Add priority into Run Handler thread pool options. Run handler thread pool will take consideration of the priorities of requests when scheduling ops.

PiperOrigin-RevId: 299199567
Change-Id: I22b99f3bc9e708573d09c56194baaaba7948291e

tensorflow/core/common_runtime/direct_session.cc

@@ -593,7 +593,9 @@ Status DirectSession::RunInternal(
   if (ShouldUseRunHandlerPool(run_options) &&
       run_options.experimental().use_run_handler_pool()) {
     VLOG(1) << "Using RunHandler to scheduler inter-op closures.";
-    handler = GetOrCreateRunHandlerPool(options_)->Get(step_id, call_timeout);
+    handler = GetOrCreateRunHandlerPool(options_)->Get(
+        step_id, call_timeout,
+        run_options.experimental().run_handler_pool_options());
     if (!handler) {
       return errors::DeadlineExceeded(
           "Could not obtain RunHandler for request after waiting for ",

tensorflow/core/framework/run_handler.cc

@@ -19,6 +19,7 @@ limitations under the License.
 
 #include <algorithm>
 #include <cmath>
+#include <list>
 #include <memory>
 
 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
@@ -841,12 +842,15 @@ class RunHandler::Impl {
   void ScheduleInterOpClosure(std::function<void()> fn);
   void ScheduleIntraOpClosure(std::function<void()> fn);
 
-  void Reset(int64 step_id);
+  void Reset(int64 step_id,
+             const RunOptions::Experimental::RunHandlerPoolOptions& options);
 
   RunHandlerPool::Impl* pool_impl() { return pool_impl_; }
 
   ThreadWorkSource* tws() { return &tws_; }
 
+  int64 priority() { return options_.priority(); }
+
  private:
   class ThreadPoolInterfaceWrapper : public thread::ThreadPoolInterface {
    public:
@@ -866,6 +870,7 @@ class RunHandler::Impl {
   int64 step_id_;
   std::unique_ptr<thread::ThreadPoolInterface> thread_pool_interface_;
   ThreadWorkSource tws_;
+  RunOptions::Experimental::RunHandlerPoolOptions options_;
 };
 
 // Contains shared state across all run handlers present in the pool. Also
@@ -891,7 +896,6 @@ class RunHandlerPool::Impl {
             std::vector<double>({1}))) {
     VLOG(1) << "Creating a RunHandlerPool with max handlers: " << max_handlers_;
     free_handlers_.reserve(max_handlers_);
-    sorted_active_handlers_.reserve(max_handlers_);
     handlers_.reserve(max_handlers_);
     for (int i = 0; i < max_handlers_; ++i) {
       handlers_.emplace_back(new RunHandler::Impl(this));
@@ -928,7 +932,9 @@ class RunHandlerPool::Impl {
     return !free_handlers_.empty();
   }
 
-  std::unique_ptr<RunHandler> Get(int64 step_id, int64 timeout_in_ms)
+  std::unique_ptr<RunHandler> Get(
+      int64 step_id, int64 timeout_in_ms,
+      const RunOptions::Experimental::RunHandlerPoolOptions& options)
       TF_LOCKS_EXCLUDED(mu_) {
     thread_local std::unique_ptr<Eigen::MaxSizeVector<ThreadWorkSource*>>
         thread_work_sources =
@@ -960,17 +966,24 @@ class RunHandlerPool::Impl {
       // Remove the last entry from free_handlers_ and add to the end of
       // sorted_active_handlers_.
       handler_impl = free_handlers_.back();
-      handler_impl->Reset(step_id);
-      // Sortedness isn't violated if we simply add at the end of the list,
-      // since handlers are expected to be obtained in increasing order of time.
-      sorted_active_handlers_.push_back(handler_impl);
-      DCHECK_LE(sorted_active_handlers_.size(), max_handlers_);
+      handler_impl->Reset(step_id, options);
       free_handlers_.pop_back();
 
-      num_active_requests = sorted_active_handlers_.size();
+      num_active_requests = sorted_active_handlers_.size() + 1;
       thread_work_sources->resize(num_active_requests);
+      int priority = options.priority();
+      auto it = sorted_active_handlers_.cbegin();
+      bool new_handler_inserted = false;
       for (int i = 0; i < num_active_requests; ++i) {
-        (*thread_work_sources)[i] = sorted_active_handlers_[i]->tws();
+        if (!new_handler_inserted && (it == sorted_active_handlers_.cend() ||
+                                      priority > (*it)->priority())) {
+          sorted_active_handlers_.insert(it, handler_impl);
+          new_handler_inserted = true;
+          // Point to the newly added handler.
+          --it;
+        }
+        (*thread_work_sources)[i] = (*it)->tws();
+        ++it;
       }
       version = ++version_;
     }
@@ -1009,6 +1022,16 @@ class RunHandlerPool::Impl {
     // requests will trigger recomputation.
   }
 
+  std::vector<int64> GetActiveHandlerPrioritiesForTesting()
+      TF_LOCKS_EXCLUDED(mu_) {
+    mutex_lock l(mu_);
+    std::vector<int64> ret;
+    for (const auto& handler_impl : sorted_active_handlers_) {
+      ret.push_back(handler_impl->priority());
+    }
+    return ret;
+  }
+
  private:
   void RecomputePoolStats(
       int num_active_requests, uint64 version,
@@ -1029,9 +1052,13 @@ class RunHandlerPool::Impl {
   // Thread compatible part used only by lock under RunHandlerPool.
   // Handlers are sorted by start time.
   // TODO(azaks): sort by the remaining latency budget.
-  std::vector<RunHandler::Impl*> sorted_active_handlers_ TF_GUARDED_BY(mu_);
+  // TODO(chaox): Consider other data structure for maintaining the sorted
+  // active handlers if the searching overhead(currently O(n)) becomes the
+  // bottleneck.
+  std::list<RunHandler::Impl*> sorted_active_handlers_ TF_GUARDED_BY(mu_);
   std::vector<RunHandler::Impl*> free_handlers_ TF_GUARDED_BY(mu_);
   std::vector<std::unique_ptr<RunHandler::Impl>> handlers_ TF_GUARDED_BY(mu_);
+
   // Histogram of elapsed runtime of every handler (in ms).
   histogram::Histogram time_hist_ TF_GUARDED_BY(mu_);
 
@@ -1092,16 +1119,17 @@ void RunHandlerPool::Impl::LogInfo() {
     uint64 now = tensorflow::Env::Default()->NowMicros();
     string times_str = "";
     string ids_str = "";
+    auto it = sorted_active_handlers_.cbegin();
     for (int i = 0; i < num_active_requests; ++i) {
       if (i > 0) {
         times_str += " ";
         ids_str += " ";
       }
 
-      times_str += strings::StrCat(
-          (now - sorted_active_handlers_[i]->start_time_us()) / 1000.0, " ms.");
-      ids_str +=
-          strings::StrCat(sorted_active_handlers_[i]->tws()->GetTracemeId());
+      times_str +=
+          strings::StrCat((now - (*it)->start_time_us()) / 1000.0, " ms.");
+      ids_str += strings::StrCat((*it)->tws()->GetTracemeId());
+      ++it;
     }
     VLOG(1) << "Elapsed times are: " << times_str;
     VLOG(1) << "Step ids are: " << ids_str;
@@ -1127,7 +1155,7 @@ void RunHandler::Impl::ThreadPoolInterfaceWrapper::Schedule(
 RunHandler::Impl::Impl(RunHandlerPool::Impl* pool_impl)
     : pool_impl_(pool_impl) {
   thread_pool_interface_.reset(new ThreadPoolInterfaceWrapper(this));
-  Reset(0);
+  Reset(0, RunOptions::Experimental::RunHandlerPoolOptions());
 }
 
 void RunHandler::Impl::ScheduleInterOpClosure(std::function<void()> fn) {
@@ -1142,9 +1170,12 @@ void RunHandler::Impl::ScheduleIntraOpClosure(std::function<void()> fn) {
                                                         std::move(fn));
 }
 
-void RunHandler::Impl::Reset(int64 step_id) {
+void RunHandler::Impl::Reset(
+    int64 step_id,
+    const RunOptions::Experimental::RunHandlerPoolOptions& options) {
   start_time_us_ = tensorflow::Env::Default()->NowMicros();
   step_id_ = step_id;
+  options_ = options;
   tws_.SetTracemeId(step_id);
 }
 
@@ -1157,9 +1188,15 @@ RunHandlerPool::RunHandlerPool(int num_inter_op_threads,
 
 RunHandlerPool::~RunHandlerPool() {}
 
-std::unique_ptr<RunHandler> RunHandlerPool::Get(int64 step_id,
-                                                int64 timeout_in_ms) {
-  return impl_->Get(step_id, timeout_in_ms);
+std::unique_ptr<RunHandler> RunHandlerPool::Get(
+    int64 step_id, int64 timeout_in_ms,
+    const RunOptions::Experimental::RunHandlerPoolOptions& options) {
+  return impl_->Get(step_id, timeout_in_ms, options);
+}
+
+std::vector<int64> RunHandlerPool::GetActiveHandlerPrioritiesForTesting()
+    const {
+  return impl_->GetActiveHandlerPrioritiesForTesting();
 }
 
 RunHandler::RunHandler(Impl* impl) : impl_(impl) {}

tensorflow/core/framework/run_handler.h

@@ -62,7 +62,14 @@ class RunHandlerPool {
   // unique_ptr is destroyed.
   //
   // Will block unless there is an inactive handler.
-  std::unique_ptr<RunHandler> Get(int64 step_id = 0, int64 timeout_in_ms = 0);
+  std::unique_ptr<RunHandler> Get(
+      int64 step_id = 0, int64 timeout_in_ms = 0,
+      const RunOptions::Experimental::RunHandlerPoolOptions& options =
+          RunOptions::Experimental::RunHandlerPoolOptions());
+
+  // Get the priorities for active handlers. The return result is with the same
+  // order of the active handler list.
+  std::vector<int64> GetActiveHandlerPrioritiesForTesting() const;
 
  private:
   class Impl;

tensorflow/core/framework/run_handler_test.cc

@@ -78,6 +78,41 @@ TEST(RunHandlerUtilTest, TestBasicScheduling) {
   counter.Wait();
 }
 
+TEST(RunHandlerUtilTest, PrioritySchedulingTest) {
+  int num_threads = 2;
+  std::unique_ptr<RunHandlerPool> pool(
+      new RunHandlerPool(num_threads, num_threads));
+
+  RunOptions::Experimental::RunHandlerPoolOptions options =
+      RunOptions::Experimental::RunHandlerPoolOptions();
+  options.set_priority(2);
+  auto handler1 = pool->Get(/*step_id=*/1, /*timeout_in_ms=*/0, options);
+  options.set_priority(1);
+  auto handler2 = pool->Get(/*step_id=*/2, /*timeout_in_ms=*/0, options);
+  options.set_priority(3);
+  auto handler3 = pool->Get(/*step_id=*/3, /*timeout_in_ms=*/0, options);
+
+  // The active requests should be ordered by priorites.
+  std::vector<int64> sorted_active_list =
+      pool->GetActiveHandlerPrioritiesForTesting();
+  EXPECT_EQ(sorted_active_list.size(), 3);
+  EXPECT_EQ(sorted_active_list[0], 3);
+  EXPECT_EQ(sorted_active_list[1], 2);
+  EXPECT_EQ(sorted_active_list[2], 1);
+
+  handler1.reset();
+  options.set_priority(5);
+  auto handler4 = pool->Get(/*step_id=*/4, /*timeout_in_ms=*/0, options);
+  options.set_priority(4);
+  auto handler5 = pool->Get(/*step_id=*/5, /*timeout_in_ms=*/0, options);
+  sorted_active_list = pool->GetActiveHandlerPrioritiesForTesting();
+  EXPECT_EQ(sorted_active_list.size(), 4);
+  EXPECT_EQ(sorted_active_list[0], 5);
+  EXPECT_EQ(sorted_active_list[1], 4);
+  EXPECT_EQ(sorted_active_list[2], 3);
+  EXPECT_EQ(sorted_active_list[3], 1);
+}
+
 SessionOptions DefaultSessionOptions() {
   SessionOptions options;
   (*options.config.mutable_device_count())["CPU"] = 2;
@@ -205,6 +240,75 @@ TEST_F(RunHandlerTest, TestConcurrencyUseRunHandlerPool) {
   delete tp;
 }
 
+TEST_F(RunHandlerTest, UseRunHandlerPoolEnableSubPoolWithPriority) {
+  Initialize({3, 2, -1, 0});
+  auto session = CreateSession();
+  ASSERT_TRUE(session != nullptr);
+  EXPECT_EQ(::tensorflow::Status::OK(), session->Create(def_));
+  std::vector<std::pair<string, Tensor>> inputs;
+
+  // Request two targets: one fetch output and one non-fetched output.
+  std::vector<string> output_names = {y_ + ":0"};
+  std::vector<string> target_nodes = {y_neg_};
+  std::vector<Tensor> outputs;
+
+  // Prepares RunOptions and RunMetadata
+  RunOptions run_options;
+  run_options.mutable_experimental()->set_use_run_handler_pool(true);
+  run_options.mutable_experimental()
+      ->mutable_run_handler_pool_options()
+      ->set_priority(1);
+
+  Status s = session->Run(run_options, inputs, output_names, target_nodes,
+                          &outputs, nullptr);
+  EXPECT_EQ(::tensorflow::Status::OK(), s);
+
+  ASSERT_EQ(1, outputs.size());
+  // The first output should be initialized and have the correct
+  // output.
+  auto mat = outputs[0].matrix<float>();
+  ASSERT_TRUE(outputs[0].IsInitialized());
+  EXPECT_FLOAT_EQ(5.0, mat(0, 0));
+}
+
+TEST_F(RunHandlerTest, TestConcurrencyUseRunHandlerPoolWithPriority) {
+  Initialize({1, 2, 3, 4});
+  auto session = CreateSession();
+  ASSERT_TRUE(session != nullptr);
+  EXPECT_EQ(::tensorflow::Status::OK(), session->Create(def_));
+
+  // Fill in the input and ask for the output
+  thread::ThreadPool* tp = new thread::ThreadPool(Env::Default(), "test", 4);
+
+  // Run the graph 1000 times in 4 different threads concurrently.
+  std::vector<string> output_names = {y_ + ":0"};
+  auto fn = [&session, output_names]() {
+    for (int i = 0; i < 1000; ++i) {
+      RunOptions run_options;
+      run_options.mutable_experimental()->set_use_run_handler_pool(true);
+      run_options.mutable_experimental()
+          ->mutable_run_handler_pool_options()
+          ->set_priority(i % 4);
+      std::vector<std::pair<string, Tensor>> inputs;
+      std::vector<Tensor> outputs;
+      // Run the graph
+      Status s = session->Run(run_options, inputs, output_names, {}, &outputs,
+                              nullptr);
+      EXPECT_EQ(::tensorflow::Status::OK(), s);
+      ASSERT_EQ(1, outputs.size());
+      auto mat = outputs[0].matrix<float>();
+      EXPECT_FLOAT_EQ(3.0, mat(0, 0));
+    }
+  };
+
+  for (int i = 0; i < 4; ++i) {
+    tp->Schedule(fn);
+  }
+
+  // Wait for the functions to finish.
+  delete tp;
+}
+
 TEST_F(RunHandlerTest, TestWaitTimeout) {
   std::unique_ptr<RunHandlerPool> pool(new RunHandlerPool(1, 1));
 

tensorflow/core/protobuf/config.proto

@@ -640,6 +640,13 @@ message RunOptions {
     // and tail) latency.
     // Consider using this option for CPU-bound workloads like inference.
     bool use_run_handler_pool = 2;
+    // Options for run handler thread pool.
+    message RunHandlerPoolOptions {
+      // Priority of the request. The run handler thread pool will schedule ops
+      // based on the priority number. The larger number means higher priority.
+      int64 priority = 1;
+    }
+    RunHandlerPoolOptions run_handler_pool_options = 3;
   }
 
   Experimental experimental = 8;

tensorflow/tools/api/golden/v1/tensorflow.-run-options.-experimental.-run-handler-pool-options.pbtxt

@@ -0,0 +1,12 @@
+path: "tensorflow.RunOptions.Experimental.RunHandlerPoolOptions"
+tf_proto {
+  descriptor {
+    name: "RunHandlerPoolOptions"
+    field {
+      name: "priority"
+      number: 1
+      label: LABEL_OPTIONAL
+      type: TYPE_INT64
+    }
+  }
+}

tensorflow/tools/api/golden/v1/tensorflow.-run-options.-experimental.pbtxt

@@ -14,5 +14,21 @@ tf_proto {
       label: LABEL_OPTIONAL
       type: TYPE_BOOL
     }
+    field {
+      name: "run_handler_pool_options"
+      number: 3
+      label: LABEL_OPTIONAL
+      type: TYPE_MESSAGE
+      type_name: ".tensorflow.RunOptions.Experimental.RunHandlerPoolOptions"
+    }
+    nested_type {
+      name: "RunHandlerPoolOptions"
+      field {
+        name: "priority"
+        number: 1
+        label: LABEL_OPTIONAL
+        type: TYPE_INT64
+      }
+    }
   }
 }

tensorflow/tools/api/golden/v1/tensorflow.-run-options.pbtxt

@@ -61,6 +61,22 @@ tf_proto {
         label: LABEL_OPTIONAL
         type: TYPE_BOOL
       }
+      field {
+        name: "run_handler_pool_options"
+        number: 3
+        label: LABEL_OPTIONAL
+        type: TYPE_MESSAGE
+        type_name: ".tensorflow.RunOptions.Experimental.RunHandlerPoolOptions"
+      }
+      nested_type {
+        name: "RunHandlerPoolOptions"
+        field {
+          name: "priority"
+          number: 1
+          label: LABEL_OPTIONAL
+          type: TYPE_INT64
+        }
+      }
     }
     enum_type {
       name: "TraceLevel"