enable small and large file optimization (#11260)

- for large objects we found that 1MiB block for
  r/w respectively.
- for small objects we found that 128KiB block for
  r/w respectively.

cmd/erasure-object_test.go

@@ -290,7 +290,7 @@ func TestGetObjectNoQuorum(t *testing.T) {
 	bucket := "bucket"
 	object := "object"
 	opts := ObjectOptions{}
-	buf := make([]byte, 33<<10)
+	buf := make([]byte, 129<<10)
 	if _, err = io.ReadFull(crand.Reader, buf); err != nil {
 		t.Fatal(err)
 	}

cmd/handler-api.go

@@ -57,7 +57,7 @@ func (t *apiConfig) init(cfg api.Config, setDriveCount int) {
 		// max requests per node is calculated as
 		// total_ram / ram_per_request
 		// ram_per_request is 4MiB * setDriveCount + 2 * 10MiB (default erasure block size)
-		apiRequestsMaxPerNode = int(stats.TotalRAM / uint64(setDriveCount*(writeBlockSize+readBlockSize)+blockSizeV1*2))
+		apiRequestsMaxPerNode = int(stats.TotalRAM / uint64(setDriveCount*(blockSizeLarge+blockSizeSmall)+blockSizeV1*2))
 	} else {
 		apiRequestsMaxPerNode = cfg.RequestsMax
 		if len(globalEndpoints.Hostnames()) > 0 {

cmd/xl-storage.go

@@ -55,8 +55,8 @@ import (
 const (
 	nullVersionID     = "null"
 	diskMinTotalSpace = 900 * humanize.MiByte // Min 900MiB total space.
-	writeBlockSize    = 4 * humanize.MiByte   // Default write block size 4MiB.
-	readBlockSize     = 2 * humanize.MiByte   // Default read block size 2MiB.
+	blockSizeLarge    = 1 * humanize.MiByte   // Default r/w block size for larger objects.
+	blockSizeSmall    = 128 * humanize.KiByte // Default r/w block size for smaller objects.
 
 	// On regular files bigger than this;
 	readAheadSize = 16 << 20
@@ -65,9 +65,10 @@ const (
 	// Size of each buffer.
 	readAheadBufSize = 1 << 20
 
-	// Small file threshold below which data accompanies metadata
-	// from storage layer.
-	smallFileThreshold = 32 * humanize.KiByte
+	// Small file threshold below which data accompanies metadata from storage layer.
+	smallFileThreshold = 128 * humanize.KiByte // Optimized for NVMe/SSDs
+	// For hardrives it is possible to set this to a lower value to avoid any
+	// spike in latency. But currently we are simply keeping it optimal for SSDs.
 
 	// XL metadata file carries per object metadata.
 	xlStorageFormatFile = "xl.meta"
@@ -97,8 +98,8 @@ type xlStorage struct {
 
 	globalSync bool
 
-	wpool sync.Pool
-	rpool sync.Pool
+	poolLarge sync.Pool
+	poolSmall sync.Pool
 
 	rootDisk bool
 
@@ -261,15 +262,15 @@ func newXLStorage(ep Endpoint) (*xlStorage, error) {
 	p := &xlStorage{
 		diskPath: path,
 		endpoint: ep,
-		wpool: sync.Pool{
+		poolLarge: sync.Pool{
 			New: func() interface{} {
-				b := disk.AlignedBlock(writeBlockSize)
+				b := disk.AlignedBlock(blockSizeLarge)
 				return &b
 			},
 		},
-		rpool: sync.Pool{
+		poolSmall: sync.Pool{
 			New: func() interface{} {
-				b := disk.AlignedBlock(readBlockSize)
+				b := disk.AlignedBlock(blockSizeSmall)
 				return &b
 			},
 		},
@@ -1180,7 +1181,7 @@ func (s *xlStorage) readAllData(volumeDir, filePath string) (buf []byte, err err
 	}
 
 	atomic.AddInt32(&s.activeIOCount, 1)
-	rd := &odirectReader{f, nil, nil, true, s, nil}
+	rd := &odirectReader{f, nil, nil, true, true, s, nil}
 	defer rd.Close()
 
 	return ioutil.ReadAll(rd)
@@ -1410,6 +1411,7 @@ type odirectReader struct {
 	buf       []byte
 	bufp      *[]byte
 	freshRead bool
+	smallFile bool
 	s         *xlStorage
 	err       error
 }
@@ -1420,7 +1422,11 @@ func (o *odirectReader) Read(buf []byte) (n int, err error) {
 		return 0, o.err
 	}
 	if o.buf == nil {
-		o.bufp = o.s.rpool.Get().(*[]byte)
+		if o.smallFile {
+			o.bufp = o.s.poolSmall.Get().(*[]byte)
+		} else {
+			o.bufp = o.s.poolLarge.Get().(*[]byte)
+		}
 	}
 	if o.freshRead {
 		o.buf = *o.bufp
@@ -1449,8 +1455,14 @@ func (o *odirectReader) Read(buf []byte) (n int, err error) {
 
 // Close - Release the buffer and close the file.
 func (o *odirectReader) Close() error {
-	o.s.rpool.Put(o.bufp)
-	atomic.AddInt32(&o.s.activeIOCount, -1)
+	if o.smallFile {
+		o.s.poolSmall.Put(o.bufp)
+	} else {
+		o.s.poolLarge.Put(o.bufp)
+	}
+	defer func() {
+		atomic.AddInt32(&o.s.activeIOCount, -1)
+	}()
 	return o.f.Close()
 }
 
@@ -1472,6 +1484,7 @@ func (s *xlStorage) ReadFileStream(ctx context.Context, volume, path string, off
 	}
 
 	var file *os.File
+	// O_DIRECT only supported if offset is zero
 	if offset == 0 && globalStorageClass.GetDMA() == storageclass.DMAReadWrite {
 		file, err = disk.OpenFileDirectIO(filePath, os.O_RDONLY, 0666)
 	} else {
@@ -1512,7 +1525,10 @@ func (s *xlStorage) ReadFileStream(ctx context.Context, volume, path string, off
 
 	atomic.AddInt32(&s.activeIOCount, 1)
 	if offset == 0 && globalStorageClass.GetDMA() == storageclass.DMAReadWrite {
-		return &odirectReader{file, nil, nil, true, s, nil}, nil
+		if length <= smallFileThreshold {
+			return &odirectReader{file, nil, nil, true, true, s, nil}, nil
+		}
+		return &odirectReader{file, nil, nil, true, false, s, nil}, nil
 	}
 
 	if offset > 0 {
@@ -1525,7 +1541,9 @@ func (s *xlStorage) ReadFileStream(ctx context.Context, volume, path string, off
 		io.Reader
 		io.Closer
 	}{Reader: io.LimitReader(file, length), Closer: closeWrapper(func() error {
-		atomic.AddInt32(&s.activeIOCount, -1)
+		defer func() {
+			atomic.AddInt32(&s.activeIOCount, -1)
+		}()
 		return file.Close()
 	})}
 
@@ -1646,8 +1664,14 @@ func (s *xlStorage) CreateFile(ctx context.Context, volume, path string, fileSiz
 		w.Close()
 	}()
 
-	bufp := s.wpool.Get().(*[]byte)
-	defer s.wpool.Put(bufp)
+	var bufp *[]byte
+	if fileSize <= smallFileThreshold {
+		bufp = s.poolSmall.Get().(*[]byte)
+		defer s.poolSmall.Put(bufp)
+	} else {
+		bufp = s.poolLarge.Get().(*[]byte)
+		defer s.poolLarge.Put(bufp)
+	}
 
 	written, err := xioutil.CopyAligned(w, r, *bufp, fileSize)
 	if err != nil {