add optimizations to bring performance on unversioned READS (#20128)

allow non-inlined on disk to be inlined via
an unversioned ReadVersion() call, we only
need ReadXL() to resolve objects with multiple
versions only.

The choice of this block makes it to be dynamic
and chosen by the user via `mc admin config set`

Other bonus things

- Start measuring internode TTFB performance.
- Set TCP_NODELAY, TCP_CORK for low latency

cmd/erasure-healing-common_test.go

@@ -23,6 +23,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"testing"
 	"time"
 
@@ -147,6 +148,10 @@ func TestCommonTime(t *testing.T) {
 // TestListOnlineDisks - checks if listOnlineDisks and outDatedDisks
 // are consistent with each other.
 func TestListOnlineDisks(t *testing.T) {
+	if runtime.GOOS == globalWindowsOSName {
+		t.Skip()
+	}
+
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
@@ -230,7 +235,7 @@ func TestListOnlineDisks(t *testing.T) {
 	}
 
 	object := "object"
-	data := bytes.Repeat([]byte("a"), smallFileThreshold*16)
+	data := bytes.Repeat([]byte("a"), smallFileThreshold*32)
 	z := obj.(*erasureServerPools)
 
 	erasureDisks, err := z.GetDisks(0, 0)

cmd/erasure-object.go

@@ -283,7 +283,7 @@ func (er erasureObjects) GetObjectNInfo(ctx context.Context, bucket, object stri
 	}
 
 	if unlockOnDefer {
-		unlockOnDefer = fi.InlineData()
+		unlockOnDefer = fi.InlineData() || len(fi.Data) > 0
 	}
 
 	pr, pw := xioutil.WaitPipe()
@@ -908,6 +908,8 @@ func (er erasureObjects) getObjectFileInfo(ctx context.Context, bucket, object s
 		}
 
 		rw.Lock()
+		// when its a versioned bucket and empty versionID - at totalResp == setDriveCount
+		// we must use rawFileInfo to resolve versions to figure out the latest version.
 		if opts.VersionID == "" && totalResp == er.setDriveCount {
 			fi, onlineMeta, onlineDisks, modTime, etag, err = calcQuorum(pickLatestQuorumFilesInfo(ctx,
 				rawArr, errs, bucket, object, readData, opts.InclFreeVersions, true))
@@ -915,7 +917,7 @@ func (er erasureObjects) getObjectFileInfo(ctx context.Context, bucket, object s
 			fi, onlineMeta, onlineDisks, modTime, etag, err = calcQuorum(metaArr, errs)
 		}
 		rw.Unlock()
-		if err == nil && fi.InlineData() {
+		if err == nil && (fi.InlineData() || len(fi.Data) > 0) {
 			break
 		}
 	}
@@ -1399,7 +1401,7 @@ func (er erasureObjects) putObject(ctx context.Context, bucket string, object st
 	writers := make([]io.Writer, len(onlineDisks))
 	var inlineBuffers []*bytes.Buffer
 	if shardFileSize >= 0 {
-		if !opts.Versioned && shardFileSize < inlineBlock {
+		if !opts.Versioned && shardFileSize <= inlineBlock {
 			inlineBuffers = make([]*bytes.Buffer, len(onlineDisks))
 		} else if shardFileSize < inlineBlock/8 {
 			inlineBuffers = make([]*bytes.Buffer, len(onlineDisks))
@@ -1407,7 +1409,7 @@ func (er erasureObjects) putObject(ctx context.Context, bucket string, object st
 	} else {
 		// If compressed, use actual size to determine.
 		if sz := erasure.ShardFileSize(data.ActualSize()); sz > 0 {
-			if !opts.Versioned && sz < inlineBlock {
+			if !opts.Versioned && sz <= inlineBlock {
 				inlineBuffers = make([]*bytes.Buffer, len(onlineDisks))
 			} else if sz < inlineBlock/8 {
 				inlineBuffers = make([]*bytes.Buffer, len(onlineDisks))

cmd/storage-datatypes.go

@@ -264,6 +264,24 @@ type FileInfo struct {
 	Versioned bool `msg:"vs"`
 }
 
+func (fi FileInfo) shardSize() int64 {
+	return ceilFrac(fi.Erasure.BlockSize, int64(fi.Erasure.DataBlocks))
+}
+
+// ShardFileSize - returns final erasure size from original size.
+func (fi FileInfo) ShardFileSize(totalLength int64) int64 {
+	if totalLength == 0 {
+		return 0
+	}
+	if totalLength == -1 {
+		return -1
+	}
+	numShards := totalLength / fi.Erasure.BlockSize
+	lastBlockSize := totalLength % fi.Erasure.BlockSize
+	lastShardSize := ceilFrac(lastBlockSize, int64(fi.Erasure.DataBlocks))
+	return numShards*fi.shardSize() + lastShardSize
+}
+
 // ShallowCopy - copies minimal information for READ MRF checks.
 func (fi FileInfo) ShallowCopy() (n FileInfo) {
 	n.Volume = fi.Volume

cmd/storage-rest-client.go

@@ -499,7 +499,7 @@ var readMsgpReaderPool = sync.Pool{New: func() interface{} { return &msgp.Reader
 func msgpNewReader(r io.Reader) *msgp.Reader {
 	p := readMsgpReaderPool.Get().(*msgp.Reader)
 	if p.R == nil {
-		p.R = xbufio.NewReaderSize(r, 4<<10)
+		p.R = xbufio.NewReaderSize(r, 32<<10)
 	} else {
 		p.R.Reset(r)
 	}

cmd/xl-storage.go

@@ -1703,18 +1703,23 @@ func (s *xlStorage) ReadVersion(ctx context.Context, origvolume, volume, path, v
 			fi.Data = nil
 		}
 
+		attemptInline := fi.TransitionStatus == "" && fi.DataDir != "" && len(fi.Parts) == 1
 		// Reading data for small objects when
 		// - object has not yet transitioned
-		// - object size lesser than 128KiB
 		// - object has maximum of 1 parts
-		if fi.TransitionStatus == "" &&
+		if attemptInline {
-			fi.DataDir != "" && fi.Size <= smallFileThreshold &&
+			inlineBlock := globalStorageClass.InlineBlock()
-			len(fi.Parts) == 1 {
+			if inlineBlock <= 0 {
-			partPath := fmt.Sprintf("part.%d", fi.Parts[0].Number)
+				inlineBlock = 128 * humanize.KiByte
-			dataPath := pathJoin(volumeDir, path, fi.DataDir, partPath)
+			}
-			fi.Data, _, err = s.readAllData(ctx, volume, volumeDir, dataPath, false)
+
-			if err != nil {
+			canInline := fi.ShardFileSize(fi.Parts[0].ActualSize) <= inlineBlock
-				return FileInfo{}, err
+			if canInline {
+				dataPath := pathJoin(volumeDir, path, fi.DataDir, fmt.Sprintf("part.%d", fi.Parts[0].Number))
+				fi.Data, _, err = s.readAllData(ctx, volume, volumeDir, dataPath, false)
+				if err != nil {
+					return FileInfo{}, err
+				}
 			}
 		}
 	}
@@ -1768,7 +1773,7 @@ func (s *xlStorage) readAllData(ctx context.Context, volume, volumeDir string, f
 	}
 
 	if discard {
-		// This discard is mostly true for DELETEEs
+		// This discard is mostly true for deletes
 		// so we need to make sure we do not keep
 		// page-cache references after.
 		defer disk.Fdatasync(f)
@@ -2993,7 +2998,7 @@ func (s *xlStorage) ReadMultiple(ctx context.Context, req ReadMultipleReq, resp
 			if req.MetadataOnly {
 				data, mt, err = s.readMetadataWithDMTime(ctx, fullPath)
 			} else {
-				data, mt, err = s.readAllData(ctx, req.Bucket, volumeDir, fullPath, true)
+				data, mt, err = s.readAllData(ctx, req.Bucket, volumeDir, fullPath, false)
 			}
 			return err
 		}); err != nil {
@@ -3096,7 +3101,7 @@ func (s *xlStorage) CleanAbandonedData(ctx context.Context, volume string, path
 	}
 	baseDir := pathJoin(volumeDir, path+slashSeparator)
 	metaPath := pathutil.Join(baseDir, xlStorageFormatFile)
-	buf, _, err := s.readAllData(ctx, volume, volumeDir, metaPath, true)
+	buf, _, err := s.readAllData(ctx, volume, volumeDir, metaPath, false)
 	if err != nil {
 		return err
 	}

internal/http/dial_linux.go

@@ -49,6 +49,11 @@ func setTCPParametersFn(opts TCPOptions) func(network, address string, c syscall
 				_ = unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_RCVBUF, opts.RecvBufSize)
 			}
 
+			if opts.NoDelay {
+				_ = syscall.SetsockoptInt(fd, syscall.IPPROTO_TCP, unix.TCP_NODELAY, 1)
+				_ = syscall.SetsockoptInt(fd, syscall.SOL_TCP, unix.TCP_CORK, 0)
+			}
+
 			// Enable TCP open
 			// https://lwn.net/Articles/508865/ - 32k queue size.
 			_ = syscall.SetsockoptInt(fd, syscall.SOL_TCP, unix.TCP_FASTOPEN, 32*1024)

internal/http/listener.go

@@ -125,6 +125,7 @@ type TCPOptions struct {
 
 	SendBufSize int              // SO_SNDBUF size for the socket connection, NOTE: this sets server and client connection
 	RecvBufSize int              // SO_RECVBUF size for the socket connection, NOTE: this sets server and client connection
+	NoDelay     bool             // Indicates callers to enable TCP_NODELAY on the net.Conn
 	Interface   string           // This is a VRF device passed via `--interface` flag
 	Trace       func(msg string) // Trace when starting.
 }
@@ -136,6 +137,7 @@ func (t TCPOptions) ForWebsocket() TCPOptions {
 		Interface:   t.Interface,
 		SendBufSize: t.SendBufSize,
 		RecvBufSize: t.RecvBufSize,
+		NoDelay:     true,
 	}
 }
 

internal/rest/rpc-stats.go

@@ -22,14 +22,17 @@ import (
 	"net/http/httptrace"
 	"sync/atomic"
 	"time"
+
+	"github.com/minio/minio/internal/logger"
 )
 
 var globalStats = struct {
 	errs uint64
 
-	tcpDialErrs     uint64
+	tcpDialErrs                 uint64
-	tcpDialCount    uint64
+	tcpDialCount                uint64
-	tcpDialTotalDur uint64
+	tcpDialTotalDur             uint64
+	tcpTimeForFirstByteTotalDur uint64
 }{}
 
 // RPCStats holds information about the DHCP/TCP metrics and errors
@@ -37,6 +40,7 @@ type RPCStats struct {
 	Errs uint64
 
 	DialAvgDuration uint64
+	TTFBAvgDuration uint64
 	DialErrs        uint64
 }
 
@@ -48,6 +52,7 @@ func GetRPCStats() RPCStats {
 	}
 	if v := atomic.LoadUint64(&globalStats.tcpDialCount); v > 0 {
 		s.DialAvgDuration = atomic.LoadUint64(&globalStats.tcpDialTotalDur) / v
+		s.TTFBAvgDuration = atomic.LoadUint64(&globalStats.tcpTimeForFirstByteTotalDur) / v
 	}
 	return s
 }
@@ -55,14 +60,19 @@ func GetRPCStats() RPCStats {
 // Return a function which update the global stats related to tcp connections
 func setupReqStatsUpdate(req *http.Request) (*http.Request, func()) {
 	var dialStart, dialEnd int64
-
+	start := time.Now()
 	trace := &httptrace.ClientTrace{
+		GotFirstResponseByte: func() {
+			atomic.AddUint64(&globalStats.tcpTimeForFirstByteTotalDur, uint64(time.Since(start)))
+		},
 		ConnectStart: func(network, addr string) {
 			atomic.StoreInt64(&dialStart, time.Now().UnixNano())
 		},
 		ConnectDone: func(network, addr string, err error) {
 			if err == nil {
 				atomic.StoreInt64(&dialEnd, time.Now().UnixNano())
+			} else {
+				logger.LogOnceIf(req.Context(), logSubsys, err, req.URL.Hostname())
 			}
 		},
 	}