Fix bug: units are not unified, causing window calculation errors, fix code style

src/brpc/policy/auto_concurrency_limiter.cpp

@@ -108,13 +108,12 @@ void AutoConcurrencyLimiter::OnResponded(int error_code, int64_t latency_us) {
     if (last_sampling_time_us == 0 || 
         now_time_us - last_sampling_time_us >= 
             FLAGS_auto_cl_sampling_interval_ms * 1000) {
-        bool sample_this_call = _last_sampling_time_us.compare_exchange_weak(
-                last_sampling_time_us, now_time_us, 
-                butil::memory_order_relaxed);
+        bool sample_this_call = _last_sampling_time_us.compare_exchange_strong(
+            last_sampling_time_us, now_time_us, butil::memory_order_relaxed);
         if (sample_this_call) {
             int32_t max_concurrency = AddSample(error_code, latency_us, now_time_us);
             if (max_concurrency != 0) {
-                LOG(INFO) 
+                LOG_EVERY_N(INFO, 100) 
                     << "MaxConcurrency updated by auto limiter,"
                     << "current_max_concurrency:" << max_concurrency;
             }
@@ -137,8 +136,7 @@ int32_t AutoConcurrencyLimiter::AddSample(int error_code,
         _sw.start_time_us = sampling_time_us;
     }
 
-    if (error_code != 0 && 
-        FLAGS_auto_cl_enable_error_punish) {
+    if (error_code != 0 && FLAGS_auto_cl_enable_error_punish) {
         ++_sw.failed_count;
         _sw.total_failed_us += latency_us;
     } else if (error_code == 0) {
@@ -147,9 +145,16 @@ int32_t AutoConcurrencyLimiter::AddSample(int error_code,
     }
 
     if (_sw.succ_count + _sw.failed_count < FLAGS_auto_cl_min_sample_count) {
+        if (sampling_time_us - _sw.sampling_time_us >= 
+            FLAGS_auto_cl_sample_window_size_ms * 1000) {
+            // If the sample size is insufficient at the end of the sampling 
+            // window, discard the entire sampling window
+            ResetSampleWindow(sampling_time_us);
+        }
         return 0;
     } 
-    if (sampling_time_us - _sw.start_time_us < FLAGS_auto_cl_sample_window_size_ms &&
+    if (sampling_time_us - _sw.start_time_us < 
+        FLAGS_auto_cl_sample_window_size_ms * 1000 &&
         _sw.succ_count + _sw.failed_count < FLAGS_auto_cl_max_sample_count) {
         return 0;
     }
@@ -221,7 +226,7 @@ int32_t AutoConcurrencyLimiter::UpdateMaxConcurrency(int64_t sampling_time_us) {
         return 0;
     }
     // Remeasure min_latency when concurrency has dropped to low load
-    if (_reset_latency_us > 0 && _reset_latency_us < sampling_time_us) {
+    if (_reset_latency_us > 0 && _reset_latency_us <= sampling_time_us) {
         _min_latency_us = -1;
         _reset_latency_us = 0;
         _remeasure_start_us = NextResetTime(sampling_time_us);
@@ -233,7 +238,6 @@ int32_t AutoConcurrencyLimiter::UpdateMaxConcurrency(int64_t sampling_time_us) {
     if (_remeasure_start_us <= sampling_time_us) {
         _reset_latency_us = sampling_time_us + avg_latency;
         next_max_concurrency = _max_concurrency / 2;
-        LOG(INFO) << "Prepare" << _max_concurrency;
     } else {
         int32_t noload_concurrency = 
             std::ceil(_min_latency_us * _ema_peak_qps / 1000000);