Merge pull request #753 from dottorblaster/fix-logs-nri-benchmark

fix: prevent eBPF cgroup tracker map exhaustion in NRI benchmarks

Author: dottorblaster

hack/ci/check-nri-benchmarks.sh

@@ -9,7 +9,7 @@ REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
 THRESHOLDS_FILE="${THRESHOLDS_FILE:-${SCRIPT_DIR}/nri-benchmark-thresholds.txt}"
 BENCH_PKG="${BENCH_PKG:-./internal/nri/}"
 BENCH_COUNT="${BENCH_COUNT:-1}"
-BENCH_TOLERANCE_PERCENT="${BENCH_TOLERANCE_PERCENT:-20}"
+BENCH_TOLERANCE_PERCENT="${BENCH_TOLERANCE_PERCENT:-30}"
 UPDATE_BASELINE="${UPDATE_BASELINE:-false}"
 
 if [[ "${GITHUB_ACTIONS:-}" != "true" ]]; then

hack/ci/nri-benchmark-thresholds.txt

@@ -1,4 +1,4 @@
 # NRI callback benchmark baseline (ns/op).
 BenchmarkPluginSynchronize 16629755
-BenchmarkPluginStartContainer 38394
-BenchmarkPluginRemoveContainer 21070
+BenchmarkPluginStartContainer 30000
+BenchmarkPluginRemoveContainer 15714

internal/nri/plugin_bench_test.go

@@ -132,6 +132,23 @@ func benchContainer(b *testing.B, pod *api.PodSandbox, name string) *api.Contain
 	}
 }
 
+// benchTrackerCleanup returns a function that deletes the container's entry
+// from the cgroup tracker map (no BPF programs loaded in benchmarks).
+func benchTrackerCleanup(b *testing.B, maps *benchMaps, container *api.Container) func() {
+	b.Helper()
+
+	cgRoot := cgroups.GetCgroupResolutionPrefix()
+	cgroupPath := filepath.Join(cgRoot, container.GetLinux().GetCgroupsPath())
+	cgID, err := cgroups.GetCgroupIDFromPath(cgroupPath)
+	require.NoError(b, err, "resolve cgroup ID")
+
+	return func() {
+		if err = maps.cgTracker.Delete(&cgID); err != nil {
+			b.Fatalf("delete cgroup tracker entry: %v", err)
+		}
+	}
+}
+
 func buildSyncFixtures(b *testing.B, numPods, containersPerPod int) ([]*api.PodSandbox, []*api.Container) {
 	b.Helper()
 
@@ -183,11 +200,15 @@ func BenchmarkPluginStartContainer(b *testing.B) {
 		b.StopTimer()
 		p := newBenchPlugin(b, maps)
 		container := benchContainer(b, pod, fmt.Sprintf("start-container-%d", i))
+		cleanup := benchTrackerCleanup(b, maps, container)
 		b.StartTimer()
 
 		if err := p.StartContainer(ctx, pod, container); err != nil {
 			b.Fatalf("StartContainer: %v", err)
 		}
+
+		b.StopTimer()
+		cleanup()
 	}
 }
 
@@ -201,6 +222,8 @@ func BenchmarkPluginRemoveContainer(b *testing.B) {
 		b.StopTimer()
 		p := newBenchPlugin(b, maps)
 		container := benchContainer(b, pod, fmt.Sprintf("remove-container-%d", i))
+		cleanup := benchTrackerCleanup(b, maps, container)
+
 		if err := p.StartContainer(ctx, pod, container); err != nil {
 			b.Fatalf("seed StartContainer: %v", err)
 		}
@@ -209,5 +232,8 @@ func BenchmarkPluginRemoveContainer(b *testing.B) {
 		if err := p.RemoveContainer(ctx, pod, container); err != nil {
 			b.Fatalf("RemoveContainer: %v", err)
 		}
+
+		b.StopTimer()
+		cleanup()
 	}
 }