Add setting enforce_max_duration

Author: anh

loadgen/bindings/python_api.cc

@@ -316,6 +316,8 @@ PYBIND11_MODULE(mlperf_loadgen, m) {
       .def_readwrite("max_duration_ms", &TestSettings::max_duration_ms)
       .def_readwrite("min_query_count", &TestSettings::min_query_count)
       .def_readwrite("max_query_count", &TestSettings::max_query_count)
+      def_readwrite("enforce_max_duration",
+                    &TestSettings::enforce_max_duration)
       .def_readwrite("qsl_rng_seed", &TestSettings::qsl_rng_seed)
       .def_readwrite("sample_index_rng_seed",
                      &TestSettings::sample_index_rng_seed)

loadgen/issue_query_controller.cc

@@ -515,19 +515,21 @@ void IssueQueryController::IssueQueriesInternal(size_t query_stride,
     // Checks if we have exceeded max_duration.
     if (settings.max_duration.count() != 0 &&
         duration > settings.max_duration) {
-      LogDetail([thread_idx, duration](AsyncDetail& detail) {
+      if (settings.enforce_max_duration) {
+        LogDetail([thread_idx, duration](AsyncDetail& detail) {
 #if USE_NEW_LOGGING_FORMAT
-        std::stringstream ss;
-        ss << "IssueQueryThread " << thread_idx
-           << " Ending early: Max test duration reached." << " duration_ns "
-           << duration.count();
-        MLPERF_LOG_ERROR(detail, "error_runtime", ss.str());
+          std::stringstream ss;
+          ss << "IssueQueryThread " << thread_idx
+             << " Ending early: Max test duration reached." << " duration_ns "
+             << duration.count();
+          MLPERF_LOG_ERROR(detail, "error_runtime", ss.str());
 #else
-        detail.Error("IssueQueryThread ", std::to_string(thread_idx),
-                     " Ending early: Max test duration reached.", "duration_ns",
-                     std::to_string(duration.count()));
+          detail.Error("IssueQueryThread ", std::to_string(thread_idx),
+                       " Ending early: Max test duration reached.",
+                       "duration_ns", std::to_string(duration.count()));
 #endif
-      });
+        });
+      }
       ran_out_of_generated_queries = false;
       break;
     }

loadgen/results.cc

@@ -506,15 +506,18 @@ void PerformanceSummary::LogSummary(AsyncSummary& summary) {
   }
   bool perf_constraints_met =
       PerfConstraintsMet(&perf_constraints_recommendation);
-  bool all_constraints_met = min_duration_met && min_queries_met &&
+  bool all_constraints_met = min_duration_met &&
                              perf_constraints_met && early_stopping_met;
+  if (!settings.enforce_max_duration){
+    all_constraints_met = all_constraints_met && min_queries_met;
+  }
   summary("Result is : ", all_constraints_met ? "VALID" : "INVALID");
   if (HasPerfConstraints()) {
     summary("  Performance constraints satisfied : ",
             perf_constraints_met ? "Yes" : "NO");
   }
   summary("  Min duration satisfied : ", min_duration_met ? "Yes" : "NO");
-  summary("  Min queries satisfied : ", min_queries_met ? "Yes" : "NO");
+  summary(" Min queries satisfied : ", min_queries_met ? "Yes" : settings.enforce_max_duration ? "Skipped" : "NO");
   summary("  Early stopping satisfied: ", early_stopping_met ? "Yes" : "NO");
 
   if (!all_constraints_met) {
@@ -668,9 +671,11 @@ void PerformanceSummary::LogDetail(AsyncDetail& detail) {
                       &pr.query_latencies,
                       std::chrono::nanoseconds(settings.server_ttft_latency));
   }
-  bool all_constraints_met = min_duration_met && min_queries_met &&
+  bool all_constraints_met = min_duration_met &&
                              perf_constraints_met && early_stopping_met;
-
+  if (!settings.enforce_max_duration) {
+    all_constraints_met = all_constraints_met && min_queries_met;
+  }
   MLPERF_LOG(detail, "result_validity",
              all_constraints_met ? "VALID" : "INVALID");
   if (HasPerfConstraints()) {

loadgen/test_settings.h

@@ -196,6 +196,7 @@ struct TestSettings {
   /**@{*/
   uint64_t min_duration_ms = 10000;
   uint64_t max_duration_ms = 0;  ///< 0: Infinity.
+  bool enforce_max_duration = true;
   uint64_t min_query_count = 100;
   uint64_t max_query_count = 0;  ///< 0: Infinity.
   /**@}*/

loadgen/test_settings_internal.cc

@@ -35,6 +35,7 @@ TestSettingsInternal::TestSettingsInternal(
       target_duration(std::chrono::milliseconds(requested.min_duration_ms)),
       min_duration(std::chrono::milliseconds(requested.min_duration_ms)),
       max_duration(std::chrono::milliseconds(requested.max_duration_ms)),
+      enforce_max_duration(requested.enforce_max_duration),
       min_query_count(requested.min_query_count),
       max_query_count(requested.max_query_count),
       min_sample_count(0),
@@ -428,6 +429,7 @@ void TestSettingsInternal::LogEffectiveSettings() const {
                s.target_duration.count());
     MLPERF_LOG(detail, "effective_min_duration_ms", s.min_duration.count());
     MLPERF_LOG(detail, "effective_max_duration_ms", s.max_duration.count());
+    MLPERF_LOG(detail, "effective_enforce_max_duration", s.enforce_max_duration);
     MLPERF_LOG(detail, "effective_min_query_count", s.min_query_count);
     MLPERF_LOG(detail, "effective_max_query_count", s.max_query_count);
     MLPERF_LOG(detail, "effective_min_sample_count", s.min_sample_count);
@@ -775,6 +777,8 @@ int TestSettings::FromConfig(const std::string &path, const std::string &model,
 
   lookupkv(model, scenario, "min_duration", &min_duration_ms, nullptr);
   lookupkv(model, scenario, "max_duration", &max_duration_ms, nullptr);
+  if (lookupkv(model, scenario, "enforce_max_duration", &val, nullptr))
+    enforce_max_duration = (val == 1) ? true : false;
   lookupkv(model, scenario, "min_query_count", &min_query_count, nullptr);
   lookupkv(model, scenario, "max_query_count", &max_query_count, nullptr);
   lookupkv(model, scenario, "performance_sample_count_override",

loadgen/test_settings_internal.h

@@ -65,6 +65,7 @@ struct TestSettingsInternal {
   // duration at the end of the run.
   std::chrono::milliseconds min_duration{0};
   std::chrono::milliseconds max_duration{0};
+  bool enforce_max_duration;
   uint64_t min_query_count;
   uint64_t max_query_count;
   uint64_t min_sample_count;  // Offline only.

tools/submission/submission_checker.py

@@ -1598,6 +1598,8 @@ def check_performance_dir(
     min_query_count = mlperf_log["effective_min_query_count"]
     samples_per_query = mlperf_log["effective_samples_per_query"]
     min_duration = mlperf_log["effective_min_duration_ms"]
+    enforce_max_duration = mlperf_log.get("effective_enforce_max_duration", True)
+    min_queries_met = mlperf_log["result_min_queries_met"]
     equal_issue_used_check = (
         mlperf_log["effective_sample_concatenate_permutation"] == True
     )
@@ -1749,6 +1751,13 @@ def check_performance_dir(
         )
         is_valid = False
 
+    if not min_queries_met and not enforce_max_duration:
+        log.error(
+            "%s Loadgen needs to enforce max duration for official submissions. Enforced %s",
+            fname,
+            enforce_max_duration,
+        )
+
     inferred = False
     if scenario_fixed != scenario:
         inferred, res, inferred_valid = get_inferred_result(