From 58367e4df6a49edb4d422b97c6da5d401c52d4b5 Mon Sep 17 00:00:00 2001
From: Evgenii Stratonikov <e.stratonikov@yadro.com>
Date: Wed, 1 Feb 2023 11:58:16 +0300
Subject: [PATCH] [#2232] pilorama: Merge in-queue batches

To achieve high performance we must choose proper values for both
batch size and delay. For user operations we want to set low delay.
However it would prevent tree synchronization operations to form big
enough batches. For these operations, batching gives the most benefit
not only in terms of on-CPU execution cost, but also by speeding up
transaction persist (`fsync`).
In this commit we try merging batches that are already
_triggered_, but not yet _started to execute_. This way we can still
query batches for execution after the provided delay while also allowing
multiple formed batches to execute faster.

Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
---
 CHANGELOG.md                                |  1 +
 pkg/local_object_storage/pilorama/batch.go  | 21 +++++++++++++++------
 pkg/local_object_storage/pilorama/boltdb.go |  2 ++
 3 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3103208e2..40d86d918 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -15,6 +15,7 @@ Changelog for FrostFS Node
 - `common.PrintVerbose` prints via `cobra.Command.Printf` (#1962)
 - Env prefix in configuration changed to `FROSTFS_*` (#43)
 - Link object is broadcast throughout the whole container now (#57)
+- Pilorama now can merge multiple batches into one (#2231)
 
 ### Fixed
 - Increase payload size metric on shards' `put` operation (#1794)
diff --git a/pkg/local_object_storage/pilorama/batch.go b/pkg/local_object_storage/pilorama/batch.go
index 59712b761..43375ba1b 100644
--- a/pkg/local_object_storage/pilorama/batch.go
+++ b/pkg/local_object_storage/pilorama/batch.go
@@ -10,8 +10,11 @@ import (
 )
 
 type batch struct {
-	forest     *boltForest
-	timer      *time.Timer
+	forest *boltForest
+	timer  *time.Timer
+	// mtx protects timer and operations fields.
+	// Because mtx can be taken inside a transaction,
+	// transactions MUST NOT be executed with the mutex taken to avoid a deadlock.
 	mtx        sync.Mutex
 	start      sync.Once
 	cid        cidSDK.ID
@@ -24,16 +27,12 @@ func (b *batch) trigger() {
 	b.mtx.Lock()
 	if b.timer != nil {
 		b.timer.Stop()
-		b.timer = nil
 	}
 	b.mtx.Unlock()
 	b.start.Do(b.run)
 }
 
 func (b *batch) run() {
-	sort.Slice(b.operations, func(i, j int) bool {
-		return b.operations[i].Time < b.operations[j].Time
-	})
 	fullID := bucketName(b.cid, b.treeID)
 	err := b.forest.db.Update(func(tx *bbolt.Tx) error {
 		bLog, bTree, err := b.forest.getTreeBuckets(tx, fullID)
@@ -41,6 +40,16 @@ func (b *batch) run() {
 			return err
 		}
 
+		b.mtx.Lock()
+		b.timer = nil
+		b.mtx.Unlock()
+
+		// Sorting without a mutex is ok, because we append to this slice only if timer is non-nil.
+		// See (*boltForest).addBatch for details.
+		sort.Slice(b.operations, func(i, j int) bool {
+			return b.operations[i].Time < b.operations[j].Time
+		})
+
 		var lm Move
 		return b.forest.applyOperation(bLog, bTree, b.operations, &lm)
 	})
diff --git a/pkg/local_object_storage/pilorama/boltdb.go b/pkg/local_object_storage/pilorama/boltdb.go
index 0f546ef61..247d07d28 100644
--- a/pkg/local_object_storage/pilorama/boltdb.go
+++ b/pkg/local_object_storage/pilorama/boltdb.go
@@ -377,7 +377,9 @@ func (t *boltForest) addBatch(d CIDDescriptor, treeID string, m *Move, ch chan e
 		results:    []chan<- error{ch},
 		operations: []*Move{m},
 	}
+	b.mtx.Lock()
 	b.timer = time.AfterFunc(t.db.MaxBatchDelay, b.trigger)
+	b.mtx.Unlock()
 	t.batches = append(t.batches, b)
 	t.mtx.Unlock()
 }