Use precise memory orderings in load_tester

tools/load_tester.cpp

@@ -401,6 +401,7 @@ private:
       double current_min = g_min_latency.load(std::memory_order_relaxed);
       while (latency < current_min &&
              !g_min_latency.compare_exchange_weak(current_min, latency,
+                                                  std::memory_order_relaxed,
                                                   std::memory_order_relaxed)) {
         // Retry if another thread updated min_latency
       }
@@ -408,6 +409,7 @@ private:
       double current_max = g_max_latency.load(std::memory_order_relaxed);
       while (latency > current_max &&
              !g_max_latency.compare_exchange_weak(current_max, latency,
+                                                  std::memory_order_relaxed,
                                                   std::memory_order_relaxed)) {
         // Retry if another thread updated max_latency
       }
@@ -637,7 +639,7 @@ int main(int argc, char *argv[]) {
       int epollfd = g_epoll_fds[i]; // Each thread uses its own epoll instance
       pthread_setname_np(pthread_self(),
                          ("network-" + std::to_string(i)).c_str());
-      while (g_connect_threads.load() != 0) {
+      while (g_connect_threads.load(std::memory_order_acquire) != 0) {
         struct epoll_event events[256]; // Use a reasonable max size
         int batch_size = std::min(int(sizeof(events) / sizeof(events[0])),
                                   g_config.event_batch_size);
@@ -779,7 +781,7 @@ int main(int argc, char *argv[]) {
           continue;
         }
       }
-      g_connect_threads.fetch_sub(1);
+      g_connect_threads.fetch_sub(1, std::memory_order_release);
     });
   }