();
|
|
|
|
assert(height >= 1 && height <= kMaxHeight_);
|
2015-11-24 22:01:09 +01:00
|
|
|
|
|
|
|
if (height > GetMaxHeight()) {
|
|
|
|
for (int i = GetMaxHeight(); i < height; i++) {
|
|
|
|
prev_[i] = head_;
|
|
|
|
}
|
|
|
|
|
|
|
|
// It is ok to mutate max_height_ without any synchronization
|
|
|
|
// with concurrent readers. A concurrent reader that observes
|
|
|
|
// the new value of max_height_ will see either the old value of
|
|
|
|
// new level pointers from head_ (nullptr), or a new value set in
|
|
|
|
// the loop below. In the former case the reader will
|
|
|
|
// immediately drop to the next level since nullptr sorts after all
|
|
|
|
// keys. In the latter case the reader will use the new node.
|
|
|
|
max_height_.store(height, std::memory_order_relaxed);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 0; i < height; i++) {
|
|
|
|
// NoBarrier_SetNext() suffices since we will add a barrier when
|
|
|
|
// we publish a pointer to "x" in prev[i].
|
|
|
|
x->NoBarrier_SetNext(i, prev_[i]->NoBarrier_Next(i));
|
|
|
|
prev_[i]->SetNext(i, x);
|
|
|
|
}
|
|
|
|
prev_[0] = x;
|
|
|
|
prev_height_ = height;
|
|
|
|
}
|
|
|
|
|
2015-11-24 22:29:50 +01:00
|
|
|
template <class Comparator>
|
|
|
|
bool InlineSkipList<Comparator>::Contains(const char* key) const {
|
2015-11-24 22:01:09 +01:00
|
|
|
Node* x = FindGreaterOrEqual(key);
|
InlineSkipList part 3/3 - new skiplist type that colocates key and node
Summary:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
2015-11-19 23:24:29 +01:00
|
|
|
if (x != nullptr && Equal(key, x->Key())) {
|
2015-11-24 22:01:09 +01:00
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace rocksdb
|