diff --git a/python/paddle/distributed/utils.py b/python/paddle/distributed/utils.py
index 3b55ec7ffce4df7e96034df067f667f2c807943f..511d501a2206e3ce106b7b76a5f3463b48353ff4 100644
--- a/python/paddle/distributed/utils.py
+++ b/python/paddle/distributed/utils.py
@@ -252,15 +252,12 @@ def get_cluster(node_ips, node_ip, paddle_ports, selected_gpus):
 def terminate_local_procs(procs):
     for p in procs:
         if p.proc.poll() is None:
-            # subprocess need to release resource(e.g. shared memory)
-            # use join to wait subprocess releasing
-            p.proc.join(timeout=1)
+            p.proc.terminate()
             p.log_fn.close()
             logger.debug("terminate process id:{}".format(p.proc.pid))
 
-    # wait all process terminiated
-    # time.sleep(3)
-
+    #wait all process terminiated
+    time.sleep(3)
     for step in range(0, 50):
         alive = False
         for p in procs: