sstable: reduce block cache memory fragmentation

Currently, the sstable writer contains heuristics to flush sstable
blocks once the size reaches a specified threshold. In CRDB this is
defined as 32KiB. However, when these blocks are loaded into memory
additional metadata is allocated sometimes exceeding the 32KiB threshold.
Since CRDB uses jemalloc, these allocations use a 40KiB size class which
leads to significant internal fragmentation. In addition, since the
system is unaware of these size classes we cannot design heuristics that
prioritize reducing memory fragmentation. Reducing internal
fragmentation can help reduce CRDB's memory footprint. This commit
decrements the target block size to prevent internal fragmentation for
small key-value pairs and adds support for optionally specifying size
classes to enable a new set of heuristics that will reduce internal
fragmentation for workloads with larger key-value pairs.

Fixes: #999.

data_test.go

@@ -21,6 +21,7 @@ import (
 	"github.com/cockroachdb/errors"
 	"github.com/cockroachdb/pebble/bloom"
 	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/cache"
 	"github.com/cockroachdb/pebble/internal/humanize"
 	"github.com/cockroachdb/pebble/internal/keyspan"
 	"github.com/cockroachdb/pebble/internal/private"
@@ -523,8 +524,8 @@ func runBuildRemoteCmd(td *datadriven.TestData, d *DB, storage remote.Storage) e
 		// Force two-level indexes if not already forced on or off.
 		blockSize = 5
 	}
-	writeOpts.BlockSize = int(blockSize)
-	writeOpts.IndexBlockSize = int(blockSize)
+	writeOpts.BlockSize = int(blockSize) + cache.ValueMetadataSize
+	writeOpts.IndexBlockSize = writeOpts.BlockSize
 
 	f, err := storage.CreateObject(path)
 	if err != nil {
@@ -1427,7 +1428,7 @@ func parseDBOptionsArgs(opts *Options, args []datadriven.CmdArg) error {
 				return err
 			}
 			for i := range opts.Levels {
-				opts.Levels[i].BlockSize = v
+				opts.Levels[i].BlockSize = v + cache.ValueMetadataSize
 			}
 		case "cache-size":
 			if opts.Cache != nil {
@@ -1445,7 +1446,7 @@ func parseDBOptionsArgs(opts *Options, args []datadriven.CmdArg) error {
 				return err
 			}
 			for i := range opts.Levels {
-				opts.Levels[i].IndexBlockSize = v
+				opts.Levels[i].IndexBlockSize = v + cache.ValueMetadataSize
 			}
 		case "target-file-size":
 			v, err := strconv.Atoi(cmdArg.Vals[0])

internal/cache/value_cgo.go

@@ -0,0 +1,43 @@
+// Copyright 2024 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build ((!invariants && !tracing) || race) && cgo
+// +build !invariants,!tracing race
+// +build cgo
+
+package cache
+
+import (
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/manual"
+)
+
+// ValueMetadataSize denotes the number of bytes of metadata allocated for a
+// cache entry.
+const ValueMetadataSize = int(unsafe.Sizeof(Value{}))
+
+func newValue(n int) *Value {
+	if n == 0 {
+		return nil
+	}
+
+	// When we're not performing leak detection, the lifetime of the returned
+	// Value is exactly the lifetime of the backing buffer and we can manually
+	// allocate both.
+	b := manual.New(ValueMetadataSize + n)
+	v := (*Value)(unsafe.Pointer(&b[0]))
+	v.buf = b[ValueMetadataSize:]
+	v.ref.init(1)
+	return v
+}
+
+func (v *Value) free() {
+	// When we're not performing leak detection, the Value and buffer were
+	// allocated contiguously.
+	n := ValueMetadataSize + cap(v.buf)
+	buf := (*[manual.MaxArrayLen]byte)(unsafe.Pointer(v))[:n:n]
+	v.buf = nil
+	manual.Free(buf)
+}

internal/cache/value_invariants.go

@@ -15,6 +15,10 @@ import (
 	"github.com/cockroachdb/pebble/internal/manual"
 )
 
+// ValueMetadataSize denotes the number of bytes of metadata allocated for a
+// cache entry.
+const ValueMetadataSize = 0
+
 // newValue creates a Value with a manually managed buffer of size n.
 //
 // This definition of newValue is used when either the "invariants" or

internal/cache/value_normal.go

@@ -2,56 +2,27 @@
 // of this source code is governed by a BSD-style license that can be found in
 // the LICENSE file.
 
-//go:build (!invariants && !tracing) || race
+//go:build ((!invariants && !tracing) || race) && !cgo
 // +build !invariants,!tracing race
+// +build !cgo
 
 package cache
 
-import (
-	"unsafe"
-
-	"github.com/cockroachdb/pebble/internal/manual"
-)
-
-const valueSize = int(unsafe.Sizeof(Value{}))
+// ValueMetadataSize denotes the number of bytes of metadata allocated for a
+// cache entry.
+const ValueMetadataSize = 0
 
 func newValue(n int) *Value {
 	if n == 0 {
 		return nil
 	}
 
-	if !cgoEnabled {
-		// If Cgo is disabled then all memory is allocated from the Go heap and we
-		// can't play the trick below to combine the Value and buffer allocation.
-		v := &Value{buf: make([]byte, n)}
-		v.ref.init(1)
-		return v
-	}
-
-	// When we're not performing leak detection, the lifetime of the returned
-	// Value is exactly the lifetime of the backing buffer and we can manually
-	// allocate both.
-	//
-	// TODO(peter): It may be better to separate the allocation of the value and
-	// the buffer in order to reduce internal fragmentation in malloc. If the
-	// buffer is right at a power of 2, adding valueSize might push the
-	// allocation over into the next larger size.
-	b := manual.New(valueSize + n)
-	v := (*Value)(unsafe.Pointer(&b[0]))
-	v.buf = b[valueSize:]
+	// Since Cgo is disabled then all memory is allocated from the Go heap we
+	// can't play the trick below to combine the Value and buffer allocation.
+	v := &Value{buf: make([]byte, n)}
 	v.ref.init(1)
 	return v
 }
 
 func (v *Value) free() {
-	if !cgoEnabled {
-		return
-	}
-
-	// When we're not performing leak detection, the Value and buffer were
-	// allocated contiguously.
-	n := valueSize + cap(v.buf)
-	buf := (*[manual.MaxArrayLen]byte)(unsafe.Pointer(v))[:n:n]
-	v.buf = nil
-	manual.Free(buf)
 }

iterator_test.go

@@ -22,6 +22,7 @@ import (
 	"github.com/cockroachdb/errors"
 	"github.com/cockroachdb/pebble/internal/base"
 	"github.com/cockroachdb/pebble/internal/bytealloc"
+	"github.com/cockroachdb/pebble/internal/cache"
 	"github.com/cockroachdb/pebble/internal/invalidating"
 	"github.com/cockroachdb/pebble/internal/manifest"
 	"github.com/cockroachdb/pebble/internal/testkeys"
@@ -1190,7 +1191,10 @@ func TestIteratorBlockIntervalFilter(t *testing.T) {
 			FormatMajorVersion:      internalFormatNewest,
 			BlockPropertyCollectors: bpCollectors,
 		}
-		lo := LevelOptions{BlockSize: 1, IndexBlockSize: 1}
+		lo := LevelOptions{
+			BlockSize:      1 + cache.ValueMetadataSize,
+			IndexBlockSize: 1 + cache.ValueMetadataSize,
+		}
 		opts.Levels = append(opts.Levels, lo)
 
 		// Automatic compactions may compact away tombstones from L6, making

options.go

@@ -1060,6 +1060,12 @@ type Options struct {
 	// to temporarily persist data spilled to disk for row-oriented SQL query execution.
 	EnableSQLRowSpillMetrics bool
 
+	// AllocatorSizeClasses provides a sorted list containing the supported size
+	// classes of the underlying memory allocator. This provides hints to the
+	// sstable block writer's flushing policy to select block sizes that
+	// preemptively reduce internal fragmentation when loaded into the block cache.
+	AllocatorSizeClasses []int
+
 	// private options are only used by internal tests or are used internally
 	// for facilitating upgrade paths of unconfigurable functionality.
 	private struct {
@@ -1970,6 +1976,7 @@ func (o *Options) MakeWriterOptions(level int, format sstable.TableFormat) sstab
 	writerOpts.FilterPolicy = levelOpts.FilterPolicy
 	writerOpts.FilterType = levelOpts.FilterType
 	writerOpts.IndexBlockSize = levelOpts.IndexBlockSize
+	writerOpts.AllocatorSizeClasses = o.AllocatorSizeClasses
 	return writerOpts
 }
 

sstable/data_test.go

@@ -41,6 +41,7 @@ func optsFromArgs(td *datadriven.TestData, writerOpts *WriterOptions) error {
 			if err != nil {
 				return err
 			}
+			writerOpts.BlockSize += cache.ValueMetadataSize
 		case "index-block-size":
 			if len(arg.Vals) != 1 {
 				return errors.Errorf("%s: arg %s expects 1 value", td.Cmd, arg.Key)
@@ -50,6 +51,7 @@ func optsFromArgs(td *datadriven.TestData, writerOpts *WriterOptions) error {
 			if err != nil {
 				return err
 			}
+			writerOpts.IndexBlockSize += cache.ValueMetadataSize
 		case "filter":
 			writerOpts.FilterPolicy = bloom.FilterPolicy(10)
 		case "comparer-split-4b-suffix":

sstable/options.go

@@ -230,15 +230,24 @@ type WriterOptions struct {
 	// 750MB sstables -- see
 	// https://github.com/cockroachdb/cockroach/issues/117113).
 	DisableValueBlocks bool
+
+	// AllocatorSizeClasses provides a sorted list containing the supported size
+	// classes of the underlying memory allocator. This provides hints to the
+	// writer's flushing policy to select block sizes that preemptively reduce
+	// internal fragmentation when loaded into the block cache.
+	AllocatorSizeClasses []int
 }
 
 func (o WriterOptions) ensureDefaults() WriterOptions {
 	if o.BlockRestartInterval <= 0 {
 		o.BlockRestartInterval = base.DefaultBlockRestartInterval
 	}
-	if o.BlockSize <= 0 {
+	// The target block size is decremented to reduce internal fragmentation when
+	// blocks are loaded into the block cache.
+	if o.BlockSize <= cache.ValueMetadataSize {
 		o.BlockSize = base.DefaultBlockSize
 	}
+	o.BlockSize -= cache.ValueMetadataSize
 	if o.BlockSizeThreshold <= 0 {
 		o.BlockSizeThreshold = base.DefaultBlockSizeThreshold
 	}
@@ -248,8 +257,10 @@ func (o WriterOptions) ensureDefaults() WriterOptions {
 	if o.Compression <= DefaultCompression || o.Compression >= NCompression {
 		o.Compression = SnappyCompression
 	}
-	if o.IndexBlockSize <= 0 {
+	if o.IndexBlockSize <= cache.ValueMetadataSize {
 		o.IndexBlockSize = o.BlockSize
+	} else {
+		o.IndexBlockSize -= cache.ValueMetadataSize
 	}
 	if o.MergerName == "" {
 		o.MergerName = base.DefaultMerger.Name

sstable/reader_test.go

@@ -1406,8 +1406,8 @@ func TestReaderChecksumErrors(t *testing.T) {
 						}
 
 						w := NewWriter(objstorageprovider.NewFileWritable(f), WriterOptions{
-							BlockSize:      blockSize,
-							IndexBlockSize: indexBlockSize,
+							BlockSize:      blockSize + cache.ValueMetadataSize,
+							IndexBlockSize: indexBlockSize + cache.ValueMetadataSize,
 							Checksum:       checksumType,
 						})
 						require.NoError(t, w.Set(bytes.Repeat([]byte("a"), blockSize), nil))

sstable/test_fixtures.go

@@ -17,6 +17,7 @@ import (
 
 	"github.com/cockroachdb/pebble/bloom"
 	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/cache"
 	"github.com/cockroachdb/pebble/objstorage/objstorageprovider"
 	"github.com/cockroachdb/pebble/vfs"
 )
@@ -266,8 +267,8 @@ func (tf TestFixtureInfo) Build(fs vfs.FS, filename string) error {
 }
 
 const fixtureDefaultIndexBlockSize = math.MaxInt32
-const fixtureSmallIndexBlockSize = 128
-const fixtureBlockSize = 2048
+const fixtureSmallIndexBlockSize = 128 + cache.ValueMetadataSize
+const fixtureBlockSize = 2048 + cache.ValueMetadataSize
 const fixtureFormat = TableFormatPebblev1
 
 var fixtureComparer = func() *Comparer {

sstable/testdata/flush_heuristics

@@ -0,0 +1,54 @@
+# Block size exceeds target block size.
+build key-size=0 val-size=0 block-size=64 target-size=64 threshold=59
+----
+true
+
+# Block size does not exceed threshold size.
+build key-size=0 val-size=0 block-size=59 target-size=64 threshold=59
+----
+false
+
+# New block size exceeds the target size.
+build key-size=1 val-size=1 block-size=60 target-size=64 threshold=32
+----
+true
+
+# New block size does not exceed the target size.
+build key-size=1 val-size=1 block-size=40 target-size=64 threshold=32
+----
+false
+
+# New block size does not exceed the target size with hints enabled.
+build key-size=1 val-size=1 block-size=36 target-size=64 threshold=0 hints=8,16,32,64,128
+----
+false
+
+# New block size reduces internal fragmentation.
+build key-size=1 val-size=60 block-size=38 target-size=64 threshold=0 hints=8,16,32,64,128
+----
+false
+
+# New block size increases internal fragmentation.
+build key-size=1 val-size=40 block-size=38 target-size=64 threshold=0 hints=8,16,32,64,128
+----
+true
+
+# Block size target exceeded with hints enabled.
+build key-size=1 val-size=1 block-size=64 target-size=64 threshold=0 hints=8,16,32,64,128
+----
+true
+
+# Block size target exceeded, however, new block would reduce internal fragmentation.
+build key-size=1 val-size=1 block-size=70 target-size=64 threshold=0 hints=8,16,32,64,128
+----
+false
+
+# Fall back to heuristics with hints disabled when size class is limited.
+build key-size=1 val-size=1 block-size=59 target-size=64 threshold=59 hints=8,16,32
+----
+false
+
+# Flush when new size class could not be computed.
+build key-size=1 val-size=60 block-size=50 target-size=64 threshold=0 hints=8,16,32,64
+----
+true