Add benchmark for disabled zstd concurrency limit

zstd.go

@@ -20,6 +20,7 @@ var zstdAvailableEncoders sync.Map
 var zstdCheckedOutEncoders int
 var zstdMutex = &sync.Mutex{}
 var zstdEncoderReturned = sync.NewCond(zstdMutex)
+var zstdTestingDisableConcurrencyLimit bool
 
 func getZstdEncoderChannel(params ZstdEncoderParams) chan *zstd.Encoder {
 	if c, ok := zstdAvailableEncoders.Load(params); ok {
@@ -47,7 +48,7 @@ func getZstdEncoder(params ZstdEncoderParams) *zstd.Encoder {
 	defer zstdMutex.Unlock()
 
 	limit := runtime.GOMAXPROCS(0)
-	for zstdCheckedOutEncoders >= limit {
+	for zstdCheckedOutEncoders >= limit && !zstdTestingDisableConcurrencyLimit {
 		zstdEncoderReturned.Wait()
 		limit = runtime.GOMAXPROCS(0)
 	}

zstd_test.go

@@ -33,6 +33,9 @@ func BenchmarkZstdMemoryConsumption(b *testing.B) {
 		// drain the buffered encoder so that we get a fresh one for the next run
 		zstdAvailableEncoders.Delete(params)
 	}
+
+	b.ReportMetric(float64(cpus), "(gomaxprocs)")
+	b.ReportMetric(float64(1), "(goroutines)")
 }
 
 // BenchmarkZstdMemoryConsumptionConcurrency benchmarks the memory consumption of the zstd encoder under the following constraints
@@ -49,11 +52,12 @@ func BenchmarkZstdMemoryConsumptionConcurrency(b *testing.B) {
 	}
 
 	cpus := 4
-	goroutines := 100
+	goroutines := 256
 
 	gomaxprocsBackup := runtime.GOMAXPROCS(cpus)
 	defer runtime.GOMAXPROCS(gomaxprocsBackup)
 
+	b.ReportMetric(float64(cpus), "(gomaxprocs)")
 	b.ResetTimer()
 	b.SetBytes(int64(len(buf) * goroutines))
 	b.ReportAllocs()
@@ -74,4 +78,52 @@ func BenchmarkZstdMemoryConsumptionConcurrency(b *testing.B) {
 		}
 		done.Wait()
 	}
+
+	b.ReportMetric(float64(cpus), "(gomaxprocs)")
+	b.ReportMetric(float64(goroutines), "(goroutines)")
+}
+
+// BenchmarkZstdMemoryNoConcurrencyLimit benchmarks the encoder behavior when the concurrency limit is disabled.
+func BenchmarkZstdMemoryNoConcurrencyLimit(b *testing.B) {
+	zstdTestingDisableConcurrencyLimit = true
+	defer func() {
+		zstdTestingDisableConcurrencyLimit = false
+	}()
+
+	params := ZstdEncoderParams{Level: 9}
+	buf := make([]byte, 1024*1024)
+	for i := 0; i < len(buf); i++ {
+		buf[i] = byte((i / 256) + (i * 257))
+	}
+
+	cpus := 4
+	goroutines := 256
+
+	gomaxprocsBackup := runtime.GOMAXPROCS(cpus)
+	defer runtime.GOMAXPROCS(gomaxprocsBackup)
+
+	b.ReportMetric(float64(cpus), "(gomaxprocs)")
+	b.ResetTimer()
+	b.SetBytes(int64(len(buf) * goroutines))
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		// create n goroutines, wait until all start and then signal them to start compressing
+		var start sync.WaitGroup
+		var done sync.WaitGroup
+		start.Add(goroutines)
+		done.Add(goroutines)
+		for j := 0; j < goroutines; j++ {
+			go func() {
+				start.Done()
+				start.Wait()
+				_, _ = zstdCompress(params, nil, buf)
+				done.Done()
+			}()
+			zstdAvailableEncoders.Delete(params)
+		}
+		done.Wait()
+	}
+
+	b.ReportMetric(float64(cpus), "(gomaxprocs)")
+	b.ReportMetric(float64(goroutines), "(goroutines)")
 }