Simplify profiling and benchmarks.

- Tweak the benchmarks to work with b.N where appropriate. In many cases,
  b.N was simply being ignored. This creates an implicit dependency in the
  user passing a reasonable benchtime (less than or equal to the actual
  runtime of the test, or using the X syntax) otherwise the test runs
  forever.
- In cases where the above is impossible, explicitly set benchtime from
  the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
  just add a hey benchmark. The platforms benchmarks already include a
  native platform, and thus these benchmarks are incredibly confusing.
  (In other words, BenchmarkNginxReverse has nothing to do with an nginx
  benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
  to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
  runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
  need to be started early, and others will need to happen at the end.

PiperOrigin-RevId: 349495377

21 files changed, 228 insertions, 318 deletions

diff --git a/test/benchmarks/README.md b/test/benchmarks/README.md
index d1bbabf6f..1bfb4a129 100644
--- a/test/benchmarks/README.md
+++ b/test/benchmarks/README.md
@@ -81,11 +81,8 @@ benchmarks.
 In general, benchmarks should look like this:
 
 ```golang
-
-var h harness.Harness
-
 func BenchmarkMyCoolOne(b *testing.B) {
-  machine, err := h.GetMachine()
+  machine, err := harness.GetMachine()
   // check err
   defer machine.CleanUp()
 
@@ -95,14 +92,14 @@ func BenchmarkMyCoolOne(b *testing.B) {
 
   b.ResetTimer()
 
-  //Respect b.N.
+  // Respect b.N.
   for i := 0; i < b.N; i++ {
     out, err := container.Run(ctx, dockerutil.RunOpts{
       Image: "benchmarks/my-cool-image",
       Env: []string{"MY_VAR=awesome"},
       other options...see dockerutil
     }, "sh", "-c", "echo MY_VAR")
-    //check err
+    // check err...
     b.StopTimer()
 
     // Do parsing and reporting outside of the timer.
@@ -114,16 +111,13 @@ func BenchmarkMyCoolOne(b *testing.B) {
 }
 
 func TestMain(m *testing.M) {
-    h.Init()
+    harness.Init()
     os.Exit(m.Run())
 }
 ```
 
 Some notes on the above:
 
-*   The harness is initiated in the TestMain method and made global to test
-    module. The harness will handle any presetup that needs to happen with
-    flags, remote virtual machines (eventually), and other services.
 *   Respect `b.N` in that users of the benchmark may want to "run for an hour"
     or something of the sort.
 *   Use the `b.ReportMetric()` method to report custom metrics.
diff --git a/test/benchmarks/base/size_test.go b/test/benchmarks/base/size_test.go
index acc49cc7c..452926e5f 100644
--- a/test/benchmarks/base/size_test.go
+++ b/test/benchmarks/base/size_test.go
@@ -26,12 +26,10 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var testHarness harness.Harness
-
 // BenchmarkSizeEmpty creates N empty containers and reads memory usage from
 // /proc/meminfo.
 func BenchmarkSizeEmpty(b *testing.B) {
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -81,7 +79,7 @@ func BenchmarkSizeEmpty(b *testing.B) {
 // BenchmarkSizeNginx starts N containers running Nginx, checks that they're
 // serving, and checks memory used based on /proc/meminfo.
 func BenchmarkSizeNginx(b *testing.B) {
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -126,7 +124,7 @@ func BenchmarkSizeNginx(b *testing.B) {
 // BenchmarkSizeNode starts N containers running a Node app, checks that
 // they're serving, and checks memory used based on /proc/meminfo.
 func BenchmarkSizeNode(b *testing.B) {
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -178,6 +176,6 @@ func BenchmarkSizeNode(b *testing.B) {
 
 // TestMain is the main method for package network.
 func TestMain(m *testing.M) {
-	testHarness.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/base/startup_test.go b/test/benchmarks/base/startup_test.go
index 8ef9f99c4..05a43ad17 100644
--- a/test/benchmarks/base/startup_test.go
+++ b/test/benchmarks/base/startup_test.go
@@ -25,11 +25,9 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/harness"
 )
 
-var testHarness harness.Harness
-
 // BenchmarkStartEmpty times startup time for an empty container.
 func BenchmarkStartupEmpty(b *testing.B) {
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -53,7 +51,7 @@ func BenchmarkStartupEmpty(b *testing.B) {
 // Time is measured from start until the first request is served.
 func BenchmarkStartupNginx(b *testing.B) {
 	// The machine to hold Nginx and the Node Server.
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -76,7 +74,7 @@ func BenchmarkStartupNginx(b *testing.B) {
 // Time is measured from start until the first request is served.
 // Note that the Node app connects to a Redis instance before serving.
 func BenchmarkStartupNode(b *testing.B) {
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -126,8 +124,8 @@ func runServerWorkload(ctx context.Context, b *testing.B, args base.ServerArgs)
 				return fmt.Errorf("failed to get ip from server: %v", err)
 			}
 
-			harness.DebugLog(b, "Waiting for container to start.")
 			// Wait until the Client sees the server as up.
+			harness.DebugLog(b, "Waiting for container to start.")
 			if err := harness.WaitUntilServing(ctx, args.Machine, servingIP, args.Port); err != nil {
 				return fmt.Errorf("failed to wait for serving: %v", err)
 			}
@@ -141,6 +139,6 @@ func runServerWorkload(ctx context.Context, b *testing.B, args base.ServerArgs)
 
 // TestMain is the main method for package network.
 func TestMain(m *testing.M) {
-	testHarness.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/base/sysbench_test.go b/test/benchmarks/base/sysbench_test.go
index bbb797e14..80569687c 100644
--- a/test/benchmarks/base/sysbench_test.go
+++ b/test/benchmarks/base/sysbench_test.go
@@ -23,8 +23,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var testHarness harness.Harness
-
 type testCase struct {
 	name string
 	test tools.Sysbench
@@ -32,42 +30,34 @@ type testCase struct {
 
 // BenchmarSysbench runs sysbench on the runtime.
 func BenchmarkSysbench(b *testing.B) {
-
 	testCases := []testCase{
 		testCase{
 			name: "CPU",
 			test: &tools.SysbenchCPU{
-				Base: tools.SysbenchBase{
+				SysbenchBase: tools.SysbenchBase{
 					Threads: 1,
-					Time:    5,
 				},
-				MaxPrime: 50000,
 			},
 		},
 		testCase{
 			name: "Memory",
 			test: &tools.SysbenchMemory{
-				Base: tools.SysbenchBase{
+				SysbenchBase: tools.SysbenchBase{
 					Threads: 1,
 				},
-				BlockSize: "1M",
-				TotalSize: "500G",
 			},
 		},
 		testCase{
 			name: "Mutex",
 			test: &tools.SysbenchMutex{
-				Base: tools.SysbenchBase{
+				SysbenchBase: tools.SysbenchBase{
 					Threads: 8,
 				},
-				Loops: 1,
-				Locks: 10000000,
-				Num:   4,
 			},
 		},
 	}
 
-	machine, err := testHarness.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -87,12 +77,15 @@ func BenchmarkSysbench(b *testing.B) {
 			sysbench := machine.GetContainer(ctx, b)
 			defer sysbench.CleanUp(ctx)
 
+			cmd := tc.test.MakeCmd(b)
+			b.ResetTimer()
 			out, err := sysbench.Run(ctx, dockerutil.RunOpts{
 				Image: "benchmarks/sysbench",
-			}, tc.test.MakeCmd()...)
+			}, cmd...)
 			if err != nil {
 				b.Fatalf("failed to run sysbench: %v: logs:%s", err, out)
 			}
+			b.StopTimer()
 			tc.test.Report(b, out)
 		})
 	}
diff --git a/test/benchmarks/database/redis_test.go b/test/benchmarks/database/redis_test.go
index f8075a04b..f3c4522ac 100644
--- a/test/benchmarks/database/redis_test.go
+++ b/test/benchmarks/database/redis_test.go
@@ -25,8 +25,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // All possible operations from redis. Note: "ping" will
 // run both PING_INLINE and PING_BUILD.
 var operations []string = []string{
@@ -52,13 +50,13 @@ var operations []string = []string{
 // BenchmarkRedis runs redis-benchmark against a redis instance and reports
 // data in queries per second. Each is reported by named operation (e.g. LPUSH).
 func BenchmarkRedis(b *testing.B) {
-	clientMachine, err := h.GetMachine()
+	clientMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
 	defer clientMachine.CleanUp()
 
-	serverMachine, err := h.GetMachine()
+	serverMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -67,7 +65,6 @@ func BenchmarkRedis(b *testing.B) {
 	// Redis runs on port 6379 by default.
 	port := 6379
 	ctx := context.Background()
-
 	for _, operation := range operations {
 		param := tools.Parameter{
 			Name:  "operation",
@@ -107,23 +104,19 @@ func BenchmarkRedis(b *testing.B) {
 				b.Fatalf("failed to start redis with: %v", err)
 			}
 
+			client := clientMachine.GetNativeContainer(ctx, b)
+			defer client.CleanUp(ctx)
+
 			redis := tools.Redis{
 				Operation: operation,
 			}
-
-			// Reset profiles and timer to begin the measurement.
-			server.RestartProfiles()
 			b.ResetTimer()
-			client := clientMachine.GetNativeContainer(ctx, b)
-			defer client.CleanUp(ctx)
 			out, err := client.Run(ctx, dockerutil.RunOpts{
 				Image: "benchmarks/redis",
 			}, redis.MakeCmd(ip, serverPort, b.N /*requests*/)...)
 			if err != nil {
 				b.Fatalf("redis-benchmark failed with: %v", err)
 			}
-
-			// Stop time while we parse results.
 			b.StopTimer()
 			redis.Report(b, out)
 		})
@@ -131,6 +124,6 @@ func BenchmarkRedis(b *testing.B) {
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/fs/bazel_test.go b/test/benchmarks/fs/bazel_test.go
index 3fb4da9d1..8baeff0db 100644
--- a/test/benchmarks/fs/bazel_test.go
+++ b/test/benchmarks/fs/bazel_test.go
@@ -25,8 +25,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // Note: CleanCache versions of this test require running with root permissions.
 func BenchmarkBuildABSL(b *testing.B) {
 	runBuildBenchmark(b, "benchmarks/absl", "/abseil-cpp", "absl/base/...")
@@ -41,7 +39,7 @@ func BenchmarkBuildRunsc(b *testing.B) {
 func runBuildBenchmark(b *testing.B, image, workdir, target string) {
 	b.Helper()
 	// Get a machine from the Harness on which to run.
-	machine, err := h.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -102,21 +100,20 @@ func runBuildBenchmark(b *testing.B, image, workdir, target string) {
 				prefix = "/tmp"
 			}
 
-			// Restart profiles after the copy.
-			container.RestartProfiles()
 			b.ResetTimer()
+			b.StopTimer()
+
 			// Drop Caches and bazel clean should happen inside the loop as we may use
 			// time options with b.N. (e.g. Run for an hour.)
 			for i := 0; i < b.N; i++ {
-				b.StopTimer()
 				// Drop Caches for clear cache runs.
 				if bm.clearCache {
 					if err := harness.DropCaches(machine); err != nil {
 						b.Skipf("failed to drop caches: %v. You probably need root.", err)
 					}
 				}
-				b.StartTimer()
 
+				b.StartTimer()
 				got, err := container.Exec(ctx, dockerutil.ExecOpts{
 					WorkDir: prefix + workdir,
 				}, "bazel", "build", "-c", "opt", target)
@@ -138,7 +135,6 @@ func runBuildBenchmark(b *testing.B, image, workdir, target string) {
 						b.Fatalf("build failed with: %v", err)
 					}
 				}
-				b.StartTimer()
 			}
 		})
 	}
@@ -146,6 +142,7 @@ func runBuildBenchmark(b *testing.B, image, workdir, target string) {
 
 // TestMain is the main method for package fs.
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
+	harness.SetFixedBenchmarks()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/fs/fio_test.go b/test/benchmarks/fs/fio_test.go
index 96340373c..0c772b768 100644
--- a/test/benchmarks/fs/fio_test.go
+++ b/test/benchmarks/fs/fio_test.go
@@ -27,42 +27,50 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // BenchmarkFio runs fio on the runtime under test. There are 4 basic test
 // cases each run on a tmpfs mount and a bind mount. Fio requires root so that
 // caches can be dropped.
 func BenchmarkFio(b *testing.B) {
 	testCases := []tools.Fio{
 		tools.Fio{
-			Test:      "write",
-			Size:      "5G",
-			Blocksize: "1M",
-			Iodepth:   4,
+			Test:      "write4K",
+			Size:      b.N,
+			BlockSize: 4,
+			IODepth:   4,
+		},
+		tools.Fio{
+			Test:      "write1M",
+			Size:      b.N,
+			BlockSize: 1024,
+			IODepth:   4,
+		},
+		tools.Fio{
+			Test:      "read4K",
+			Size:      b.N,
+			BlockSize: 4,
+			IODepth:   4,
 		},
 		tools.Fio{
-			Test:      "read",
-			Size:      "5G",
-			Blocksize: "1M",
-			Iodepth:   4,
+			Test:      "read1M",
+			Size:      b.N,
+			BlockSize: 1024,
+			IODepth:   4,
 		},
 		tools.Fio{
-			Test:      "randwrite",
-			Size:      "5G",
-			Blocksize: "4K",
-			Iodepth:   4,
-			Time:      30,
+			Test:      "randwrite4K",
+			Size:      b.N,
+			BlockSize: 4,
+			IODepth:   4,
 		},
 		tools.Fio{
-			Test:      "randread",
-			Size:      "5G",
-			Blocksize: "4K",
-			Iodepth:   4,
-			Time:      30,
+			Test:      "randread4K",
+			Size:      b.N,
+			BlockSize: 4,
+			IODepth:   4,
 		},
 	}
 
-	machine, err := h.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -116,7 +124,7 @@ func BenchmarkFio(b *testing.B) {
 
 				// For reads, we need a file to read so make one inside the container.
 				if strings.Contains(tc.Test, "read") {
-					fallocateCmd := fmt.Sprintf("fallocate -l %s %s", tc.Size, outfile)
+					fallocateCmd := fmt.Sprintf("fallocate -l %dK %s", tc.Size, outfile)
 					if out, err := container.Exec(ctx, dockerutil.ExecOpts{},
 						strings.Split(fallocateCmd, " ")...); err != nil {
 						b.Fatalf("failed to create readable file on mount: %v, %s", err, out)
@@ -128,22 +136,24 @@ func BenchmarkFio(b *testing.B) {
 					b.Skipf("failed to drop caches with %v. You probably need root.", err)
 				}
 				cmd := tc.MakeCmd(outfile)
-				container.RestartProfiles()
+
 				b.ResetTimer()
+				b.StopTimer()
+
 				for i := 0; i < b.N; i++ {
+					if err := harness.DropCaches(machine); err != nil {
+						b.Fatalf("failed to drop caches: %v", err)
+					}
+
 					// Run fio.
+					b.StartTimer()
 					data, err := container.Exec(ctx, dockerutil.ExecOpts{}, cmd...)
 					if err != nil {
 						b.Fatalf("failed to run cmd %v: %v", cmd, err)
 					}
 					b.StopTimer()
+					b.SetBytes(1024 * 1024) // Bytes for go reporting (Size is in megabytes).
 					tc.Report(b, data)
-					// If b.N is used (i.e. we run for an hour), we should drop caches
-					// after each run.
-					if err := harness.DropCaches(machine); err != nil {
-						b.Fatalf("failed to drop caches: %v", err)
-					}
-					b.StartTimer()
 				}
 			})
 		}
@@ -185,6 +195,6 @@ func makeMount(machine harness.Machine, mountType mount.Type, target string) (mo
 
 // TestMain is the main method for package fs.
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/harness/harness.go b/test/benchmarks/harness/harness.go
index 4c6e724aa..a853b7ba8 100644
--- a/test/benchmarks/harness/harness.go
+++ b/test/benchmarks/harness/harness.go
@@ -28,12 +28,8 @@ var (
 	debug = flag.Bool("debug", false, "turns on debug messages for individual benchmarks")
 )
 
-// Harness is a handle for managing state in benchmark runs.
-type Harness struct {
-}
-
 // Init performs any harness initilialization before runs.
-func (h *Harness) Init() error {
+func Init() error {
 	flag.Usage = func() {
 		fmt.Fprintf(os.Stderr, "Usage: %s -- --test.bench=<regex>\n", os.Args[0])
 		flag.PrintDefaults()
@@ -47,7 +43,15 @@ func (h *Harness) Init() error {
 	return nil
 }
 
+// SetFixedBenchmarks causes all benchmarks to run once.
+//
+// This must be set if they cannot scale with N. Note that this uses 1ns
+// instead of 1x due to https://github.com/golang/go/issues/32051.
+func SetFixedBenchmarks() {
+	flag.Set("test.benchtime", "1ns")
+}
+
 // GetMachine returns this run's implementation of machine.
-func (h *Harness) GetMachine() (Machine, error) {
+func GetMachine() (Machine, error) {
 	return &localMachine{}, nil
 }
diff --git a/test/benchmarks/media/ffmpeg_test.go b/test/benchmarks/media/ffmpeg_test.go
index a462ec2a6..1b99a319a 100644
--- a/test/benchmarks/media/ffmpeg_test.go
+++ b/test/benchmarks/media/ffmpeg_test.go
@@ -23,12 +23,10 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/harness"
 )
 
-var h harness.Harness
-
 // BenchmarkFfmpeg runs ffmpeg in a container and records runtime.
 // BenchmarkFfmpeg should run as root to drop caches.
 func BenchmarkFfmpeg(b *testing.B) {
-	machine, err := h.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -38,24 +36,26 @@ func BenchmarkFfmpeg(b *testing.B) {
 	cmd := strings.Split("ffmpeg -i video.mp4 -c:v libx264 -preset veryslow output.mp4", " ")
 
 	b.ResetTimer()
+	b.StopTimer()
+
 	for i := 0; i < b.N; i++ {
-		b.StopTimer()
 		container := machine.GetContainer(ctx, b)
 		defer container.CleanUp(ctx)
 		if err := harness.DropCaches(machine); err != nil {
 			b.Skipf("failed to drop caches: %v. You probably need root.", err)
 		}
-		b.StartTimer()
 
+		b.StartTimer()
 		if _, err := container.Run(ctx, dockerutil.RunOpts{
 			Image: "benchmarks/ffmpeg",
 		}, cmd...); err != nil {
 			b.Fatalf("failed to run container: %v", err)
 		}
+		b.StopTimer()
 	}
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/ml/tensorflow_test.go b/test/benchmarks/ml/tensorflow_test.go
index a55329d82..b0e0c4720 100644
--- a/test/benchmarks/ml/tensorflow_test.go
+++ b/test/benchmarks/ml/tensorflow_test.go
@@ -22,8 +22,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/harness"
 )
 
-var h harness.Harness
-
 // BenchmarkTensorflow runs workloads from a TensorFlow tutorial.
 // See: https://github.com/aymericdamien/TensorFlow-Examples
 func BenchmarkTensorflow(b *testing.B) {
@@ -38,7 +36,7 @@ func BenchmarkTensorflow(b *testing.B) {
 		"NeuralNetwork":        "3_NeuralNetworks/neural_network.py",
 	}
 
-	machine, err := h.GetMachine()
+	machine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -49,15 +47,17 @@ func BenchmarkTensorflow(b *testing.B) {
 			ctx := context.Background()
 
 			b.ResetTimer()
+			b.StopTimer()
+
 			for i := 0; i < b.N; i++ {
-				b.StopTimer()
 				container := machine.GetContainer(ctx, b)
 				defer container.CleanUp(ctx)
 				if err := harness.DropCaches(machine); err != nil {
 					b.Skipf("failed to drop caches: %v. You probably need root.", err)
 				}
-				b.StartTimer()
 
+				// Run tensorflow.
+				b.StartTimer()
 				if out, err := container.Run(ctx, dockerutil.RunOpts{
 					Image:   "benchmarks/tensorflow",
 					Env:     []string{"PYTHONPATH=$PYTHONPATH:/TensorFlow-Examples/examples"},
@@ -65,13 +65,14 @@ func BenchmarkTensorflow(b *testing.B) {
 				}, "python", workload); err != nil {
 					b.Fatalf("failed to run container: %v logs: %s", err, out)
 				}
+				b.StopTimer()
 			}
 		})
 	}
-
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
+	harness.SetFixedBenchmarks()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/network/httpd_test.go b/test/benchmarks/network/httpd_test.go
index b07274662..629127250 100644
--- a/test/benchmarks/network/httpd_test.go
+++ b/test/benchmarks/network/httpd_test.go
@@ -23,8 +23,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // see Dockerfile '//images/benchmarks/httpd'.
 var httpdDocs = map[string]string{
 	"notfound": "notfound",
@@ -38,13 +36,7 @@ var httpdDocs = map[string]string{
 // BenchmarkHttpd iterates over different sized payloads and concurrency, testing
 // how well the runtime handles sending different payload sizes.
 func BenchmarkHttpd(b *testing.B) {
-	benchmarkHttpdDocSize(b, false /* reverse */)
-}
-
-// BenchmarkReverseHttpd iterates over different sized payloads, testing
-// how well the runtime handles receiving different payload sizes.
-func BenchmarkReverseHttpd(b *testing.B) {
-	benchmarkHttpdDocSize(b, true /* reverse */)
+	benchmarkHttpdDocSize(b)
 }
 
 // BenchmarkContinuousHttpd runs specific benchmarks for continous jobs.
@@ -52,20 +44,12 @@ func BenchmarkReverseHttpd(b *testing.B) {
 func BenchmarkContinuousHttpd(b *testing.B) {
 	sizes := []string{"10Kb", "100Kb", "1Mb"}
 	threads := []int{1, 25, 100, 1000}
-	benchmarkHttpdContinuous(b, threads, sizes, false /*reverse*/)
-}
-
-// BenchmarkContinuousHttpdReverse runs specific benchmarks for continous jobs.
-// The runtime under test is the client downloading from a runc server.
-func BenchmarkContinuousHttpdReverse(b *testing.B) {
-	sizes := []string{"10Kb", "100Kb", "1Mb"}
-	threads := []int{1, 25, 100, 1000}
-	benchmarkHttpdContinuous(b, threads, sizes, true /*reverse*/)
+	benchmarkHttpdContinuous(b, threads, sizes)
 }
 
 // benchmarkHttpdDocSize iterates through all doc sizes, running subbenchmarks
 // for each size.
-func benchmarkHttpdDocSize(b *testing.B, reverse bool) {
+func benchmarkHttpdDocSize(b *testing.B) {
 	b.Helper()
 	for size, filename := range httpdDocs {
 		concurrency := []int{1, 25, 50, 100, 1000}
@@ -82,25 +66,20 @@ func benchmarkHttpdDocSize(b *testing.B, reverse bool) {
 			if err != nil {
 				b.Fatalf("Failed to parse parameters: %v", err)
 			}
-			requests := b.N
-			if requests < c {
-				b.Logf("b.N is %d must be greater than threads %d. Consider running with --test.benchtime=Nx where N >= %d", b.N, c, c)
-				requests = c
-			}
 			b.Run(name, func(b *testing.B) {
 				hey := &tools.Hey{
-					Requests:    requests,
+					Requests:    b.N,
 					Concurrency: c,
 					Doc:         filename,
 				}
-				runHttpd(b, hey, reverse)
+				runHttpd(b, hey)
 			})
 		}
 	}
 }
 
 // benchmarkHttpdContinuous iterates through given sizes and concurrencies.
-func benchmarkHttpdContinuous(b *testing.B, concurrency []int, sizes []string, reverse bool) {
+func benchmarkHttpdContinuous(b *testing.B, concurrency []int, sizes []string) {
 	for _, size := range sizes {
 		filename := httpdDocs[size]
 		for _, c := range concurrency {
@@ -118,26 +97,20 @@ func benchmarkHttpdContinuous(b *testing.B, concurrency []int, sizes []string, r
 			if err != nil {
 				b.Fatalf("Failed to parse parameters: %v", err)
 			}
-
-			requests := b.N
-			if requests < c {
-				b.Logf("b.N is %d must be greater than threads %d. Consider running with --test.benchtime=Nx where N >= %d", b.N, c, c)
-				requests = c
-			}
 			b.Run(name, func(b *testing.B) {
 				hey := &tools.Hey{
-					Requests:    requests,
+					Requests:    b.N,
 					Concurrency: c,
 					Doc:         filename,
 				}
-				runHttpd(b, hey, reverse)
+				runHttpd(b, hey)
 			})
 		}
 	}
 }
 
 // runHttpd configures the static serving methods to run httpd.
-func runHttpd(b *testing.B, hey *tools.Hey, reverse bool) {
+func runHttpd(b *testing.B, hey *tools.Hey) {
 	// httpd runs on port 80.
 	port := 80
 	httpdRunOpts := dockerutil.RunOpts{
@@ -153,10 +126,10 @@ func runHttpd(b *testing.B, hey *tools.Hey, reverse bool) {
 		},
 	}
 	httpdCmd := []string{"sh", "-c", "mkdir -p /tmp/html; cp -r /local/* /tmp/html/.; apache2 -X"}
-	runStaticServer(b, h, httpdRunOpts, httpdCmd, port, hey, reverse)
+	runStaticServer(b, httpdRunOpts, httpdCmd, port, hey)
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/network/iperf_test.go b/test/benchmarks/network/iperf_test.go
index 9d64db943..5e81149fe 100644
--- a/test/benchmarks/network/iperf_test.go
+++ b/test/benchmarks/network/iperf_test.go
@@ -24,20 +24,18 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 func BenchmarkIperf(b *testing.B) {
 	iperf := tools.Iperf{
-		Time: b.N, // time in seconds to run client.
+		Num: b.N,
 	}
 
-	clientMachine, err := h.GetMachine()
+	clientMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
 	defer clientMachine.CleanUp()
 
-	serverMachine, err := h.GetMachine()
+	serverMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
@@ -94,12 +92,9 @@ func BenchmarkIperf(b *testing.B) {
 			if err := harness.WaitUntilServing(ctx, clientMachine, ip, servingPort); err != nil {
 				b.Fatalf("failed to wait for server: %v", err)
 			}
+
 			// Run the client.
 			b.ResetTimer()
-
-			// Restart the server profiles. If the server isn't being profiled
-			// this does nothing.
-			server.RestartProfiles()
 			out, err := client.Run(ctx, dockerutil.RunOpts{
 				Image: "benchmarks/iperf",
 			}, iperf.MakeCmd(ip, servingPort)...)
@@ -113,6 +108,6 @@ func BenchmarkIperf(b *testing.B) {
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/network/network.go b/test/benchmarks/network/network.go
index b18bc2b3c..d61002cea 100644
--- a/test/benchmarks/network/network.go
+++ b/test/benchmarks/network/network.go
@@ -25,33 +25,26 @@ import (
 )
 
 // runStaticServer runs static serving workloads (httpd, nginx).
-func runStaticServer(b *testing.B, h harness.Harness, serverOpts dockerutil.RunOpts, serverCmd []string, port int, hey *tools.Hey, reverse bool) {
+func runStaticServer(b *testing.B, serverOpts dockerutil.RunOpts, serverCmd []string, port int, hey *tools.Hey) {
 	ctx := context.Background()
 
 	// Get two machines: a client and server.
-	clientMachine, err := h.GetMachine()
+	clientMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
 	defer clientMachine.CleanUp()
 
-	serverMachine, err := h.GetMachine()
+	serverMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine: %v", err)
 	}
 	defer serverMachine.CleanUp()
 
-	// Make the containers. 'reverse=true' specifies that the client should use the
-	// runtime under test.
-	var client, server *dockerutil.Container
-	if reverse {
-		client = clientMachine.GetContainer(ctx, b)
-		server = serverMachine.GetNativeContainer(ctx, b)
-	} else {
-		client = clientMachine.GetNativeContainer(ctx, b)
-		server = serverMachine.GetContainer(ctx, b)
-	}
+	// Make the containers.
+	client := clientMachine.GetNativeContainer(ctx, b)
 	defer client.CleanUp(ctx)
+	server := serverMachine.GetContainer(ctx, b)
 	defer server.CleanUp(ctx)
 
 	// Start the server.
@@ -73,16 +66,15 @@ func runStaticServer(b *testing.B, h harness.Harness, serverOpts dockerutil.RunO
 
 	// Make sure the server is serving.
 	harness.WaitUntilServing(ctx, clientMachine, ip, servingPort)
+
+	// Run the client.
 	b.ResetTimer()
-	server.RestartProfiles()
 	out, err := client.Run(ctx, dockerutil.RunOpts{
 		Image: "benchmarks/hey",
 	}, hey.MakeCmd(ip, servingPort)...)
 	if err != nil {
 		b.Fatalf("run failed with: %v", err)
 	}
-
 	b.StopTimer()
 	hey.Report(b, out)
-	b.StartTimer()
 }
diff --git a/test/benchmarks/network/nginx_test.go b/test/benchmarks/network/nginx_test.go
index 87449612a..74f3578fc 100644
--- a/test/benchmarks/network/nginx_test.go
+++ b/test/benchmarks/network/nginx_test.go
@@ -23,8 +23,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // see Dockerfile '//images/benchmarks/nginx'.
 var nginxDocs = map[string]string{
 	"notfound": "notfound",
@@ -38,14 +36,8 @@ var nginxDocs = map[string]string{
 // BenchmarkNginxDocSize iterates over different sized payloads, testing how
 // well the runtime handles sending different payload sizes.
 func BenchmarkNginxDocSize(b *testing.B) {
-	benchmarkNginxDocSize(b, false /* reverse */, true /* tmpfs */)
-	benchmarkNginxDocSize(b, false /* reverse */, false /* tmpfs */)
-}
-
-// BenchmarkReverseNginxDocSize iterates over different sized payloads, testing
-// how well the runtime handles receiving different payload sizes.
-func BenchmarkReverseNginxDocSize(b *testing.B) {
-	benchmarkNginxDocSize(b, true /* reverse */, true /* tmpfs */)
+	benchmarkNginxDocSize(b, true /* tmpfs */)
+	benchmarkNginxDocSize(b, false /* tmpfs */)
 }
 
 // BenchmarkContinuousNginx runs specific benchmarks for continous jobs.
@@ -53,20 +45,12 @@ func BenchmarkReverseNginxDocSize(b *testing.B) {
 func BenchmarkContinuousNginx(b *testing.B) {
 	sizes := []string{"10Kb", "100Kb", "1Mb"}
 	threads := []int{1, 25, 100, 1000}
-	benchmarkNginxContinuous(b, threads, sizes, false /*reverse*/)
-}
-
-// BenchmarkContinuousNginxReverse runs specific benchmarks for continous jobs.
-// The runtime under test is the client downloading from a runc server.
-func BenchmarkContinuousNginxReverse(b *testing.B) {
-	sizes := []string{"10Kb", "100Kb", "1Mb"}
-	threads := []int{1, 25, 100, 1000}
-	benchmarkNginxContinuous(b, threads, sizes, true /*reverse*/)
+	benchmarkNginxContinuous(b, threads, sizes)
 }
 
 // benchmarkNginxDocSize iterates through all doc sizes, running subbenchmarks
 // for each size.
-func benchmarkNginxDocSize(b *testing.B, reverse, tmpfs bool) {
+func benchmarkNginxDocSize(b *testing.B, tmpfs bool) {
 	for size, filename := range nginxDocs {
 		concurrency := []int{1, 25, 50, 100, 1000}
 		for _, c := range concurrency {
@@ -91,26 +75,20 @@ func benchmarkNginxDocSize(b *testing.B, reverse, tmpfs bool) {
 			if err != nil {
 				b.Fatalf("Failed to parse parameters: %v", err)
 			}
-
-			requests := b.N
-			if requests < c {
-				b.Logf("b.N is %d must be greater than threads %d. Consider running with --test.benchtime=Nx where N >= %d", b.N, c, c)
-				requests = c
-			}
 			b.Run(name, func(b *testing.B) {
 				hey := &tools.Hey{
-					Requests:    requests,
+					Requests:    b.N,
 					Concurrency: c,
 					Doc:         filename,
 				}
-				runNginx(b, hey, reverse, tmpfs)
+				runNginx(b, hey, tmpfs)
 			})
 		}
 	}
 }
 
 // benchmarkNginxContinuous iterates through given sizes and concurrencies on a tmpfs mount.
-func benchmarkNginxContinuous(b *testing.B, concurrency []int, sizes []string, reverse bool) {
+func benchmarkNginxContinuous(b *testing.B, concurrency []int, sizes []string) {
 	for _, size := range sizes {
 		filename := nginxDocs[size]
 		for _, c := range concurrency {
@@ -133,25 +111,20 @@ func benchmarkNginxContinuous(b *testing.B, concurrency []int, sizes []string, r
 			if err != nil {
 				b.Fatalf("Failed to parse parameters: %v", err)
 			}
-			requests := b.N
-			if requests < c {
-				b.Logf("b.N is %d must be greater than threads %d. Consider running with --test.benchtime=Nx where N >= %d", b.N, c, c)
-				requests = c
-			}
 			b.Run(name, func(b *testing.B) {
 				hey := &tools.Hey{
-					Requests:    requests,
+					Requests:    b.N,
 					Concurrency: c,
 					Doc:         filename,
 				}
-				runNginx(b, hey, reverse, true /*tmpfs*/)
+				runNginx(b, hey, true /*tmpfs*/)
 			})
 		}
 	}
 }
 
 // runNginx configures the static serving methods to run httpd.
-func runNginx(b *testing.B, hey *tools.Hey, reverse, tmpfs bool) {
+func runNginx(b *testing.B, hey *tools.Hey, tmpfs bool) {
 	// nginx runs on port 80.
 	port := 80
 	nginxRunOpts := dockerutil.RunOpts{
@@ -165,10 +138,10 @@ func runNginx(b *testing.B, hey *tools.Hey, reverse, tmpfs bool) {
 	}
 
 	// Command copies nginxDocs to tmpfs serving directory and runs nginx.
-	runStaticServer(b, h, nginxRunOpts, nginxCmd, port, hey, reverse)
+	runStaticServer(b, nginxRunOpts, nginxCmd, port, hey)
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/network/node_test.go b/test/benchmarks/network/node_test.go
index 3e837a9e4..a1fc82f95 100644
--- a/test/benchmarks/network/node_test.go
+++ b/test/benchmarks/network/node_test.go
@@ -25,8 +25,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // BenchmarkNode runs requests using 'hey' against a Node server run on
 // 'runtime'. The server responds to requests by grabbing some data in a
 // redis instance and returns the data in its reponse. The test loops through
@@ -42,14 +40,9 @@ func BenchmarkNode(b *testing.B) {
 		if err != nil {
 			b.Fatalf("Failed to parse parameters: %v", err)
 		}
-		requests := b.N
-		if requests < c {
-			b.Logf("b.N is %d must be greater than threads %d. Consider running with --test.benchtime=Nx where N >= %d", b.N, c, c)
-			requests = c
-		}
 		b.Run(name, func(b *testing.B) {
 			hey := &tools.Hey{
-				Requests:    requests,
+				Requests:    b.N,
 				Concurrency: c,
 			}
 			runNode(b, hey)
@@ -62,14 +55,14 @@ func runNode(b *testing.B, hey *tools.Hey) {
 	b.Helper()
 
 	// The machine to hold Redis and the Node Server.
-	serverMachine, err := h.GetMachine()
+	serverMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
 	defer serverMachine.CleanUp()
 
 	// The machine to run 'hey'.
-	clientMachine, err := h.GetMachine()
+	clientMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -124,10 +117,8 @@ func runNode(b *testing.B, hey *tools.Hey) {
 
 	heyCmd := hey.MakeCmd(servingIP, servingPort)
 
-	nodeApp.RestartProfiles()
-	b.ResetTimer()
-
 	// the client should run on Native.
+	b.ResetTimer()
 	client := clientMachine.GetNativeContainer(ctx, b)
 	out, err := client.Run(ctx, dockerutil.RunOpts{
 		Image: "benchmarks/hey",
@@ -137,11 +128,10 @@ func runNode(b *testing.B, hey *tools.Hey) {
 	}
 
 	// Stop the timer to parse the data and report stats.
-	b.StopTimer()
 	hey.Report(b, out)
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/network/ruby_test.go b/test/benchmarks/network/ruby_test.go
index c89672873..b7ec16e0a 100644
--- a/test/benchmarks/network/ruby_test.go
+++ b/test/benchmarks/network/ruby_test.go
@@ -26,8 +26,6 @@ import (
 	"gvisor.dev/gvisor/test/benchmarks/tools"
 )
 
-var h harness.Harness
-
 // BenchmarkRuby runs requests using 'hey' against a ruby application server.
 // On start, ruby app generates some random data and pushes it to a redis
 // instance. On a request, the app grabs for random entries from the redis
@@ -43,14 +41,9 @@ func BenchmarkRuby(b *testing.B) {
 		if err != nil {
 			b.Fatalf("Failed to parse parameters: %v", err)
 		}
-		requests := b.N
-		if requests < c {
-			b.Logf("b.N is %d must be greater than threads %d. Consider running with --test.benchtime=Nx where N >= %d", b.N, c, c)
-			requests = c
-		}
 		b.Run(name, func(b *testing.B) {
 			hey := &tools.Hey{
-				Requests:    requests,
+				Requests:    b.N,
 				Concurrency: c,
 			}
 			runRuby(b, hey)
@@ -61,14 +54,14 @@ func BenchmarkRuby(b *testing.B) {
 // runRuby runs the test for a given # of requests and concurrency.
 func runRuby(b *testing.B, hey *tools.Hey) {
 	// The machine to hold Redis and the Ruby Server.
-	serverMachine, err := h.GetMachine()
+	serverMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
 	defer serverMachine.CleanUp()
 
 	// The machine to run 'hey'.
-	clientMachine, err := h.GetMachine()
+	clientMachine, err := harness.GetMachine()
 	if err != nil {
 		b.Fatalf("failed to get machine with: %v", err)
 	}
@@ -130,10 +123,9 @@ func runRuby(b *testing.B, hey *tools.Hey) {
 		b.Fatalf("failed to wait until  serving: %v", err)
 	}
 	heyCmd := hey.MakeCmd(servingIP, servingPort)
-	rubyApp.RestartProfiles()
-	b.ResetTimer()
 
 	// the client should run on Native.
+	b.ResetTimer()
 	client := clientMachine.GetNativeContainer(ctx, b)
 	defer client.CleanUp(ctx)
 	out, err := client.Run(ctx, dockerutil.RunOpts{
@@ -142,14 +134,11 @@ func runRuby(b *testing.B, hey *tools.Hey) {
 	if err != nil {
 		b.Fatalf("hey container failed: %v logs: %s", err, out)
 	}
-
-	// Stop the timer to parse the data and report stats.
 	b.StopTimer()
 	hey.Report(b, out)
-	b.StartTimer()
 }
 
 func TestMain(m *testing.M) {
-	h.Init()
+	harness.Init()
 	os.Exit(m.Run())
 }
diff --git a/test/benchmarks/tools/fio.go b/test/benchmarks/tools/fio.go
index f5f60fa84..f6324c3ab 100644
--- a/test/benchmarks/tools/fio.go
+++ b/test/benchmarks/tools/fio.go
@@ -25,25 +25,20 @@ import (
 // Fio makes 'fio' commands and parses their output.
 type Fio struct {
 	Test      string // test to run: read, write, randread, randwrite.
-	Size      string // total size to be read/written of format N[GMK] (e.g. 5G).
-	Blocksize string // blocksize to be read/write of format N[GMK] (e.g. 4K).
-	Iodepth   int    // iodepth for reads/writes.
-	Time      int    // time to run the test in seconds, usually for rand(read/write).
+	Size      int    // total size to be read/written in megabytes.
+	BlockSize int    // block size to be read/written in kilobytes.
+	IODepth   int    // I/O depth for reads/writes.
 }
 
 // MakeCmd makes a 'fio' command.
 func (f *Fio) MakeCmd(filename string) []string {
 	cmd := []string{"fio", "--output-format=json", "--ioengine=sync"}
 	cmd = append(cmd, fmt.Sprintf("--name=%s", f.Test))
-	cmd = append(cmd, fmt.Sprintf("--size=%s", f.Size))
-	cmd = append(cmd, fmt.Sprintf("--blocksize=%s", f.Blocksize))
+	cmd = append(cmd, fmt.Sprintf("--size=%dM", f.Size))
+	cmd = append(cmd, fmt.Sprintf("--blocksize=%dK", f.BlockSize))
 	cmd = append(cmd, fmt.Sprintf("--filename=%s", filename))
-	cmd = append(cmd, fmt.Sprintf("--iodepth=%d", f.Iodepth))
+	cmd = append(cmd, fmt.Sprintf("--iodepth=%d", f.IODepth))
 	cmd = append(cmd, fmt.Sprintf("--rw=%s", f.Test))
-	if f.Time != 0 {
-		cmd = append(cmd, "--time_based")
-		cmd = append(cmd, fmt.Sprintf("--runtime=%d", f.Time))
-	}
 	return cmd
 }
 
diff --git a/test/benchmarks/tools/hey.go b/test/benchmarks/tools/hey.go
index b8cb938fe..de908feeb 100644
--- a/test/benchmarks/tools/hey.go
+++ b/test/benchmarks/tools/hey.go
@@ -19,7 +19,6 @@ import (
 	"net"
 	"regexp"
 	"strconv"
-	"strings"
 	"testing"
 )
 
@@ -32,8 +31,16 @@ type Hey struct {
 
 // MakeCmd returns a 'hey' command.
 func (h *Hey) MakeCmd(ip net.IP, port int) []string {
-	return strings.Split(fmt.Sprintf("hey -n %d -c %d http://%s:%d/%s",
-		h.Requests, h.Concurrency, ip, port, h.Doc), " ")
+	c := h.Concurrency
+	if c > h.Requests {
+		c = h.Requests
+	}
+	return []string{
+		"hey",
+		"-n", fmt.Sprintf("%d", h.Requests),
+		"-c", fmt.Sprintf("%d", c),
+		fmt.Sprintf("http://%s:%d/%s", ip.String(), port, h.Doc),
+	}
 }
 
 // Report parses output from 'hey' and reports metrics.
diff --git a/test/benchmarks/tools/iperf.go b/test/benchmarks/tools/iperf.go
index 891d32704..abf296731 100644
--- a/test/benchmarks/tools/iperf.go
+++ b/test/benchmarks/tools/iperf.go
@@ -19,19 +19,27 @@ import (
 	"net"
 	"regexp"
 	"strconv"
-	"strings"
 	"testing"
 )
 
+const length = 64 * 1024
+
 // Iperf is for the client side of `iperf`.
 type Iperf struct {
-	Time int
+	Num int
 }
 
 // MakeCmd returns a iperf client command.
 func (i *Iperf) MakeCmd(ip net.IP, port int) []string {
-	// iperf report in Kb realtime
-	return strings.Split(fmt.Sprintf("iperf -f K --realtime --time %d --client %s --port %d", i.Time, ip, port), " ")
+	return []string{
+		"iperf",
+		"--format", "K", // Output in KBytes.
+		"--realtime", // Measured in realtime.
+		"--num", fmt.Sprintf("%d", i.Num),
+		"--length", fmt.Sprintf("%d", length),
+		"--client", ip.String(),
+		"--port", fmt.Sprintf("%d", port),
+	}
 }
 
 // Report parses output from iperf client and reports metrics.
@@ -42,6 +50,7 @@ func (i *Iperf) Report(b *testing.B, output string) {
 	if err != nil {
 		b.Fatalf("failed to parse bandwitdth from %s: %v", output, err)
 	}
+	b.SetBytes(length) // Measure Bytes/sec for b.N, although below is iperf output.
 	ReportCustomMetric(b, bW*1024, "bandwidth" /*metric name*/, "bytes_per_second" /*unit*/)
 }
 
diff --git a/test/benchmarks/tools/redis.go b/test/benchmarks/tools/redis.go
index a42e3456e..12fdbc7cc 100644
--- a/test/benchmarks/tools/redis.go
+++ b/test/benchmarks/tools/redis.go
@@ -19,7 +19,6 @@ import (
 	"net"
 	"regexp"
 	"strconv"
-	"strings"
 	"testing"
 )
 
@@ -33,13 +32,25 @@ func (r *Redis) MakeCmd(ip net.IP, port, requests int) []string {
 	// There is no -t PING_BULK for redis-benchmark, so adjust the command in that case.
 	// Note that "ping" will run both PING_INLINE and PING_BULK.
 	if r.Operation == "PING_BULK" {
-		return strings.Split(
-			fmt.Sprintf("redis-benchmark --csv -t ping -h %s -p %d -n %d", ip, port, requests), " ")
+		return []string{
+			"redis-benchmark",
+			"--csv",
+			"-t", "ping",
+			"-h", ip.String(),
+			"-p", fmt.Sprintf("%d", port),
+			"-n", fmt.Sprintf("%d", requests),
+		}
 	}
 
 	// runs redis-benchmark -t operation for 100K requests against server.
-	return strings.Split(
-		fmt.Sprintf("redis-benchmark --csv -t %s -h %s -p %d -n %d", r.Operation, ip, port, requests), " ")
+	return []string{
+		"redis-benchmark",
+		"--csv",
+		"-t", r.Operation,
+		"-h", ip.String(),
+		"-p", fmt.Sprintf("%d", port),
+		"-n", fmt.Sprintf("%d", requests),
+	}
 }
 
 // Report parses output from redis-benchmark client and reports metrics.
diff --git a/test/benchmarks/tools/sysbench.go b/test/benchmarks/tools/sysbench.go
index 7ccacd8ff..2b8e6c8aa 100644
--- a/test/benchmarks/tools/sysbench.go
+++ b/test/benchmarks/tools/sysbench.go
@@ -18,58 +18,46 @@ import (
 	"fmt"
 	"regexp"
 	"strconv"
-	"strings"
 	"testing"
 )
 
-var warmup = "sysbench --threads=8 --memory-total-size=5G memory run > /dev/null &&"
-
 // Sysbench represents a 'sysbench' command.
 type Sysbench interface {
-	MakeCmd() []string // Makes a sysbench command.
-	flags() []string
-	Report(*testing.B, string) // Reports results contained in string.
+	// MakeCmd constructs the relevant command line.
+	MakeCmd(*testing.B) []string
+
+	// Report reports relevant custom metrics.
+	Report(*testing.B, string)
 }
 
 // SysbenchBase is the top level struct for sysbench and holds top-level arguments
 // for sysbench. See: 'sysbench --help'
 type SysbenchBase struct {
-	Threads int // number of Threads for the test.
-	Time    int // time limit for test in seconds.
+	// Threads is the number of threads for the test.
+	Threads int
 }
 
 // baseFlags returns top level flags.
-func (s *SysbenchBase) baseFlags() []string {
+func (s *SysbenchBase) baseFlags(b *testing.B) []string {
 	var ret []string
 	if s.Threads > 0 {
 		ret = append(ret, fmt.Sprintf("--threads=%d", s.Threads))
 	}
-	if s.Time > 0 {
-		ret = append(ret, fmt.Sprintf("--time=%d", s.Time))
-	}
+	ret = append(ret, "--time=0") // Ensure events is used.
+	ret = append(ret, fmt.Sprintf("--events=%d", b.N))
 	return ret
 }
 
 // SysbenchCPU is for 'sysbench [flags] cpu run' and holds CPU specific arguments.
 type SysbenchCPU struct {
-	Base     SysbenchBase
-	MaxPrime int // upper limit for primes generator [10000].
+	SysbenchBase
 }
 
 // MakeCmd makes commands for SysbenchCPU.
-func (s *SysbenchCPU) MakeCmd() []string {
-	cmd := []string{warmup, "sysbench"}
-	cmd = append(cmd, s.flags()...)
-	cmd = append(cmd, "cpu run")
-	return []string{"sh", "-c", strings.Join(cmd, " ")}
-}
-
-// flags makes flags for SysbenchCPU cmds.
-func (s *SysbenchCPU) flags() []string {
-	cmd := s.Base.baseFlags()
-	if s.MaxPrime > 0 {
-		return append(cmd, fmt.Sprintf("--cpu-max-prime=%d", s.MaxPrime))
-	}
+func (s *SysbenchCPU) MakeCmd(b *testing.B) []string {
+	cmd := []string{"sysbench"}
+	cmd = append(cmd, s.baseFlags(b)...)
+	cmd = append(cmd, "cpu", "run")
 	return cmd
 }
 
@@ -96,9 +84,9 @@ func (s *SysbenchCPU) parseEvents(data string) (float64, error) {
 
 // SysbenchMemory is for 'sysbench [FLAGS] memory run' and holds Memory specific arguments.
 type SysbenchMemory struct {
-	Base          SysbenchBase
-	BlockSize     string // size of test memory block [1K].
-	TotalSize     string // size of data to transfer [100G].
+	SysbenchBase
+	BlockSize     int    // size of test memory block in megabytes [1].
+	TotalSize     int    // size of data to transfer in gigabytes [100].
 	Scope         string // memory access scope {global, local} [global].
 	HugeTLB       bool   // allocate memory from HugeTLB [off].
 	OperationType string // type of memory ops {read, write, none} [write].
@@ -106,21 +94,21 @@ type SysbenchMemory struct {
 }
 
 // MakeCmd makes commands for SysbenchMemory.
-func (s *SysbenchMemory) MakeCmd() []string {
-	cmd := []string{warmup, "sysbench"}
-	cmd = append(cmd, s.flags()...)
-	cmd = append(cmd, "memory run")
-	return []string{"sh", "-c", strings.Join(cmd, " ")}
+func (s *SysbenchMemory) MakeCmd(b *testing.B) []string {
+	cmd := []string{"sysbench"}
+	cmd = append(cmd, s.flags(b)...)
+	cmd = append(cmd, "memory", "run")
+	return cmd
 }
 
 // flags makes flags for SysbenchMemory cmds.
-func (s *SysbenchMemory) flags() []string {
-	cmd := s.Base.baseFlags()
-	if s.BlockSize != "" {
-		cmd = append(cmd, fmt.Sprintf("--memory-block-size=%s", s.BlockSize))
+func (s *SysbenchMemory) flags(b *testing.B) []string {
+	cmd := s.baseFlags(b)
+	if s.BlockSize != 0 {
+		cmd = append(cmd, fmt.Sprintf("--memory-block-size=%dM", s.BlockSize))
 	}
-	if s.TotalSize != "" {
-		cmd = append(cmd, fmt.Sprintf("--memory-total-size=%s", s.TotalSize))
+	if s.TotalSize != 0 {
+		cmd = append(cmd, fmt.Sprintf("--memory-total-size=%dG", s.TotalSize))
 	}
 	if s.Scope != "" {
 		cmd = append(cmd, fmt.Sprintf("--memory-scope=%s", s.Scope))
@@ -147,7 +135,7 @@ func (s *SysbenchMemory) Report(b *testing.B, output string) {
 	ReportCustomMetric(b, result, "memory_operations" /*metric name*/, "ops_per_second" /*unit*/)
 }
 
-var memoryOperationsRE = regexp.MustCompile(`Total\soperations:\s+\d*\s*\((\d*\.\d*)\sper\ssecond\)`)
+var memoryOperationsRE = regexp.MustCompile(`Total\s+operations:\s+\d+\s+\((\s*\d+\.\d+\s*)\s+per\s+second\)`)
 
 // parseOperations parses memory operations per second form sysbench memory ouput.
 func (s *SysbenchMemory) parseOperations(data string) (float64, error) {
@@ -160,24 +148,24 @@ func (s *SysbenchMemory) parseOperations(data string) (float64, error) {
 
 // SysbenchMutex is for 'sysbench [FLAGS] mutex run' and holds Mutex specific arguments.
 type SysbenchMutex struct {
-	Base  SysbenchBase
+	SysbenchBase
 	Num   int // total size of mutex array [4096].
-	Locks int // number of mutex locks per thread [50K].
-	Loops int // number of loops to do outside mutex lock [10K].
+	Locks int // number of mutex locks per thread [50000].
+	Loops int // number of loops to do outside mutex lock [10000].
 }
 
 // MakeCmd makes commands for SysbenchMutex.
-func (s *SysbenchMutex) MakeCmd() []string {
-	cmd := []string{warmup, "sysbench"}
-	cmd = append(cmd, s.flags()...)
-	cmd = append(cmd, "mutex run")
-	return []string{"sh", "-c", strings.Join(cmd, " ")}
+func (s *SysbenchMutex) MakeCmd(b *testing.B) []string {
+	cmd := []string{"sysbench"}
+	cmd = append(cmd, s.flags(b)...)
+	cmd = append(cmd, "mutex", "run")
+	return cmd
 }
 
 // flags makes flags for SysbenchMutex commands.
-func (s *SysbenchMutex) flags() []string {
+func (s *SysbenchMutex) flags(b *testing.B) []string {
 	var cmd []string
-	cmd = append(cmd, s.Base.baseFlags()...)
+	cmd = append(cmd, s.baseFlags(b)...)
 	if s.Num > 0 {
 		cmd = append(cmd, fmt.Sprintf("--mutex-num=%d", s.Num))
 	}