initial merge of latency vs load to qps_json_driver.cc

test/cpp/qps/qps_json_driver.cc

@@ -50,13 +50,100 @@ DEFINE_string(scenarios_json, "",
               "JSON string containing an array of Scenario objects");
 DEFINE_bool(quit, false, "Quit the workers");
 
+DEFINE_bool(search, flase, "Search for offered load setting that achieves targeted cpu load");
+
 DEFINE_double(initial_offered_load, 1000.0, "Set up for intial offered load");
 
 DEFINE_double(targeted_cpu_load, 99.0, "targeted cpu load");
 
+DEFINE_double(precision, 500, "final search result precision");
+
 namespace grpc {
 namespace testing {
 
+static std::unique_ptr<ScenarioResult> RunAndReport(const Scenario& scenario,
+                                                    bool* success) {
+  std::cerr << "RUNNING SCENARIO: " << scenario.name() << "\n";
+  auto result =
+      RunScenario(scenario.client_config(), scenario.num_clients(),
+                  scenario.server_config(), scenario.num_servers(),
+                  scenario.warmup_seconds(), scenario.benchmark_seconds(),
+                  scenario.spawn_local_worker_count());
+
+  // Amend the result with scenario config. Eventually we should adjust
+  // RunScenario contract so we don't need to touch the result here.
+  result->mutable_scenario()->CopyFrom(scenario);
+
+  GetReporter()->ReportQPS(*result);
+  GetReporter()->ReportQPSPerCore(*result);
+  GetReporter()->ReportLatency(*result);
+  GetReporter()->ReportTimes(*result);
+  GetReporter()->ReportCpuUsage(*result);
+
+  for (int i = 0; *success && i < result->client_success_size(); i++) {
+    *success = result->client_success(i);
+  }
+  for (int i = 0; *success && i < result->server_success_size(); i++) {
+    *success = result->server_success(i);
+  }
+
+  return result;
+}
+
+static double GetCpuLoad(Scenario * scenario, double offered_load, bool* success) {
+  scenario->mutable_client_config()->mutable_load_params()->mutable_poisson()->
+    set_offered_load(offered_load);
+  auto result = RunAndReport(*scenario, success);
+  return result->summary().server_cpu_usage();
+}
+
+static double BinarySearch(Scenario * scenario, double targeted_cpu_load,
+                        double low_offered_load, double high_offered_load, bool* success) {
+  while (low <= high - FLAGS_precision) {
+    double mid = low + (high - low) /2;
+    double current_cpu_load = GetCpuLoad(scenario, mid, success);
+    gpr_log(GPR_INFO, "binary search: current_offered_load %.0f", mid);
+    if (!*success) {
+      gpr_log(GPR_ERROR, "Client/Server Failure");
+      break;
+    }
+    if (targeted_cpu_load < current_cpu_load) {
+      high = mid -1;
+    }
+    else if (targeted_cpu_load > current_cpu_load) {
+      low = mid + 1;
+    }
+    else {
+      high = mid - 1;
+    }
+  }
+
+  return low;
+}
+
+static double SearchOfferedLoad(double initial_offered_load, double targeted_cpu_load,
+                                Scenario * scenario, bool* success) {
+    std::cerr << "RUNNING SCENARIO: " << scenario->name() << "\n";
+    double current_offered_load = initial_offered_load;
+    double current_cpu_load = GetCpuLoad(scenario, current_offered_load, success);
+    if (current_cpu_load > targeted_cpu_load) {
+      gpr_log(GPR_ERROR, "Initial offered load too high");
+      return -1;
+    }
+
+    while (*success && (current_cpu_load < targeted_cpu_load)) {
+      current_offered_load *= 2;
+      current_cpu_load = GetCpuLoad(scenario, current_offered_load, success);
+      gpr_log(GPR_INFO, "do while: current_offered_load %f", current_offered_load);
+    }
+
+    double targeted_offered_load = BinarySearch(scenario, targeted_cpu_load,
+                                                current_offered_load / 2,
+                                                current_offered_load, success);
+
+    return targeted_offered_load;
+}
+
 static bool QpsDriver() {
   grpc::string json;
 
@@ -97,29 +184,14 @@ static bool QpsDriver() {
   GPR_ASSERT(scenarios.scenarios_size() > 0);
 
   for (int i = 0; i < scenarios.scenarios_size(); i++) {
+    if (!FLAGS_search) {
       const Scenario &scenario = scenarios.scenarios(i);
-    std::cerr << "RUNNING SCENARIO: " << scenario.name() << "\n";
-    auto result =
-        RunScenario(scenario.client_config(), scenario.num_clients(),
-                    scenario.server_config(), scenario.num_servers(),
-                    scenario.warmup_seconds(), scenario.benchmark_seconds(),
-                    scenario.spawn_local_worker_count());
-
-    // Amend the result with scenario config. Eventually we should adjust
-    // RunScenario contract so we don't need to touch the result here.
-    result->mutable_scenario()->CopyFrom(scenario);
-
-    GetReporter()->ReportQPS(*result);
-    GetReporter()->ReportQPSPerCore(*result);
-    GetReporter()->ReportLatency(*result);
-    GetReporter()->ReportTimes(*result);
-    GetReporter()->ReportCpuUsage(*result);
-
-    for (int i = 0; success && i < result->client_success_size(); i++) {
-      success = result->client_success(i);
+      RunAndReport(scenario, &success);
     }
-    for (int i = 0; success && i < result->server_success_size(); i++) {
-      success = result->server_success(i);
+    else {
+      Scenario *scenario = scenarios.mutable_scenarios(i);
+      double targeted_offered_load = SearchOfferedLoad(FLAGS_initial_offered_load, FLAGS_targeted_cpu_load, scenario, &success);
+      gpr_log(GPR_INFO, "targeted_offered_load %f", targeted_offered_load);
     }
   }
   return success;