diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py
index d0a72b92d9d8b35c5087f44947505e83d55e6cf5..5f3ce300acc44ad8d2898c27296b866c403f3cc8 100644
--- a/benchmark/fluid/fluid_benchmark.py
+++ b/benchmark/fluid/fluid_benchmark.py
@@ -168,7 +168,7 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
     startup_exe = fluid.Executor(place)
     startup_exe.run(startup_prog)
     strategy = fluid.ExecutionStrategy()
-    strategy.num_threads = 0  #args.cpus
+    strategy.num_threads = args.cpus
     strategy.allow_op_delay = False
     build_strategy = fluid.BuildStrategy()
     if args.reduce_strategy == "reduce":
@@ -188,8 +188,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
         num_trainers = 1
         trainer_id = 0
 
-    print('Use parallel_executor')
-    strategy.type = 2
     exe = fluid.ParallelExecutor(
         True,
         avg_loss.name,
diff --git a/benchmark/fluid/models/resnet.py b/benchmark/fluid/models/resnet.py
index 947c497ce2b9becd1282b62187f743008985a17f..f692e7722a1c9a54a4509ce7c78cc68e1f28da74 100644
--- a/benchmark/fluid/models/resnet.py
+++ b/benchmark/fluid/models/resnet.py
@@ -172,7 +172,7 @@ def get_model(args, is_train, main_prog, startup_prog):
     reader, dshape, class_dim = _model_reader_dshape_classdim(args, is_train)
 
     pyreader = None
-    trainer_count = int(os.getenv("PADDLE_TRAINERS", 1))
+    trainer_count = int(os.getenv("PADDLE_TRAINERS"))
     with fluid.program_guard(main_prog, startup_prog):
         with fluid.unique_name.guard():
             if args.use_reader_op: