VictoriaMetrics/lib/streamaggr/dedup.go
Roman Khavronenko 7cb894a777
lib/streamaggr: reduce number of inuse objects (#6402)
The main change is getting rid of interning of sample key. It was
discovered that for cases with many unique time series aggregated by
vmagent interned keys could grow up to hundreds of millions of objects.
This has negative impact on the following aspects:
1. It slows down garbage collection cycles, as GC has to scan all inuse
objects periodically. The higher is the number of inuse objects, the
longer it takes/the more CPU it takes.
2. It slows down the hot path of samples aggregation where each key
needs to be looked up in the map first.

The change makes code more fragile, but suppose to provide performance
optimization for heavy-loaded vmagents with stream aggregation enabled.

---------

Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2024-06-07 15:45:52 +02:00

225 lines
4.5 KiB
Go

package streamaggr
import (
"strings"
"sync"
"unsafe"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/cespare/xxhash/v2"
)
const dedupAggrShardsCount = 128
type dedupAggr struct {
shards []dedupAggrShard
}
type dedupAggrShard struct {
dedupAggrShardNopad
// The padding prevents false sharing on widespread platforms with
// 128 mod (cache line size) = 0 .
_ [128 - unsafe.Sizeof(dedupAggrShardNopad{})%128]byte
}
type dedupAggrShardNopad struct {
mu sync.Mutex
m map[string]*dedupAggrSample
}
type dedupAggrSample struct {
value float64
timestamp int64
}
func newDedupAggr() *dedupAggr {
shards := make([]dedupAggrShard, dedupAggrShardsCount)
return &dedupAggr{
shards: shards,
}
}
func (da *dedupAggr) sizeBytes() uint64 {
n := uint64(unsafe.Sizeof(*da))
for i := range da.shards {
n += da.shards[i].sizeBytes()
}
return n
}
func (da *dedupAggr) itemsCount() uint64 {
n := uint64(0)
for i := range da.shards {
n += da.shards[i].itemsCount()
}
return n
}
func (das *dedupAggrShard) sizeBytes() uint64 {
das.mu.Lock()
n := uint64(unsafe.Sizeof(*das))
for k, s := range das.m {
n += uint64(len(k)) + uint64(unsafe.Sizeof(k)+unsafe.Sizeof(s))
}
das.mu.Unlock()
return n
}
func (das *dedupAggrShard) itemsCount() uint64 {
das.mu.Lock()
n := uint64(len(das.m))
das.mu.Unlock()
return n
}
func (da *dedupAggr) pushSamples(samples []pushSample) {
pss := getPerShardSamples()
shards := pss.shards
for _, sample := range samples {
h := xxhash.Sum64(bytesutil.ToUnsafeBytes(sample.key))
idx := h % uint64(len(shards))
shards[idx] = append(shards[idx], sample)
}
for i, shardSamples := range shards {
if len(shardSamples) == 0 {
continue
}
da.shards[i].pushSamples(shardSamples)
}
putPerShardSamples(pss)
}
func getDedupFlushCtx() *dedupFlushCtx {
v := dedupFlushCtxPool.Get()
if v == nil {
return &dedupFlushCtx{}
}
return v.(*dedupFlushCtx)
}
func putDedupFlushCtx(ctx *dedupFlushCtx) {
ctx.reset()
dedupFlushCtxPool.Put(ctx)
}
var dedupFlushCtxPool sync.Pool
type dedupFlushCtx struct {
samples []pushSample
}
func (ctx *dedupFlushCtx) reset() {
clear(ctx.samples)
ctx.samples = ctx.samples[:0]
}
func (da *dedupAggr) flush(f func(samples []pushSample), resetState bool) {
var wg sync.WaitGroup
for i := range da.shards {
flushConcurrencyCh <- struct{}{}
wg.Add(1)
go func(shard *dedupAggrShard) {
defer func() {
<-flushConcurrencyCh
wg.Done()
}()
ctx := getDedupFlushCtx()
shard.flush(ctx, f, resetState)
putDedupFlushCtx(ctx)
}(&da.shards[i])
}
wg.Wait()
}
type perShardSamples struct {
shards [][]pushSample
}
func (pss *perShardSamples) reset() {
shards := pss.shards
for i, shardSamples := range shards {
if len(shardSamples) > 0 {
clear(shardSamples)
shards[i] = shardSamples[:0]
}
}
}
func getPerShardSamples() *perShardSamples {
v := perShardSamplesPool.Get()
if v == nil {
return &perShardSamples{
shards: make([][]pushSample, dedupAggrShardsCount),
}
}
return v.(*perShardSamples)
}
func putPerShardSamples(pss *perShardSamples) {
pss.reset()
perShardSamplesPool.Put(pss)
}
var perShardSamplesPool sync.Pool
func (das *dedupAggrShard) pushSamples(samples []pushSample) {
das.mu.Lock()
defer das.mu.Unlock()
m := das.m
if m == nil {
m = make(map[string]*dedupAggrSample, len(samples))
das.m = m
}
for _, sample := range samples {
s, ok := m[sample.key]
if !ok {
m[strings.Clone(sample.key)] = &dedupAggrSample{
value: sample.value,
timestamp: sample.timestamp,
}
continue
}
// Update the existing value according to logic described at https://docs.victoriametrics.com/#deduplication
if sample.timestamp > s.timestamp || (sample.timestamp == s.timestamp && sample.value > s.value) {
s.value = sample.value
s.timestamp = sample.timestamp
}
}
}
func (das *dedupAggrShard) flush(ctx *dedupFlushCtx, f func(samples []pushSample), resetState bool) {
das.mu.Lock()
m := das.m
if resetState && len(m) > 0 {
das.m = make(map[string]*dedupAggrSample, len(m))
}
das.mu.Unlock()
if len(m) == 0 {
return
}
dstSamples := ctx.samples
for key, s := range m {
dstSamples = append(dstSamples, pushSample{
key: key,
value: s.value,
timestamp: s.timestamp,
})
// Limit the number of samples per each flush in order to limit memory usage.
if len(dstSamples) >= 100_000 {
f(dstSamples)
clear(dstSamples)
dstSamples = dstSamples[:0]
}
}
f(dstSamples)
ctx.samples = dstSamples
}