Performance drop when object size is large

cherry-pick from https://github.com/mulbc/gosbench/pull/10/files

Signed-off-by: Liang Zheng <zhengliang0901@gmail.com>

Readme.md

@@ -106,6 +106,13 @@ Note: Due to the constant distribution, we will only consider the `_min` values.
 
 This will cause each workers to search for pre-existing files in the buckets `myBucket-0` and `myBucket-1` and read 10 objects from these buckets. If there are less than 10 objects in any of these buckets, some objects will be read multiple times. The object size given in your config will be ignored when reading pre-existing files.
 
+## Cosbench vs Gosbench benchmark comparision
+When a new tool is presented, it’s essential to compare it to existing tools for accuracy. For this reason, we ran a comparision between Cosbench and Gosbench. Both benchmarks were tasked to do a 100% write test and 100% read test on 4KB, 16KB, 256KB, 1MB, 4MB objects for 60 seconds each. The tests were to run on one RGW using S3 protocol in ceph storage clusteri, also run in the test configuration in parallel. Figure below show writing and reading, respectively. From these charts, it’s apparent that the performance metrics for all objects are similar for both tools. 
+
+![Latency](examples/Latency.png)
+
+![Bandwidth](examples/Bandwidth.png)
+
 ## Contributing
 
 * Be aware that this repo uses pre-commit hooks - install them via `pre-commit install`

worker/main.go

@@ -8,6 +8,7 @@ import (
 	"math/rand"
 	"net"
 	"os"
+	"runtime"
 	"sync"
 	"time"
 
@@ -21,6 +22,7 @@ var prometheusPort int
 var debug, trace bool
 
 func init() {
+	runtime.GOMAXPROCS(runtime.NumCPU())
 	log.SetFormatter(&log.TextFormatter{
 		FullTimestamp: true,
 	})

worker/s3.go

@@ -159,17 +159,23 @@ func listObjects(service *s3.S3, prefix string, bucket string) (*s3.ListObjectsV
 	return result, err
 }
 
-func getObject(service *s3.S3, objectName string, bucket string) error {
-	// Create a downloader with the session and custom options
-	downloader := s3manager.NewDownloaderWithClient(service)
-	buf := aws.NewWriteAtBuffer([]byte{})
-	_, err := downloader.DownloadWithContext(ctx, buf, &s3.GetObjectInput{
+func getObject(service *s3.S3, objectName string, bucket string, objectSize uint64) error {
+	// Remove the allocation of buffer
+	result, err := svc.GetObjectWithContext(ctx, &s3.GetObjectInput{
 		Bucket: &bucket,
 		Key:    &objectName,
-	}, func(d *s3manager.Downloader) {
-		d.PartSize = 64 * 1024 * 1024 // 64MB parts
 	})
-	return err
+	if err != nil {
+		return err
+	}
+	numBytes, err := io.Copy(io.Discard, result.Body)
+	if err != nil {
+		return err
+	}
+	if numBytes != int64(objectSize) {
+		return fmt.Errorf("Expected object length %d is not matched to actual object length %d", objectSize, numBytes)
+	}
+	return nil
 }
 
 func deleteObject(service *s3.S3, objectName string, bucket string) error {

worker/workItems.go

@@ -126,7 +126,7 @@ func (op *Stopper) Prepare() error {
 func (op *ReadOperation) Do() error {
 	log.WithField("bucket", op.Bucket).WithField("object", op.ObjectName).WithField("Preexisting?", op.WorksOnPreexistingObject).Debug("Doing ReadOperation")
 	start := time.Now()
-	err := getObject(svc, op.ObjectName, op.Bucket)
+	err := getObject(svc, op.ObjectName, op.Bucket, op.ObjectSize)
 	duration := time.Since(start)
 	promLatency.WithLabelValues(op.TestName, "GET").Observe(float64(duration.Milliseconds()))
 	if err != nil {