rocksdb/table/multiget_context.h
anand76 35cdd3e71e MultiGet async IO across multiple levels (#10535)
Summary:
This PR exploits parallelism in MultiGet across levels. It applies only to the coroutine version of MultiGet. Previously, MultiGet file reads from SST files in the same level were parallelized. With this PR, MultiGet batches with keys distributed across multiple levels are read in parallel. This is accomplished by splitting the keys not present in a level (determined by bloom filtering) into a separate batch, and processing the new batch in parallel with the original batch.

Pull Request resolved: https://github.com/facebook/rocksdb/pull/10535

Test Plan:
1. Ensure existing MultiGet unit tests pass, updating them as necessary
2. New unit tests - TODO
3. Run stress test - TODO

No noticeable regression (<1%) without async IO -
Without PR: `multireadrandom :       7.261 micros/op 1101724 ops/sec 60.007 seconds 66110936 operations;  571.6 MB/s (8168992 of 8168992 found)`
With PR: `multireadrandom :       7.305 micros/op 1095167 ops/sec 60.007 seconds 65717936 operations;  568.2 MB/s (8271992 of 8271992 found)`

For a fully cached DB, but with async IO option on, no regression observed (<1%) -
Without PR: `multireadrandom :       5.201 micros/op 1538027 ops/sec 60.005 seconds 92288936 operations;  797.9 MB/s (11540992 of 11540992 found) `
With PR: `multireadrandom :       5.249 micros/op 1524097 ops/sec 60.005 seconds 91452936 operations;  790.7 MB/s (11649992 of 11649992 found) `

Reviewed By: akankshamahajan15

Differential Revision: D38774009

Pulled By: anand1976

fbshipit-source-id: c955e259749f1c091590ade73105b3ee46cd0007
2022-08-19 16:52:52 -07:00

404 lines
13 KiB
C++

// 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).
#pragma once
#include <algorithm>
#include <array>
#include <string>
#include "db/dbformat.h"
#include "db/lookup_key.h"
#include "db/merge_context.h"
#include "rocksdb/env.h"
#include "rocksdb/statistics.h"
#include "rocksdb/types.h"
#include "util/async_file_reader.h"
#include "util/autovector.h"
#include "util/math.h"
#include "util/single_thread_executor.h"
namespace ROCKSDB_NAMESPACE {
class GetContext;
struct KeyContext {
const Slice* key;
LookupKey* lkey;
Slice ukey_with_ts;
Slice ukey_without_ts;
Slice ikey;
ColumnFamilyHandle* column_family;
Status* s;
MergeContext merge_context;
SequenceNumber max_covering_tombstone_seq;
bool key_exists;
bool is_blob_index;
void* cb_arg;
PinnableSlice* value;
std::string* timestamp;
GetContext* get_context;
KeyContext(ColumnFamilyHandle* col_family, const Slice& user_key,
PinnableSlice* val, std::string* ts, Status* stat)
: key(&user_key),
lkey(nullptr),
column_family(col_family),
s(stat),
max_covering_tombstone_seq(0),
key_exists(false),
is_blob_index(false),
cb_arg(nullptr),
value(val),
timestamp(ts),
get_context(nullptr) {}
KeyContext() = default;
};
// The MultiGetContext class is a container for the sorted list of keys that
// we need to lookup in a batch. Its main purpose is to make batch execution
// easier by allowing various stages of the MultiGet lookups to operate on
// subsets of keys, potentially non-contiguous. In order to accomplish this,
// it defines the following classes -
//
// MultiGetContext::Range
// MultiGetContext::Range::Iterator
// MultiGetContext::Range::IteratorWrapper
//
// Here is an example of how this can be used -
//
// {
// MultiGetContext ctx(...);
// MultiGetContext::Range range = ctx.GetMultiGetRange();
//
// // Iterate to determine some subset of the keys
// MultiGetContext::Range::Iterator start = range.begin();
// MultiGetContext::Range::Iterator end = ...;
//
// // Make a new range with a subset of keys
// MultiGetContext::Range subrange(range, start, end);
//
// // Define an auxillary vector, if needed, to hold additional data for
// // each key
// std::array<Foo, MultiGetContext::MAX_BATCH_SIZE> aux;
//
// // Iterate over the subrange and the auxillary vector simultaneously
// MultiGetContext::Range::Iterator iter = subrange.begin();
// for (; iter != subrange.end(); ++iter) {
// KeyContext& key = *iter;
// Foo& aux_key = aux_iter[iter.index()];
// ...
// }
// }
class MultiGetContext {
public:
// Limit the number of keys in a batch to this number. Benchmarks show that
// there is negligible benefit for batches exceeding this. Keeping this < 32
// simplifies iteration, as well as reduces the amount of stack allocations
// that need to be performed
static const int MAX_BATCH_SIZE = 32;
// A bitmask of at least MAX_BATCH_SIZE - 1 bits, so that
// Mask{1} << MAX_BATCH_SIZE is well defined
using Mask = uint64_t;
static_assert(MAX_BATCH_SIZE < sizeof(Mask) * 8);
MultiGetContext(autovector<KeyContext*, MAX_BATCH_SIZE>* sorted_keys,
size_t begin, size_t num_keys, SequenceNumber snapshot,
const ReadOptions& read_opts, FileSystem* fs,
Statistics* stats)
: num_keys_(num_keys),
value_mask_(0),
value_size_(0),
lookup_key_ptr_(reinterpret_cast<LookupKey*>(lookup_key_stack_buf))
#if USE_COROUTINES
,
reader_(fs, stats),
executor_(reader_)
#endif // USE_COROUTINES
{
(void)fs;
(void)stats;
assert(num_keys <= MAX_BATCH_SIZE);
if (num_keys > MAX_LOOKUP_KEYS_ON_STACK) {
lookup_key_heap_buf.reset(new char[sizeof(LookupKey) * num_keys]);
lookup_key_ptr_ = reinterpret_cast<LookupKey*>(
lookup_key_heap_buf.get());
}
for (size_t iter = 0; iter != num_keys_; ++iter) {
// autovector may not be contiguous storage, so make a copy
sorted_keys_[iter] = (*sorted_keys)[begin + iter];
sorted_keys_[iter]->lkey = new (&lookup_key_ptr_[iter])
LookupKey(*sorted_keys_[iter]->key, snapshot, read_opts.timestamp);
sorted_keys_[iter]->ukey_with_ts = sorted_keys_[iter]->lkey->user_key();
sorted_keys_[iter]->ukey_without_ts = StripTimestampFromUserKey(
sorted_keys_[iter]->lkey->user_key(),
read_opts.timestamp == nullptr ? 0 : read_opts.timestamp->size());
sorted_keys_[iter]->ikey = sorted_keys_[iter]->lkey->internal_key();
sorted_keys_[iter]->timestamp = (*sorted_keys)[begin + iter]->timestamp;
sorted_keys_[iter]->get_context =
(*sorted_keys)[begin + iter]->get_context;
}
}
~MultiGetContext() {
for (size_t i = 0; i < num_keys_; ++i) {
lookup_key_ptr_[i].~LookupKey();
}
}
#if USE_COROUTINES
SingleThreadExecutor& executor() { return executor_; }
AsyncFileReader& reader() { return reader_; }
#endif // USE_COROUTINES
private:
static const int MAX_LOOKUP_KEYS_ON_STACK = 16;
alignas(alignof(LookupKey))
char lookup_key_stack_buf[sizeof(LookupKey) * MAX_LOOKUP_KEYS_ON_STACK];
std::array<KeyContext*, MAX_BATCH_SIZE> sorted_keys_;
size_t num_keys_;
Mask value_mask_;
uint64_t value_size_;
std::unique_ptr<char[]> lookup_key_heap_buf;
LookupKey* lookup_key_ptr_;
#if USE_COROUTINES
AsyncFileReader reader_;
SingleThreadExecutor executor_;
#endif // USE_COROUTINES
public:
// MultiGetContext::Range - Specifies a range of keys, by start and end index,
// from the parent MultiGetContext. Each range contains a bit vector that
// indicates whether the corresponding keys need to be processed or skipped.
// A Range object can be copy constructed, and the new object inherits the
// original Range's bit vector. This is useful for progressively skipping
// keys as the lookup goes through various stages. For example, when looking
// up keys in the same SST file, a Range is created excluding keys not
// belonging to that file. A new Range is then copy constructed and individual
// keys are skipped based on bloom filter lookup.
class Range {
public:
// MultiGetContext::Range::Iterator - A forward iterator that iterates over
// non-skippable keys in a Range, as well as keys whose final value has been
// found. The latter is tracked by MultiGetContext::value_mask_
class Iterator {
public:
// -- iterator traits
using self_type = Iterator;
using value_type = KeyContext;
using reference = KeyContext&;
using pointer = KeyContext*;
using difference_type = int;
using iterator_category = std::forward_iterator_tag;
Iterator(const Range* range, size_t idx)
: range_(range), ctx_(range->ctx_), index_(idx) {
while (index_ < range_->end_ &&
(Mask{1} << index_) &
(range_->ctx_->value_mask_ | range_->skip_mask_ |
range_->invalid_mask_))
index_++;
}
Iterator(const Iterator&) = default;
Iterator(const Iterator& other, const Range* range)
: range_(range), ctx_(other.ctx_), index_(other.index_) {
assert(range->ctx_ == other.ctx_);
}
Iterator& operator=(const Iterator&) = default;
Iterator& operator++() {
while (++index_ < range_->end_ &&
(Mask{1} << index_) &
(range_->ctx_->value_mask_ | range_->skip_mask_ |
range_->invalid_mask_))
;
return *this;
}
bool operator==(Iterator other) const {
assert(range_->ctx_ == other.range_->ctx_);
return index_ == other.index_;
}
bool operator!=(Iterator other) const {
assert(range_->ctx_ == other.range_->ctx_);
return index_ != other.index_;
}
KeyContext& operator*() {
assert(index_ < range_->end_ && index_ >= range_->start_);
return *(ctx_->sorted_keys_[index_]);
}
KeyContext* operator->() {
assert(index_ < range_->end_ && index_ >= range_->start_);
return ctx_->sorted_keys_[index_];
}
size_t index() { return index_; }
private:
friend Range;
const Range* range_;
const MultiGetContext* ctx_;
size_t index_;
};
Range(const Range& mget_range,
const Iterator& first,
const Iterator& last) {
ctx_ = mget_range.ctx_;
if (first == last) {
// This means create an empty range based on mget_range. So just
// set start_ and and_ to the same value
start_ = mget_range.start_;
end_ = start_;
} else {
start_ = first.index_;
end_ = last.index_;
}
skip_mask_ = mget_range.skip_mask_;
invalid_mask_ = mget_range.invalid_mask_;
assert(start_ < 64);
assert(end_ < 64);
}
Range() = default;
Iterator begin() const { return Iterator(this, start_); }
Iterator end() const { return Iterator(this, end_); }
bool empty() const { return RemainingMask() == 0; }
void SkipIndex(size_t index) { skip_mask_ |= Mask{1} << index; }
void SkipKey(const Iterator& iter) { SkipIndex(iter.index_); }
bool IsKeySkipped(const Iterator& iter) const {
return skip_mask_ & (Mask{1} << iter.index_);
}
// Update the value_mask_ in MultiGetContext so its
// immediately reflected in all the Range Iterators
void MarkKeyDone(Iterator& iter) {
ctx_->value_mask_ |= (Mask{1} << iter.index_);
}
bool CheckKeyDone(Iterator& iter) const {
return ctx_->value_mask_ & (Mask{1} << iter.index_);
}
uint64_t KeysLeft() const { return BitsSetToOne(RemainingMask()); }
void AddSkipsFrom(const Range& other) {
assert(ctx_ == other.ctx_);
skip_mask_ |= other.skip_mask_;
}
uint64_t GetValueSize() { return ctx_->value_size_; }
void AddValueSize(uint64_t value_size) { ctx_->value_size_ += value_size; }
MultiGetContext* context() const { return ctx_; }
Range Suffix(const Range& other) const {
size_t other_last = other.FindLastRemaining();
size_t my_last = FindLastRemaining();
if (my_last > other_last) {
return Range(*this, Iterator(this, other_last),
Iterator(this, my_last));
} else {
return Range(*this, begin(), begin());
}
}
// The += operator expands the number of keys in this range. The expansion
// is always to the right, i.e start of the additional range >= end of
// current range. There should be no overlap. Any skipped keys in rhs are
// marked as invalid in the invalid_mask_.
Range& operator+=(const Range& rhs) {
assert(rhs.start_ >= end_);
// Check for non-overlapping ranges and adjust invalid_mask_ accordingly
if (end_ < rhs.start_) {
invalid_mask_ |= RangeMask(end_, rhs.start_);
skip_mask_ |= RangeMask(end_, rhs.start_);
}
start_ = std::min<size_t>(start_, rhs.start_);
end_ = std::max<size_t>(end_, rhs.end_);
skip_mask_ |= rhs.skip_mask_ & RangeMask(rhs.start_, rhs.end_);
invalid_mask_ |= (rhs.invalid_mask_ | rhs.skip_mask_) &
RangeMask(rhs.start_, rhs.end_);
assert(start_ < 64);
assert(end_ < 64);
return *this;
}
// The -= operator removes keys from this range. The removed keys should
// come from a range completely overlapping the current range. The removed
// keys are marked invalid in the invalid_mask_.
Range& operator-=(const Range& rhs) {
assert(start_ <= rhs.start_ && end_ >= rhs.end_);
skip_mask_ |= (~rhs.skip_mask_ | rhs.invalid_mask_) &
RangeMask(rhs.start_, rhs.end_);
invalid_mask_ |= (~rhs.skip_mask_ | rhs.invalid_mask_) &
RangeMask(rhs.start_, rhs.end_);
return *this;
}
// Return a complement of the current range
Range operator~() {
Range res = *this;
res.skip_mask_ = ~skip_mask_ & RangeMask(start_, end_);
return res;
}
private:
friend MultiGetContext;
MultiGetContext* ctx_;
size_t start_;
size_t end_;
Mask skip_mask_;
Mask invalid_mask_;
Range(MultiGetContext* ctx, size_t num_keys)
: ctx_(ctx),
start_(0),
end_(num_keys),
skip_mask_(0),
invalid_mask_(0) {
assert(num_keys < 64);
}
static Mask RangeMask(size_t start, size_t end) {
return (((Mask{1} << (end - start)) - 1) << start);
}
Mask RemainingMask() const {
return (((Mask{1} << end_) - 1) & ~((Mask{1} << start_) - 1) &
~(ctx_->value_mask_ | skip_mask_));
}
size_t FindLastRemaining() const {
Mask mask = RemainingMask();
size_t index = (mask >>= start_) ? start_ : 0;
while (mask >>= 1) {
index++;
}
return index;
}
};
// Return the initial range that encompasses all the keys in the batch
Range GetMultiGetRange() { return Range(this, num_keys_); }
};
} // namespace ROCKSDB_NAMESPACE