ebpf: reducing ebpf call overhead by using sampling instead of tracin…

…g every calls (#928) * ebpf: reducing ebpf call overhead by using sampling instead of tracing every calls Signed-off-by: Huamin Chen <[email protected]> * review feedback: setting sample_rate to 1 Signed-off-by: Huamin Chen <[email protected]> --------- Signed-off-by: Huamin Chen <[email protected]>
sustainable-computing-io · Sep 13, 2023 · 5f33240 · 5f33240
1 parent d8f89ea
commit 5f33240
Show file tree

Hide file tree

Showing 8 changed files with 58 additions and 2 deletions.
diff --git a/bpfassets/libbpf/bpf.o/amd64_kepler.bpf.o b/bpfassets/libbpf/bpf.o/amd64_kepler.bpf.o
diff --git a/bpfassets/libbpf/src/kepler.bpf.c b/bpfassets/libbpf/src/kepler.bpf.c
@@ -44,6 +44,10 @@ BPF_ARRAY(cache_miss, u64);
 // cpu freq counters
 BPF_ARRAY(cpu_freq_array, u32);
 
+// setting sample rate or counter to 0 will make compiler to remove the code entirely.
+int sample_rate = 1;
+int counter = 1;
+
 static inline u64 get_on_cpu_time(u32 cur_pid, u32 prev_pid, u64 cur_ts)
 {
     u64 cpu_time = 0;
@@ -207,6 +211,13 @@ SEC("tracepoint/sched/sched_switch")
 int kepler_trace(struct sched_switch_args *ctx)
 {
     u32 next_pid = ctx->next_pid; // the new pid that is to be scheduled
+
+    if (counter > 0)
+    {
+        counter--;
+        return 0;
+    }
+    counter = sample_rate;
     u32 cur_pid = bpf_get_current_pid_tgid();
     u64 cgroup_id = bpf_get_current_cgroup_id(); // the cgroup id is the cgroup id of the running process (this is not next_pid or prev_pid)
     u64 cur_ts = bpf_ktime_get_ns();
@@ -260,4 +271,4 @@ int kepler_irq_trace(struct trace_event_raw_softirq *ctx)
     return 0;
 }
 
-char _license[] SEC("license") = "GPL";
+char _license[] SEC("license") = "GPL";
diff --git a/bpfassets/perf_event/perf_event.c b/bpfassets/perf_event/perf_event.c
@@ -71,6 +71,11 @@ BPF_ARRAY(cache_miss, u64, NUM_CPUS);
 // cpu freq counters
 BPF_ARRAY(cpu_freq_array, u32, NUM_CPUS);
 
+#ifndef SAMPLE_RATE
+#define SAMPLE_RATE 0
+#endif
+BPF_HASH(sample_rate, u32, u32);
+
 static inline u64 get_on_cpu_time(pid_t cur_pid, u32 prev_pid, u32 cpu_id, u64 cur_ts)
 {
     u64 cpu_time = 0;
@@ -192,6 +197,19 @@ static inline u64 get_on_cpu_avg_freq(u32 *cpu_id, u64 on_cpu_cycles_delta, u64
 // int kprobe__finish_task_switch(switch_args *ctx)
 int kprobe__finish_task_switch(struct pt_regs *ctx, struct task_struct *prev)
 {
+    u32 initial = SAMPLE_RATE, *sample_counter_value, sample_counter_key = 1234;
+    // only do sampling if sample rate is set
+    if (initial != 0) {
+        sample_counter_value = sample_rate.lookup_or_try_init(&sample_counter_key, &initial);
+        if (sample_counter_value > 0) {
+            if (*sample_counter_value > 0) {
+                (*sample_counter_value)--;
+                return 0;
+            }
+        }
+        sample_rate.update(&sample_counter_key, &initial);
+    }
+
     pid_t cur_pid = bpf_get_current_pid_tgid();
 #ifdef SET_GROUP_ID
     u64 cgroup_id = bpf_get_current_cgroup_id();

diff --git a/pkg/bpfassets/attacher/bcc_attacher.go b/pkg/bpfassets/attacher/bcc_attacher.go
@@ -144,10 +144,11 @@ func attachBccModule() (*BccModuleTables, error) {
 		// so if /proc/cpuinfo is available, we can get the number of all CPUs
 		cores = int(cpu.TotalThreads)
 	}
-
+	bpfSampleRate := config.BPFSampleRate
 	options := []string{
 		"-DMAP_SIZE=" + strconv.Itoa(MapSize),
 		"-DNUM_CPUS=" + strconv.Itoa(cores),
+		"-DSAMPLE_RATE=" + strconv.Itoa(bpfSampleRate),
 	}
 	if config.EnabledEBPFCgroupID {
 		options = append(options, "-DSET_GROUP_ID")

diff --git a/pkg/bpfassets/attacher/libbpf_attacher.go b/pkg/bpfassets/attacher/libbpf_attacher.go
@@ -106,7 +106,13 @@ func attachLibbpfModule() (*bpf.Module, error) {
 			klog.Infof("failed to resize array %s: %v\n", arrayName, err)
 		}
 	}
+	// set the sample rate, this must be done before loading the object
+	sampleRate := config.BPFSampleRate
 
+	err = libbpfModule.InitGlobalVariable("sample_rate", int32(sampleRate))
+	if err != nil {
+		return nil, fmt.Errorf("failed to set sample rate: %v", err)
+	}
 	err = libbpfModule.BPFLoadObject()
 
 	// attach sched_switch tracepoint to kepler_trace function

diff --git a/pkg/bpfassets/perf_event_bindata.go b/pkg/bpfassets/perf_event_bindata.go
diff --git a/pkg/collector/metric/container_metric.go b/pkg/collector/metric/container_metric.go
@@ -271,6 +271,9 @@ func (c *ContainerMetrics) UpdateCgroupMetrics() error {
 }
 
 func (c *ContainerMetrics) GetDynEnergyStat(component string) (energyStat *types.UInt64Stat) {
+	if c == nil {
+		return
+	}
 	switch component {
 	case PKG:
 		return c.DynEnergyInPkg

diff --git a/pkg/config/config.go b/pkg/config/config.go
@@ -83,6 +83,7 @@ var (
 	BindAddressKey  = "BIND_ADDRESS"
 	CPUArchOverride = getConfig("CPU_ARCH_OVERRIDE", "")
 	MaxLookupRetry  = getIntConfig("MAX_LOOKUP_RETRY", defaultMaxLookupRetry)
+	BPFSampleRate   = getIntConfig("EXPERIMENTAL_BPF_SAMPLE_RATE", 0)
 
 	EstimatorModel        = getConfig("ESTIMATOR_MODEL", defaultMetricValue)         // auto-select
 	EstimatorSelectFilter = getConfig("ESTIMATOR_SELECT_FILTER", defaultMetricValue) // no filter
@@ -149,6 +150,7 @@ func logBoolConfigs() {
 		klog.V(5).Infof("EXPOSE_KUBELET_METRICS: %t", ExposeKubeletMetrics)
 		klog.V(5).Infof("EXPOSE_IRQ_COUNTER_METRICS: %t", ExposeIRQCounterMetrics)
 		klog.V(5).Infof("EXPOSE_ESTIMATED_IDLE_POWER_METRICS: %t. This only impacts when the power is estimated using pre-prained models. Estimated idle power is meaningful only when Kepler is running on bare-metal or with a single virtual machine (VM) on the node.", ExposeEstimatedIdlePowerMetrics)
+		klog.V(5).Infof("EXPERIMENTAL_BPF_SAMPLE_RATE: %t", BPFSampleRate)
 	}
 }