Add counters for rmw atomic operations to microbenchmarks

1 files changed, 51 insertions, 0 deletions

diff --git a/test/cpp/microbenchmarks/bm_closure.cc b/test/cpp/microbenchmarks/bm_closure.cc
index 80d6610e13..16d05781bb 100644
--- a/test/cpp/microbenchmarks/bm_closure.cc
+++ b/test/cpp/microbenchmarks/bm_closure.cc
@@ -43,13 +43,46 @@ extern "C" {
 
 #include "third_party/benchmark/include/benchmark/benchmark.h"
 
+#include <sstream>
+
+#ifdef GPR_LOW_LEVEL_COUNTERS
+extern "C" gpr_atm gpr_mu_locks;
+#endif
+
 static class InitializeStuff {
  public:
   InitializeStuff() { grpc_init(); }
   ~InitializeStuff() { grpc_shutdown(); }
 } initialize_stuff;
 
+class TrackCounters {
+ public:
+  TrackCounters(benchmark::State& state) : state_(state) {}
+
+  ~TrackCounters() {
+    std::ostringstream out;
+#ifdef GPR_LOW_LEVEL_COUNTERS
+    out << " locks/iter:" << ((double)(gpr_atm_no_barrier_load(&gpr_mu_locks) -
+                                       mu_locks_at_start_) /
+                              (double)state_.iterations())
+        << " atm_rmw/iter:"
+        << ((double)(gpr_atm_no_barrier_load(&gpr_counter_rmw) -
+                     rmw_at_start_) /
+            (double)state_.iterations());
+#endif
+    state_.SetLabel(out.str());
+  }
+
+ private:
+  benchmark::State& state_;
+#ifdef GPR_LOW_LEVEL_COUNTERS
+  const size_t mu_locks_at_start_ = gpr_atm_no_barrier_load(&gpr_mu_locks);
+  const size_t rmw_at_start_ = gpr_atm_no_barrier_load(&gpr_counter_rmw);
+#endif
+};
+
 static void BM_NoOpExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   while (state.KeepRunning()) {
     grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
     grpc_exec_ctx_finish(&exec_ctx);
@@ -58,6 +91,7 @@ static void BM_NoOpExecCtx(benchmark::State& state) {
 BENCHMARK(BM_NoOpExecCtx);
 
 static void BM_WellFlushed(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   while (state.KeepRunning()) {
     grpc_exec_ctx_flush(&exec_ctx);
@@ -69,6 +103,7 @@ BENCHMARK(BM_WellFlushed);
 static void DoNothing(grpc_exec_ctx* exec_ctx, void* arg, grpc_error* error) {}
 
 static void BM_ClosureInitAgainstExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_closure c;
   while (state.KeepRunning()) {
     benchmark::DoNotOptimize(
@@ -78,6 +113,7 @@ static void BM_ClosureInitAgainstExecCtx(benchmark::State& state) {
 BENCHMARK(BM_ClosureInitAgainstExecCtx);
 
 static void BM_ClosureInitAgainstCombiner(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_combiner* combiner = grpc_combiner_create(NULL);
   grpc_closure c;
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
@@ -91,6 +127,7 @@ static void BM_ClosureInitAgainstCombiner(benchmark::State& state) {
 BENCHMARK(BM_ClosureInitAgainstCombiner);
 
 static void BM_ClosureRunOnExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_closure c;
   grpc_closure_init(&c, DoNothing, NULL, grpc_schedule_on_exec_ctx);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
@@ -103,6 +140,7 @@ static void BM_ClosureRunOnExecCtx(benchmark::State& state) {
 BENCHMARK(BM_ClosureRunOnExecCtx);
 
 static void BM_ClosureCreateAndRun(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   while (state.KeepRunning()) {
     grpc_closure_run(&exec_ctx, grpc_closure_create(DoNothing, NULL,
@@ -114,6 +152,7 @@ static void BM_ClosureCreateAndRun(benchmark::State& state) {
 BENCHMARK(BM_ClosureCreateAndRun);
 
 static void BM_ClosureInitAndRun(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   grpc_closure c;
   while (state.KeepRunning()) {
@@ -126,6 +165,7 @@ static void BM_ClosureInitAndRun(benchmark::State& state) {
 BENCHMARK(BM_ClosureInitAndRun);
 
 static void BM_ClosureSchedOnExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_closure c;
   grpc_closure_init(&c, DoNothing, NULL, grpc_schedule_on_exec_ctx);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
@@ -138,6 +178,7 @@ static void BM_ClosureSchedOnExecCtx(benchmark::State& state) {
 BENCHMARK(BM_ClosureSchedOnExecCtx);
 
 static void BM_ClosureSched2OnExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_closure c1;
   grpc_closure c2;
   grpc_closure_init(&c1, DoNothing, NULL, grpc_schedule_on_exec_ctx);
@@ -153,6 +194,7 @@ static void BM_ClosureSched2OnExecCtx(benchmark::State& state) {
 BENCHMARK(BM_ClosureSched2OnExecCtx);
 
 static void BM_ClosureSched3OnExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_closure c1;
   grpc_closure c2;
   grpc_closure c3;
@@ -171,6 +213,7 @@ static void BM_ClosureSched3OnExecCtx(benchmark::State& state) {
 BENCHMARK(BM_ClosureSched3OnExecCtx);
 
 static void BM_AcquireMutex(benchmark::State& state) {
+  TrackCounters track_counters(state);
   // for comparison with the combiner stuff below
   gpr_mu mu;
   gpr_mu_init(&mu);
@@ -185,6 +228,7 @@ static void BM_AcquireMutex(benchmark::State& state) {
 BENCHMARK(BM_AcquireMutex);
 
 static void BM_ClosureSchedOnCombiner(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_combiner* combiner = grpc_combiner_create(NULL);
   grpc_closure c;
   grpc_closure_init(&c, DoNothing, NULL,
@@ -200,6 +244,7 @@ static void BM_ClosureSchedOnCombiner(benchmark::State& state) {
 BENCHMARK(BM_ClosureSchedOnCombiner);
 
 static void BM_ClosureSched2OnCombiner(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_combiner* combiner = grpc_combiner_create(NULL);
   grpc_closure c1;
   grpc_closure c2;
@@ -219,6 +264,7 @@ static void BM_ClosureSched2OnCombiner(benchmark::State& state) {
 BENCHMARK(BM_ClosureSched2OnCombiner);
 
 static void BM_ClosureSched3OnCombiner(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_combiner* combiner = grpc_combiner_create(NULL);
   grpc_closure c1;
   grpc_closure c2;
@@ -242,6 +288,7 @@ static void BM_ClosureSched3OnCombiner(benchmark::State& state) {
 BENCHMARK(BM_ClosureSched3OnCombiner);
 
 static void BM_ClosureSched2OnTwoCombiners(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_combiner* combiner1 = grpc_combiner_create(NULL);
   grpc_combiner* combiner2 = grpc_combiner_create(NULL);
   grpc_closure c1;
@@ -263,6 +310,7 @@ static void BM_ClosureSched2OnTwoCombiners(benchmark::State& state) {
 BENCHMARK(BM_ClosureSched2OnTwoCombiners);
 
 static void BM_ClosureSched4OnTwoCombiners(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_combiner* combiner1 = grpc_combiner_create(NULL);
   grpc_combiner* combiner2 = grpc_combiner_create(NULL);
   grpc_closure c1;
@@ -323,6 +371,7 @@ class Rescheduler {
 };
 
 static void BM_ClosureReschedOnExecCtx(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   Rescheduler r(state, grpc_schedule_on_exec_ctx);
   r.ScheduleFirst(&exec_ctx);
@@ -331,6 +380,7 @@ static void BM_ClosureReschedOnExecCtx(benchmark::State& state) {
 BENCHMARK(BM_ClosureReschedOnExecCtx);
 
 static void BM_ClosureReschedOnCombiner(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   grpc_combiner* combiner = grpc_combiner_create(NULL);
   Rescheduler r(state, grpc_combiner_scheduler(combiner, false));
@@ -342,6 +392,7 @@ static void BM_ClosureReschedOnCombiner(benchmark::State& state) {
 BENCHMARK(BM_ClosureReschedOnCombiner);
 
 static void BM_ClosureReschedOnCombinerFinally(benchmark::State& state) {
+  TrackCounters track_counters(state);
   grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   grpc_combiner* combiner = grpc_combiner_create(NULL);
   Rescheduler r(state, grpc_combiner_finally_scheduler(combiner, false));