fix: increase the tiering part size to 128MiB (#19424)

also introduce 8MiB buffer to read from for
bigger parts

cmd/handler-api.go

@@ -142,7 +142,7 @@ func (t *apiConfig) init(cfg api.Config, setDriveCounts []int) {
 		// total_ram / ram_per_request
 		// ram_per_request is (2MiB+128KiB) * driveCount \
 		//    + 2 * 10MiB (default erasure block size v1) + 2 * 1MiB (default erasure block size v2)
-		blockSize := xioutil.BlockSizeLarge + xioutil.BlockSizeSmall
+		blockSize := xioutil.LargeBlock + xioutil.SmallBlock
 		apiRequestsMaxPerNode = int(maxMem / uint64(maxSetDrives*blockSize+int(blockSizeV1*2+blockSizeV2*2)))
 		if globalIsDistErasure {
 			logger.Info("Automatically configured API requests per node based on available memory on the system: %d", apiRequestsMaxPerNode)

cmd/storage-datatypes.go

@@ -450,7 +450,7 @@ func (r *RenameDataInlineHandlerParams) Recycle() {
 	if r == nil {
 		return
 	}
-	if cap(r.FI.Data) >= xioutil.BlockSizeSmall {
+	if cap(r.FI.Data) >= xioutil.SmallBlock {
 		grid.PutByteBuffer(r.FI.Data)
 		r.FI.Data = nil
 	}

cmd/storage-rest-server.go

@@ -595,6 +595,8 @@ func (s *storageRESTServer) ReadFileStreamHandler(w http.ResponseWriter, r *http
 		// Windows can lock up with this optimization, so we fall back to regular copy.
 		sr, ok := rc.(*sendFileReader)
 		if ok {
+			// Sendfile sends in 4MiB chunks per sendfile syscall which is more than enough
+			// for most setups.
 			_, err = rf.ReadFrom(sr.Reader)
 			if !xnet.IsNetworkOrHostDown(err, true) { // do not need to log disconnected clients
 				storageLogIf(r.Context(), err)

cmd/warm-backend-minio.go

@@ -42,7 +42,7 @@ const (
 	maxMultipartPutObjectSize = 1024 * 1024 * 1024 * 1024 * 5
 	maxPartsCount             = 10000
 	maxPartSize               = 1024 * 1024 * 1024 * 5
-	minPartSize               = 1024 * 1024 * 64 // chosen by us to be optimal for HDDs
+	minPartSize               = 1024 * 1024 * 128 // chosen by us to be optimal for HDDs
 )
 
 // optimalPartInfo - calculate the optimal part info for a given

cmd/xl-storage.go

@@ -2131,11 +2131,15 @@ func (s *xlStorage) writeAllDirect(ctx context.Context, filePath string, fileSiz
 
 	var bufp *[]byte
 	switch {
-	case fileSize > 0 && fileSize >= xioutil.BlockSizeReallyLarge:
+	case fileSize > 0 && fileSize >= xioutil.XXLargeBlock*2:
+		// use a larger 8MiB buffer for a really really large streamsx.
+		bufp = xioutil.ODirectPoolXXLarge.Get().(*[]byte)
+		defer xioutil.ODirectPoolXXLarge.Put(bufp)
+	case fileSize > 0 && fileSize >= xioutil.XLargeBlock:
 		// use a larger 4MiB buffer for a really large streams.
 		bufp = xioutil.ODirectPoolXLarge.Get().(*[]byte)
 		defer xioutil.ODirectPoolXLarge.Put(bufp)
-	case fileSize <= xioutil.BlockSizeSmall:
+	case fileSize <= xioutil.SmallBlock:
 		bufp = xioutil.ODirectPoolSmall.Get().(*[]byte)
 		defer xioutil.ODirectPoolSmall.Put(bufp)
 	default:

internal/ioutil/ioutil.go

@@ -34,28 +34,35 @@ import (
 
 // Block sizes constant.
 const (
-	BlockSizeSmall       = 32 * humanize.KiByte // Default r/w block size for smaller objects.
-	BlockSizeLarge       = 1 * humanize.MiByte  // Default r/w block size for normal objects.
-	BlockSizeReallyLarge = 4 * humanize.MiByte  // Default r/w block size for very large objects.
+	SmallBlock   = 32 * humanize.KiByte // Default r/w block size for smaller objects.
+	LargeBlock   = 1 * humanize.MiByte  // Default r/w block size for normal objects.
+	XLargeBlock  = 4 * humanize.MiByte  // Default r/w block size for very large objects.
+	XXLargeBlock = 8 * humanize.MiByte  // Default r/w block size for very very large objects.
 )
 
 // aligned sync.Pool's
 var (
+	ODirectPoolXXLarge = sync.Pool{
+		New: func() interface{} {
+			b := disk.AlignedBlock(XXLargeBlock)
+			return &b
+		},
+	}
 	ODirectPoolXLarge = sync.Pool{
 		New: func() interface{} {
-			b := disk.AlignedBlock(BlockSizeReallyLarge)
+			b := disk.AlignedBlock(XLargeBlock)
 			return &b
 		},
 	}
 	ODirectPoolLarge = sync.Pool{
 		New: func() interface{} {
-			b := disk.AlignedBlock(BlockSizeLarge)
+			b := disk.AlignedBlock(LargeBlock)
 			return &b
 		},
 	}
 	ODirectPoolSmall = sync.Pool{
 		New: func() interface{} {
-			b := disk.AlignedBlock(BlockSizeSmall)
+			b := disk.AlignedBlock(SmallBlock)
 			return &b
 		},
 	}