VictoriaMetrics/lib/fs/reader_at.go
Aliaksandr Valialkin d120197676 lib/fs: optimize queries that read recent samples for big number of time series
Use standard copy() func instead of mmap-aware copy func for reading recently touched mmap-ed data.
This improves read performance by up to 4x.
2020-06-05 19:10:04 +03:00

222 lines
5.8 KiB
Go

package fs
import (
"flag"
"fmt"
"os"
"sync"
"sync/atomic"
"time"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
"github.com/VictoriaMetrics/metrics"
"golang.org/x/sys/unix"
)
var disableMmap = flag.Bool("fs.disableMmap", is32BitPtr, "Whether to use pread() instead of mmap() for reading data files. "+
"By default mmap() is used for 64-bit arches and pread() is used for 32-bit arches, since they cannot data files bigger than 2^32 bytes in memory")
const is32BitPtr = (^uintptr(0) >> 32) == 0
// MustReadAtCloser is rand-access read interface.
type MustReadAtCloser interface {
// MustReadAt must read len(p) bytes from offset off to p.
MustReadAt(p []byte, off int64)
// MustClose must close the reader.
MustClose()
}
// ReaderAt implements rand-access reader.
type ReaderAt struct {
f *os.File
mmapData []byte
// pageCacheBitmap holds a bitmap for recently touched pages in mmapData.
// This bitmap allows using simple copy() instead of copyMmap() for reading recently touched pages,
// which is up to 4x faster when reading small chunks of data via MustReadAt.
pageCacheBitmap atomic.Value
pageCacheBitmapWG sync.WaitGroup
stopCh chan struct{}
}
type pageCacheBitmap struct {
m []uint64
}
// MustReadAt reads len(p) bytes at off from r.
func (r *ReaderAt) MustReadAt(p []byte, off int64) {
if len(p) == 0 {
return
}
if off < 0 {
logger.Panicf("off=%d cannot be negative", off)
}
end := off + int64(len(p))
if len(r.mmapData) == 0 || (len(p) > 8*1024 && !r.isInPageCache(off, end)) {
// Read big blocks directly from file.
// This could be faster than reading these blocks from mmap,
// since it triggers less page faults.
n, err := r.f.ReadAt(p, off)
if err != nil {
logger.Panicf("FATAL: cannot read %d bytes at offset %d of file %q: %s", len(p), off, r.f.Name(), err)
}
if n != len(p) {
logger.Panicf("FATAL: unexpected number of bytes read; got %d; want %d", n, len(p))
}
if len(r.mmapData) > 0 {
r.markInPageCache(off, end)
}
} else {
if off > int64(len(r.mmapData)-len(p)) {
logger.Panicf("off=%d is out of allowed range [0...%d] for len(p)=%d", off, len(r.mmapData)-len(p), len(p))
}
src := r.mmapData[off:]
if r.isInPageCache(off, end) {
// It is safe copying the data with copy(), since it is likely it is in the page cache.
// This is up to 4x faster than copyMmap() below.
copy(p, src)
} else {
// The data may be missing in the page cache, so it is better to copy it via cgo trick
// in order to avoid P stalls in Go runtime.
// See https://medium.com/@valyala/mmap-in-go-considered-harmful-d92a25cb161d for details.
copyMmap(p, src)
r.markInPageCache(off, end)
}
}
readCalls.Inc()
readBytes.Add(len(p))
}
func (r *ReaderAt) isInPageCache(start, end int64) bool {
startBit := uint64(start) / pageSize
endBit := uint64(end) / pageSize
m := r.pageCacheBitmap.Load().(*pageCacheBitmap).m
for startBit <= endBit {
idx := startBit / 64
off := startBit % 64
if idx >= uint64(len(m)) {
return true
}
n := atomic.LoadUint64(&m[idx])
if (n>>off)&1 != 1 {
return false
}
startBit++
}
return true
}
func (r *ReaderAt) markInPageCache(start, end int64) {
startBit := uint64(start) / pageSize
endBit := uint64(end) / pageSize
m := r.pageCacheBitmap.Load().(*pageCacheBitmap).m
for startBit <= endBit {
idx := startBit / 64
off := startBit % 64
n := atomic.LoadUint64(&m[idx])
n |= 1 << off
// It is OK if multiple concurrent goroutines store the same m[idx].
atomic.StoreUint64(&m[idx], n)
startBit++
}
}
// Assume page size is 4KB
const pageSize = 4 * 1024
// MustClose closes r.
func (r *ReaderAt) MustClose() {
close(r.stopCh)
r.pageCacheBitmapWG.Wait()
fname := r.f.Name()
if len(r.mmapData) > 0 {
if err := unix.Munmap(r.mmapData); err != nil {
logger.Panicf("FATAL: cannot unmap data for file %q: %s", fname, err)
}
r.mmapData = nil
}
MustClose(r.f)
r.f = nil
readersCount.Dec()
}
// OpenReaderAt opens ReaderAt for reading from filename.
//
// MustClose must be called on the returned ReaderAt when it is no longer needed.
func OpenReaderAt(path string) (*ReaderAt, error) {
f, err := os.Open(path)
if err != nil {
return nil, fmt.Errorf("cannot open file %q for reader: %s", path, err)
}
var r ReaderAt
r.f = f
r.stopCh = make(chan struct{})
if !*disableMmap {
fi, err := f.Stat()
if err != nil {
return nil, fmt.Errorf("error in stat: %s", err)
}
size := fi.Size()
bm := &pageCacheBitmap{
m: make([]uint64, 1+size/pageSize/64),
}
r.pageCacheBitmap.Store(bm)
r.pageCacheBitmapWG.Add(1)
go func() {
defer r.pageCacheBitmapWG.Done()
pageCacheBitmapCleaner(&r.pageCacheBitmap, r.stopCh)
}()
data, err := mmapFile(f, size)
if err != nil {
MustClose(f)
return nil, fmt.Errorf("cannot init reader for %q: %s", path, err)
}
r.mmapData = data
}
readersCount.Inc()
return &r, nil
}
func pageCacheBitmapCleaner(pcbm *atomic.Value, stopCh <-chan struct{}) {
t := time.NewTimer(time.Minute)
for {
select {
case <-stopCh:
t.Stop()
return
case <-t.C:
}
bmOld := pcbm.Load().(*pageCacheBitmap)
bm := &pageCacheBitmap{
m: make([]uint64, len(bmOld.m)),
}
pcbm.Store(bm)
}
}
var (
readCalls = metrics.NewCounter(`vm_fs_read_calls_total`)
readBytes = metrics.NewCounter(`vm_fs_read_bytes_total`)
readersCount = metrics.NewCounter(`vm_fs_readers`)
)
func mmapFile(f *os.File, size int64) ([]byte, error) {
if size == 0 {
return nil, nil
}
if size < 0 {
return nil, fmt.Errorf("got negative file size: %d bytes", size)
}
if int64(int(size)) != size {
return nil, fmt.Errorf("file is too big to be mmap'ed: %d bytes", size)
}
data, err := unix.Mmap(int(f.Fd()), 0, int(size), unix.PROT_READ, unix.MAP_SHARED)
if err != nil {
return nil, fmt.Errorf("cannot mmap file with size %d: %s", size, err)
}
return data, nil
}