[benchmarks] Update BenchmarkStartEmpty benchmark.

Update the start benchmark on empty to only "Start" a container, not wait
for its completion.

TL:DR only measure the actual start call for the empty container.

Previously, we were measuring the completion of /bin/true in container
alpine AND the cleanup. This was fine until profiling started failing all
the time on ptrace. This is a cost that runc is not paying.

These changes will reduce the over all timing of the benchmark, but it will
give more sane results.

Instead, use "Spawn" which is similar to `docker run --detach alpine
/bin/sleep 100`. Call sleep so containers stick around long enough
for the profiler to read profile data from them.

PiperOrigin-RevId: 390705431

1 files changed, 8 insertions, 4 deletions

diff --git a/test/benchmarks/base/startup_test.go b/test/benchmarks/base/startup_test.go
index 05a43ad17..197241622 100644
--- a/test/benchmarks/base/startup_test.go
+++ b/test/benchmarks/base/startup_test.go
@@ -34,15 +34,19 @@ func BenchmarkStartupEmpty(b *testing.B) {
 	defer machine.CleanUp()
 
 	ctx := context.Background()
+	b.StopTimer()
+	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		harness.DebugLog(b, "Running container: %d", i)
 		container := machine.GetContainer(ctx, b)
-		defer container.CleanUp(ctx)
-		if _, err := container.Run(ctx, dockerutil.RunOpts{
+		b.StartTimer()
+		if err := container.Spawn(ctx, dockerutil.RunOpts{
 			Image: "benchmarks/alpine",
-		}, "true"); err != nil {
-			b.Fatalf("failed to run container: %v", err)
+		}, "sleep", "100"); err != nil {
+			b.Fatalf("failed to start container: %v", err)
 		}
+		b.StopTimer()
+		container.CleanUp(ctx)
 		harness.DebugLog(b, "Ran container: %d", i)
 	}
 }