Combined writing for FSTree

CHANGELOG.md

@@ -4,6 +4,7 @@ Changelog for NeoFS Node
 ## [Unreleased]
 
 ### Added
+- More effective FSTree writer for HDDs, new configuration options for it (#2814)
 
 ### Fixed
 

cmd/neofs-lens/internal/storage/root.go

@@ -9,7 +9,6 @@
 	engineconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine"
 	shardconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard"
 	fstreeconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/fstree"
-	peapodconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/peapod"
 	"github.com/nspcc-dev/neofs-node/cmd/neofs-node/storage"
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/blobstor"
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/blobstor/fstree"
@@ -119,15 +118,18 @@
 			sCfg.Typ = storagesCfg[i].Type()
 			sCfg.Path = storagesCfg[i].Path()
 			sCfg.Perm = storagesCfg[i].Perm()
+			sCfg.FlushInterval = storagesCfg[i].FlushInterval()
 
 			switch storagesCfg[i].Type() {
 			case fstree.Type:
 				sub := fstreeconfig.From((*config.Config)(storagesCfg[i]))
 				sCfg.Depth = sub.Depth()
 				sCfg.NoSync = sub.NoSync()
+				sCfg.CombinedCountLimit = sub.CombinedCountLimit()
+				sCfg.CombinedSizeLimit = sub.CombinedSizeLimit()
+				sCfg.CombinedSizeThreshold = sub.CombinedSizeThreshold()
 			case peapod.Type:
-				peapodCfg := peapodconfig.From((*config.Config)(storagesCfg[i]))
-				sCfg.FlushInterval = peapodCfg.FlushInterval()
+				// Nothing peapod-specific, but it should work.
 			default:
 				return fmt.Errorf("can't initiate storage. invalid storage type: %s", storagesCfg[i].Type())
 			}
@@ -193,7 +195,11 @@
 						fstree.WithPath(sRead.Path),
 						fstree.WithPerm(sRead.Perm),
 						fstree.WithDepth(sRead.Depth),
-						fstree.WithNoSync(sRead.NoSync)),
+						fstree.WithNoSync(sRead.NoSync),
+						fstree.WithCombinedCountLimit(sRead.CombinedCountLimit),
+						fstree.WithCombinedSizeLimit(sRead.CombinedSizeLimit),
+						fstree.WithCombinedSizeThreshold(sRead.CombinedSizeThreshold),
+						fstree.WithCombinedWriteInterval(sRead.FlushInterval)),
 					Policy: func(_ *objectSDK.Object, data []byte) bool {
 						return true
 					},

cmd/neofs-lens/internal/storage/sanity.go

@@ -13,7 +13,6 @@
 	engineconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine"
 	shardconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard"
 	fstreeconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/fstree"
-	peapodconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/peapod"
 	objectcore "github.com/nspcc-dev/neofs-node/pkg/core/object"
 	commonb "github.com/nspcc-dev/neofs-node/pkg/local_object_storage/blobstor/common"
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/blobstor/compression"
@@ -77,8 +76,7 @@
 			default:
 				return fmt.Errorf("unsupported sub-storage type '%s'", subCfg.Type())
 			case peapod.Type:
-				peapodCfg := peapodconfig.From((*config.Config)(subCfg))
-				sh.p = peapod.New(subCfg.Path(), subCfg.Perm(), peapodCfg.FlushInterval())
+				sh.p = peapod.New(subCfg.Path(), subCfg.Perm(), subCfg.FlushInterval())
 
 				var compressCfg compression.Config
 				err := compressCfg.Init()

cmd/neofs-node/config.go

@@ -23,7 +23,6 @@
 	engineconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine"
 	shardconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard"
 	fstreeconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/fstree"
-	peapodconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/peapod"
 	loggerconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/logger"
 	metricsconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/metrics"
 	morphconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/morph"
@@ -224,15 +223,18 @@
 			sCfg.Typ = storagesCfg[i].Type()
 			sCfg.Path = storagesCfg[i].Path()
 			sCfg.Perm = storagesCfg[i].Perm()
+			sCfg.FlushInterval = storagesCfg[i].FlushInterval()
 
 			switch storagesCfg[i].Type() {
 			case fstree.Type:
 				sub := fstreeconfig.From((*config.Config)(storagesCfg[i]))
 				sCfg.Depth = sub.Depth()
 				sCfg.NoSync = sub.NoSync()
+				sCfg.CombinedCountLimit = sub.CombinedCountLimit()
+				sCfg.CombinedSizeLimit = sub.CombinedSizeLimit()
+				sCfg.CombinedSizeThreshold = sub.CombinedSizeThreshold()
 			case peapod.Type:
-				peapodCfg := peapodconfig.From((*config.Config)(storagesCfg[i]))
-				sCfg.FlushInterval = peapodCfg.FlushInterval()
+				// No specific configs, but it's a valid storage type.
 			default:
 				return fmt.Errorf("invalid storage type: %s", storagesCfg[i].Type())
 			}

cmd/neofs-node/config/engine/config_test.go

@@ -9,7 +9,6 @@ import (
 	engineconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine"
 	shardconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard"
 	fstreeconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/fstree"
-	peapodconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/blobstor/peapod"
 	piloramaconfig "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/engine/shard/pilorama"
 	configtest "github.com/nspcc-dev/neofs-node/cmd/neofs-node/config/test"
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/blobstor/peapod"
@@ -87,11 +86,10 @@ func TestEngineSection(t *testing.T) {
 				require.EqualValues(t, 102400, sc.SmallSizeLimit())
 
 				require.Equal(t, 2, len(ss))
-				ppd := peapodconfig.From((*config.Config)(ss[0]))
 				require.Equal(t, "tmp/0/blob/peapod.db", ss[0].Path())
 				require.EqualValues(t, 0644, ss[0].Perm())
 				require.EqualValues(t, peapod.Type, ss[0].Type())
-				require.EqualValues(t, 10*time.Millisecond, ppd.FlushInterval())
+				require.EqualValues(t, 10*time.Millisecond, ss[0].FlushInterval())
 
 				require.Equal(t, "tmp/0/blob", ss[1].Path())
 				require.EqualValues(t, 0644, ss[1].Perm())
@@ -131,11 +129,10 @@ func TestEngineSection(t *testing.T) {
 				require.EqualValues(t, 102400, sc.SmallSizeLimit())
 
 				require.Equal(t, 2, len(ss))
-				ppd := peapodconfig.From((*config.Config)(ss[0]))
 				require.Equal(t, "tmp/1/blob/peapod.db", ss[0].Path())
 				require.EqualValues(t, 0644, ss[0].Perm())
 				require.EqualValues(t, peapod.Type, ss[0].Type())
-				require.EqualValues(t, 30*time.Millisecond, ppd.FlushInterval())
+				require.EqualValues(t, 30*time.Millisecond, ss[0].FlushInterval())
 
 				require.Equal(t, "tmp/1/blob", ss[1].Path())
 				require.EqualValues(t, 0644, ss[1].Perm())

cmd/neofs-node/config/engine/shard/blobstor/fstree/config.go

@@ -1,16 +1,27 @@
 package fstree
 
 import (
+	"math"
+
 	"github.com/nspcc-dev/neofs-node/cmd/neofs-node/config"
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/blobstor/fstree"
+	"github.com/spf13/cast"
 )
 
 // Config is a wrapper over the config section
 // which provides access to FSTree configurations.
 type Config config.Config
 
-// DepthDefault is a default shallow dir depth.
-const DepthDefault = 4
+const (
+	// DepthDefault is the default shallow dir depth.
+	DepthDefault = 4
+	// CombinedCountLimitDefault is the default for the maximum number of objects to write into a single file.
+	CombinedCountLimitDefault = 128
+	// CombinedSizeLimitDefault is the default for the maximum size of the combined object file.
+	CombinedSizeLimitDefault = 8 * 1024 * 1024
+	// CombinedSizeThresholdDefault is the default for the minimal size of the object that won't be combined with others for writes.
+	CombinedSizeThresholdDefault = 128 * 1024
+)
 
 // From wraps config section into Config.
 func From(c *config.Config) *Config {
@@ -45,3 +56,41 @@
 func (x *Config) NoSync() bool {
 	return config.BoolSafe((*config.Config)(x), "no_sync")
 }
+
+// CombinedCountLimit returns the value of "combined_count_limit" config parameter.
+//
+// Returns [CombinedCountLimitDefault] if the value is missing or not a positive integer.
+func (x *Config) CombinedCountLimit() int {
+	var v = (*config.Config)(x).Value("combined_count_limit")
+	if v == nil {
+		return CombinedCountLimitDefault
+	}
+
+	i, err := cast.ToIntE(v)
+	if err != nil {
+		return CombinedCountLimitDefault
+	}
+	return i
+}
+
+// CombinedSizeLimit returns the value of "combined_size_limit" config parameter.
+//
+// Returns [CombinedSizeLimitDefault] if the value is missing, equal to 0 or not a proper size specification.
+func (x *Config) CombinedSizeLimit() int {
+	var s = config.SizeInBytesSafe((*config.Config)(x), "combined_size_limit")
+	if s == 0 || s > math.MaxInt {
+		return CombinedSizeLimitDefault
+	}
+	return int(s)
+}
+
+// CombinedSizeThreshold returns the value of "combined_size_threshold" config parameter.
+//
+// Returns [CombinedSizeThresholdDefault] if the value is missing, equal to 0 or not a proper size specification.
+func (x *Config) CombinedSizeThreshold() int {
+	var s = config.SizeInBytesSafe((*config.Config)(x), "combined_size_threshold")
+	if s == 0 || s > math.MaxInt {
+		return CombinedSizeThresholdDefault
+	}
+	return int(s)
+}

cmd/neofs-node/config/engine/shard/blobstor/storage/config.go

@@ -2,14 +2,22 @@ package storage
 
 import (
 	"io/fs"
+	"time"
 
 	"github.com/nspcc-dev/neofs-node/cmd/neofs-node/config"
 )
 
 type Config config.Config
 
-// PermDefault are default permission bits for BlobStor data.
-const PermDefault = 0o640
+// Various config defaults.
+const (
+	// PermDefault are default permission bits for BlobStor data.
+	PermDefault = 0o640
+
+	// DefaultFlushInterval is the default time interval between Peapod's batch writes
+	// to disk.
+	DefaultFlushInterval = 10 * time.Millisecond
+)
 
 func From(x *config.Config) *Config {
 	return (*Config)(x)
@@ -53,3 +61,14 @@ func (x *Config) Perm() fs.FileMode {
 
 	return fs.FileMode(p)
 }
+
+// FlushInterval returns the value of "flush_interval" config parameter.
+//
+// Returns DefaultFlushInterval if the value is not a positive duration.
+func (x *Config) FlushInterval() time.Duration {
+	d := config.DurationSafe((*config.Config)(x), "flush_interval")
+	if d > 0 {
+		return d
+	}
+	return DefaultFlushInterval
+}

cmd/neofs-node/storage.go

@@ -153,7 +153,11 @@
 						fstree.WithPath(sRead.Path),
 						fstree.WithPerm(sRead.Perm),
 						fstree.WithDepth(sRead.Depth),
-						fstree.WithNoSync(sRead.NoSync)),
+						fstree.WithNoSync(sRead.NoSync),
+						fstree.WithCombinedCountLimit(sRead.CombinedCountLimit),
+						fstree.WithCombinedSizeLimit(sRead.CombinedSizeLimit),
+						fstree.WithCombinedSizeThreshold(sRead.CombinedSizeThreshold),
+						fstree.WithCombinedWriteInterval(sRead.FlushInterval)),
 					Policy: func(_ *objectSDK.Object, data []byte) bool {
 						return true
 					},

cmd/neofs-node/storage/config.go

@@ -53,16 +53,17 @@ type ShardCfg struct {
 }
 type SubStorageCfg struct {
 	// common for all storages
-	Typ  string
-	Path string
-	Perm fs.FileMode
+	Typ           string
+	Path          string
+	Perm          fs.FileMode
+	FlushInterval time.Duration
 
 	// tree-specific (FS)
-	Depth  uint64
-	NoSync bool
-
-	// Peapod-specific
-	FlushInterval time.Duration
+	Depth                 uint64
+	NoSync                bool
+	CombinedCountLimit    int
+	CombinedSizeLimit     int
+	CombinedSizeThreshold int
 }
 
 // ID returns persistent id of a shard. It is different from the ID used in runtime

config/example/node.json

@@ -199,7 +199,11 @@
             "path": "tmp/1/blob",
             "no_sync": true,
             "perm": "0644",
-            "depth": 5
+            "depth": 5,
+            "flush_interval": "20ms",
+            "combined_count_limit": 64,
+            "combined_size_limit": "16M",
+            "combined_size_threshold": "512K"
           }
         ],
         "pilorama": {

config/example/node.yaml

@@ -196,6 +196,10 @@ storage:
         - type: fstree
           path: tmp/1/blob  # blobstor path
           no_sync: true
+          flush_interval: 20ms # time interval between combined file writes to disk (defaults to 10ms)
+          combined_count_limit: 64 # number of small objects to write into a single file (defaults to 128)
+          combined_size_limit: 16M # limit for the multi-object file size (defaults to 8M)
+          combined_size_threshold: 512K # threshold for combined object writing (defaults to 128K)
 
       pilorama:
         path: tmp/1/blob/pilorama.db

docs/storage-node-configuration.md

@@ -180,36 +180,35 @@ Currently only 2 types are supported: `fstree` and `peapod`.
 blobstor:
   - type: peapod
     path: /path/to/peapod.db
-    depth: 1
-    width: 4
   - type: fstree
     path: /path/to/blobstor
     perm: 0644
-    size: 4194304
     depth: 1
-    width: 4
-    opened_cache_capacity: 50
 ```
 
 #### Common options for sub-storages
 | Parameter                           | Type                                          | Default value | Description                                                                                                                                                                                                       |
 |-------------------------------------|-----------------------------------------------|---------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
 | `path`                              | `string`                                      |               | Path to the root of the blobstor.                                                                                                                                                                                 |
 | `perm`                              | file mode                                     | `0640`        | Default permission for created files and directories.                                                                                                                                                             |
+| `flush_interval`                    | `duration`                                    | `10ms`        | Time interval between batch writes to disk.                                                                                                                                                                       |
 
 #### `fstree` type options
-| Parameter           | Type      | Default value | Description                                           |
-|---------------------|-----------|---------------|-------------------------------------------------------|
-| `path`              | `string`  |               | Path to the root of the blobstor.                     |
-| `perm`              | file mode | `0640`        | Default permission for created files and directories. |
-| `depth`             | `int`     | `4`           | File-system tree depth.                               |
+| Parameter                 | Type      | Default value | Description                                                                                                                  |
+|---------------------------|-----------|---------------|------------------------------------------------------------------------------------------------------------------------------|
+| `path`                    | `string`  |               | Path to the root of the blobstor.                                                                                            |
+| `perm`                    | file mode | `0640`        | Default permission for created files and directories.                                                                        |
+| `depth`                   | `int`     | `4`           | File-system tree depth.                                                                                                      |
+| `no_sync`                 | `bool`    | `false`       | Disable write synchronization, makes writes faster, but can lead to data loss.                                               |
+| `combined_count_limit`    | `int`     | `128`         | Maximum number of objects to write into a single file, 0 or 1 disables combined writing (disabling is recommended for SSDs). |
+| `combined_size_limit`     | `size`    | `8M`          | Maximum size of a multi-object file.                                                                                         |
+| `combined_size_threshold` | `size`    | `128K`        | Minimum size of object that won't be combined with others when writing to disk.                                              |
 
 #### `peapod` type options
 | Parameter           | Type      | Default value | Description                                           |
 |---------------------|-----------|---------------|-------------------------------------------------------|
 | `path`              | `string`  |               | Path to the Peapod database file.                     |
 | `perm`              | file mode | `0640`        | Default permission for created files and directories. |
-| `flush_interval`    | `duration`| `10ms`        | Time interval between batch writes to disk.           |
 
 ### `gc` subsection