bsc/core/state/trie_prefetcher.go

376 lines
11 KiB
Go
Raw Normal View History

// Copyright 2020 The go-ethereum Authors
// This file is part of the go-ethereum library.
//
// The go-ethereum library is free software: you can redistribute it and/or modify
// it under the terms of the GNU Lesser General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// The go-ethereum library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
package state
import (
"sync"
"github.com/ethereum/go-ethereum/common"
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
"github.com/ethereum/go-ethereum/common/gopool"
"github.com/ethereum/go-ethereum/log"
"github.com/ethereum/go-ethereum/metrics"
)
const abortChanSize = 64
var (
// triePrefetchMetricsPrefix is the prefix under which to publis the metrics.
triePrefetchMetricsPrefix = "trie/prefetch/"
)
// triePrefetcher is an active prefetcher, which receives accounts or storage
// items and does trie-loading of them. The goal is to get as much useful content
// into the caches as possible.
//
// Note, the prefetcher's API is not thread safe.
type triePrefetcher struct {
db Database // Database to fetch trie nodes through
root common.Hash // Root hash of theaccount trie for metrics
fetches map[common.Hash]Trie // Partially or fully fetcher tries
fetchers map[common.Hash]*subfetcher // Subfetchers for each trie
abortChan chan *subfetcher
closeChan chan struct{}
deliveryMissMeter metrics.Meter
accountLoadMeter metrics.Meter
accountDupMeter metrics.Meter
accountSkipMeter metrics.Meter
accountWasteMeter metrics.Meter
storageLoadMeter metrics.Meter
storageDupMeter metrics.Meter
storageSkipMeter metrics.Meter
storageWasteMeter metrics.Meter
}
// newTriePrefetcher
func newTriePrefetcher(db Database, root common.Hash, namespace string) *triePrefetcher {
prefix := triePrefetchMetricsPrefix + namespace
p := &triePrefetcher{
db: db,
root: root,
fetchers: make(map[common.Hash]*subfetcher), // Active prefetchers use the fetchers map
abortChan: make(chan *subfetcher, abortChanSize),
closeChan: make(chan struct{}),
deliveryMissMeter: metrics.GetOrRegisterMeter(prefix+"/deliverymiss", nil),
accountLoadMeter: metrics.GetOrRegisterMeter(prefix+"/account/load", nil),
accountDupMeter: metrics.GetOrRegisterMeter(prefix+"/account/dup", nil),
accountSkipMeter: metrics.GetOrRegisterMeter(prefix+"/account/skip", nil),
accountWasteMeter: metrics.GetOrRegisterMeter(prefix+"/account/waste", nil),
storageLoadMeter: metrics.GetOrRegisterMeter(prefix+"/storage/load", nil),
storageDupMeter: metrics.GetOrRegisterMeter(prefix+"/storage/dup", nil),
storageSkipMeter: metrics.GetOrRegisterMeter(prefix+"/storage/skip", nil),
storageWasteMeter: metrics.GetOrRegisterMeter(prefix+"/storage/waste", nil),
}
go p.abortLoop()
return p
}
func (p *triePrefetcher) abortLoop() {
for {
select {
case fetcher := <-p.abortChan:
fetcher.abort()
case <-p.closeChan:
// drain fetcher channel
for {
select {
case fetcher := <-p.abortChan:
fetcher.abort()
default:
return
}
}
}
}
}
// close iterates over all the subfetchers, aborts any that were left spinning
// and reports the stats to the metrics subsystem.
func (p *triePrefetcher) close() {
for _, fetcher := range p.fetchers {
p.abortChan <- fetcher // safe to do multiple times
if metrics.Enabled {
if fetcher.root == p.root {
p.accountLoadMeter.Mark(int64(len(fetcher.seen)))
p.accountDupMeter.Mark(int64(fetcher.dups))
p.accountSkipMeter.Mark(int64(len(fetcher.tasks)))
for _, key := range fetcher.used {
delete(fetcher.seen, string(key))
}
p.accountWasteMeter.Mark(int64(len(fetcher.seen)))
} else {
p.storageLoadMeter.Mark(int64(len(fetcher.seen)))
p.storageDupMeter.Mark(int64(fetcher.dups))
p.storageSkipMeter.Mark(int64(len(fetcher.tasks)))
for _, key := range fetcher.used {
delete(fetcher.seen, string(key))
}
p.storageWasteMeter.Mark(int64(len(fetcher.seen)))
}
}
}
close(p.closeChan)
// Clear out all fetchers (will crash on a second call, deliberate)
p.fetchers = nil
}
// copy creates a deep-but-inactive copy of the trie prefetcher. Any trie data
// already loaded will be copied over, but no goroutines will be started. This
// is mostly used in the miner which creates a copy of it's actively mutated
// state to be sealed while it may further mutate the state.
func (p *triePrefetcher) copy() *triePrefetcher {
copy := &triePrefetcher{
db: p.db,
root: p.root,
fetches: make(map[common.Hash]Trie), // Active prefetchers use the fetches map
deliveryMissMeter: p.deliveryMissMeter,
accountLoadMeter: p.accountLoadMeter,
accountDupMeter: p.accountDupMeter,
accountSkipMeter: p.accountSkipMeter,
accountWasteMeter: p.accountWasteMeter,
storageLoadMeter: p.storageLoadMeter,
storageDupMeter: p.storageDupMeter,
storageSkipMeter: p.storageSkipMeter,
storageWasteMeter: p.storageWasteMeter,
}
// If the prefetcher is already a copy, duplicate the data
if p.fetches != nil {
for root, fetch := range p.fetches {
copy.fetches[root] = p.db.CopyTrie(fetch)
}
return copy
}
// Otherwise we're copying an active fetcher, retrieve the current states
for root, fetcher := range p.fetchers {
copy.fetches[root] = fetcher.peek()
}
return copy
}
// prefetch schedules a batch of trie items to prefetch.
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
func (p *triePrefetcher) prefetch(root common.Hash, keys [][]byte, accountHash common.Hash) {
// If the prefetcher is an inactive one, bail out
if p.fetches != nil {
return
}
// Active fetcher, schedule the retrievals
fetcher := p.fetchers[root]
if fetcher == nil {
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
fetcher = newSubfetcher(p.db, root, accountHash)
p.fetchers[root] = fetcher
}
fetcher.schedule(keys)
}
// trie returns the trie matching the root hash, or nil if the prefetcher doesn't
// have it.
func (p *triePrefetcher) trie(root common.Hash) Trie {
// If the prefetcher is inactive, return from existing deep copies
if p.fetches != nil {
trie := p.fetches[root]
if trie == nil {
p.deliveryMissMeter.Mark(1)
return nil
}
return p.db.CopyTrie(trie)
}
// Otherwise the prefetcher is active, bail if no trie was prefetched for this root
fetcher := p.fetchers[root]
if fetcher == nil {
p.deliveryMissMeter.Mark(1)
return nil
}
// Interrupt the prefetcher if it's by any chance still running and return
// a copy of any pre-loaded trie.
p.abortChan <- fetcher // safe to do multiple times
trie := fetcher.peek()
if trie == nil {
p.deliveryMissMeter.Mark(1)
return nil
}
return trie
}
// used marks a batch of state items used to allow creating statistics as to
// how useful or wasteful the prefetcher is.
func (p *triePrefetcher) used(root common.Hash, used [][]byte) {
if fetcher := p.fetchers[root]; fetcher != nil {
fetcher.used = used
}
}
// subfetcher is a trie fetcher goroutine responsible for pulling entries for a
// single trie. It is spawned when a new root is encountered and lives until the
// main prefetcher is paused and either all requested items are processed or if
// the trie being worked on is retrieved from the prefetcher.
type subfetcher struct {
db Database // Database to load trie nodes through
root common.Hash // Root hash of the trie to prefetch
trie Trie // Trie being populated with nodes
tasks [][]byte // Items queued up for retrieval
lock sync.Mutex // Lock protecting the task queue
wake chan struct{} // Wake channel if a new task is scheduled
stop chan struct{} // Channel to interrupt processing
term chan struct{} // Channel to signal iterruption
copy chan chan Trie // Channel to request a copy of the current trie
seen map[string]struct{} // Tracks the entries already loaded
dups int // Number of duplicate preload tasks
used [][]byte // Tracks the entries used in the end
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
accountHash common.Hash
}
// newSubfetcher creates a goroutine to prefetch state items belonging to a
// particular root hash.
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
func newSubfetcher(db Database, root common.Hash, accountHash common.Hash) *subfetcher {
sf := &subfetcher{
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
db: db,
root: root,
wake: make(chan struct{}, 1),
stop: make(chan struct{}),
term: make(chan struct{}),
copy: make(chan chan Trie),
seen: make(map[string]struct{}),
accountHash: accountHash,
}
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
gopool.Submit(func() {
sf.loop()
})
return sf
}
// schedule adds a batch of trie keys to the queue to prefetch.
func (sf *subfetcher) schedule(keys [][]byte) {
// Append the tasks to the current queue
sf.lock.Lock()
sf.tasks = append(sf.tasks, keys...)
sf.lock.Unlock()
// Notify the prefetcher, it's fine if it's already terminated
select {
case sf.wake <- struct{}{}:
default:
}
}
// peek tries to retrieve a deep copy of the fetcher's trie in whatever form it
// is currently.
func (sf *subfetcher) peek() Trie {
ch := make(chan Trie)
select {
case sf.copy <- ch:
// Subfetcher still alive, return copy from it
return <-ch
case <-sf.term:
// Subfetcher already terminated, return a copy directly
if sf.trie == nil {
return nil
}
return sf.db.CopyTrie(sf.trie)
}
}
// abort interrupts the subfetcher immediately. It is safe to call abort multiple
// times but it is not thread safe.
func (sf *subfetcher) abort() {
select {
case <-sf.stop:
default:
close(sf.stop)
}
<-sf.term
}
// loop waits for new tasks to be scheduled and keeps loading them until it runs
// out of tasks or its underlying trie is retrieved for committing.
func (sf *subfetcher) loop() {
// No matter how the loop stops, signal anyone waiting that it's terminated
defer close(sf.term)
// Start by opening the trie and stop processing if it fails
[R4R] performance improvement in many aspects (#257) * focus on performance improvement in many aspects. 1. Do BlockBody verification concurrently; 2. Do calculation of intermediate root concurrently; 3. Preload accounts before processing blocks; 4. Make the snapshot layers configurable. 5. Reuse some object to reduce GC. add * rlp: improve decoder stream implementation (#22858) This commit makes various cleanup changes to rlp.Stream. * rlp: shrink Stream struct This removes a lot of unused padding space in Stream by reordering the fields. The size of Stream changes from 120 bytes to 88 bytes. Stream instances are internally cached and reused using sync.Pool, so this does not improve performance. * rlp: simplify list stack The list stack kept track of the size of the current list context as well as the current offset into it. The size had to be stored in the stack in order to subtract it from the remaining bytes of any enclosing list in ListEnd. It seems that this can be implemented in a simpler way: just subtract the size from the enclosing list context in List instead. * rlp: use atomic.Value for type cache (#22902) All encoding/decoding operations read the type cache to find the writer/decoder function responsible for a type. When analyzing CPU profiles of geth during sync, I found that the use of sync.RWMutex in cache lookups appears in the profiles. It seems we are running into CPU cache contention problems when package rlp is heavily used on all CPU cores during sync. This change makes it use atomic.Value + a writer lock instead of sync.RWMutex. In the common case where the typeinfo entry is present in the cache, we simply fetch the map and lookup the type. * rlp: optimize byte array handling (#22924) This change improves the performance of encoding/decoding [N]byte. name old time/op new time/op delta DecodeByteArrayStruct-8 336ns ± 0% 246ns ± 0% -26.98% (p=0.000 n=9+10) EncodeByteArrayStruct-8 225ns ± 1% 148ns ± 1% -34.12% (p=0.000 n=10+10) name old alloc/op new alloc/op delta DecodeByteArrayStruct-8 120B ± 0% 48B ± 0% -60.00% (p=0.000 n=10+10) EncodeByteArrayStruct-8 0.00B 0.00B ~ (all equal) * rlp: optimize big.Int decoding for size <= 32 bytes (#22927) This change grows the static integer buffer in Stream to 32 bytes, making it possible to decode 256bit integers without allocating a temporary buffer. In the recent commit 088da24, Stream struct size decreased from 120 bytes down to 88 bytes. This commit grows the struct to 112 bytes again, but the size change will not degrade performance because Stream instances are internally cached in sync.Pool. name old time/op new time/op delta DecodeBigInts-8 12.2µs ± 0% 8.6µs ± 4% -29.58% (p=0.000 n=9+10) name old speed new speed delta DecodeBigInts-8 230MB/s ± 0% 326MB/s ± 4% +42.04% (p=0.000 n=9+10) * eth/protocols/eth, les: avoid Raw() when decoding HashOrNumber (#22841) Getting the raw value is not necessary to decode this type, and decoding it directly from the stream is faster. * fix testcase * debug no lazy * fix can not repair * address comments Co-authored-by: Felix Lange <fjl@twurst.com>
2021-07-29 12:16:53 +03:00
var trie Trie
var err error
if sf.accountHash == emptyAddr {
trie, err = sf.db.OpenTrie(sf.root)
} else {
// address is useless
trie, err = sf.db.OpenStorageTrie(sf.accountHash, sf.root)
}
if err != nil {
log.Warn("Trie prefetcher failed opening trie", "root", sf.root, "err", err)
return
}
sf.trie = trie
// Trie opened successfully, keep prefetching items
for {
select {
case <-sf.wake:
// Subfetcher was woken up, retrieve any tasks to avoid spinning the lock
sf.lock.Lock()
tasks := sf.tasks
sf.tasks = nil
sf.lock.Unlock()
// Prefetch any tasks until the loop is interrupted
for i, task := range tasks {
select {
case <-sf.stop:
// If termination is requested, add any leftover back and return
sf.lock.Lock()
sf.tasks = append(sf.tasks, tasks[i:]...)
sf.lock.Unlock()
return
case ch := <-sf.copy:
// Somebody wants a copy of the current trie, grant them
ch <- sf.db.CopyTrie(sf.trie)
default:
// No termination request yet, prefetch the next entry
taskid := string(task)
if _, ok := sf.seen[taskid]; ok {
sf.dups++
} else {
sf.trie.TryGet(task)
sf.seen[taskid] = struct{}{}
}
}
}
case ch := <-sf.copy:
// Somebody wants a copy of the current trie, grant them
ch <- sf.db.CopyTrie(sf.trie)
case <-sf.stop:
// Termination is requested, abort and leave remaining tasks
return
}
}
}