Record pause time per task

Differential Revision: D55159826

velox/exec/Driver.h

@@ -110,6 +110,9 @@ struct ThreadState {
   /// driver goes off thread. This is used to track the time that a driver has
   /// continuously run on a thread for per-driver cpu time slice enforcement.
   size_t startExecTimeMs{0};
+  /// The end execution time on thread in milliseconds. It is set when the
+  /// driver goes off thread and reset when the driver gets on a thread.
+  size_t endExecTimeMs{0};
 
   bool isOnThread() const {
     return thread != std::thread::id();
@@ -118,6 +121,7 @@ struct ThreadState {
   void setThread() {
     thread = std::this_thread::get_id();
     startExecTimeMs = getCurrentTimeMs();
+    endExecTimeMs = 0;
 #if !defined(__APPLE__)
     // This is a debugging feature disabled on the Mac since syscall
     // is deprecated on that platform.
@@ -128,6 +132,7 @@ struct ThreadState {
   void clearThread() {
     thread = std::thread::id(); // no thread.
     startExecTimeMs = 0;
+    endExecTimeMs = getCurrentTimeMs();
     tid = 0;
   }
 

velox/exec/Task.cpp

@@ -876,6 +876,8 @@ void Task::resume(std::shared_ptr<Task> self) {
           if (!driver->state().hasBlockingFuture) {
             // Do not continue a Driver that is blocked on external
             // event. The Driver gets enqueued by the promise realization.
+            self->taskStats_.totalPauseTimeMs +=
+                getCurrentTimeMs() - driver->state().endExecTimeMs;
             Driver::enqueue(driver);
           }
         }

velox/exec/TaskStats.h

@@ -98,6 +98,9 @@ struct TaskStats {
   /// Drivers blocked for various reasons. Based on enum BlockingReason.
   std::unordered_map<BlockingReason, uint64_t> numBlockedDrivers;
 
+  /// Total pause time (ms) for all drivers.
+  uint64_t totalPauseTimeMs{0};
+
   /// Output buffer's memory utilization ratio measured as
   /// current buffer usage / max buffer size
   double outputBufferUtilization{0};