diff --git a/src/master/taskmgr.py b/src/master/taskmgr.py
index 6428355..3e7d06d 100644
--- a/src/master/taskmgr.py
+++ b/src/master/taskmgr.py
@@ -113,6 +113,7 @@ class SubTask():
         self.status_reason = ''
         self.try_count = 0
         self.worker = None
+        self.lock = threading.Lock()
 
     def waiting_for_retry(self,reason=""):
         self.try_count += 1
@@ -120,7 +121,7 @@ class SubTask():
         if self.status == FAILED:
             self.root_task.status = FAILED
             self.failed_reason = reason
-
+            self.root_task.failed_reason = reason
 
 class TaskReporter(MasterServicer):
 
@@ -197,6 +198,19 @@ class TaskMgr(threading.Thread):
             return new_f
         return lock
 
+    def subtask_lock(f):
+        @wraps(f)
+        def new_f(self, subtask, *args, **kwargs):
+            subtask.lock.acquire()
+            try:
+                result = f(self, subtask, *args, **kwargs)
+            except Exception as err:
+                subtask.lock.release()
+                raise err
+            subtask.lock.release()
+            return result
+        return new_f
+
     def run(self):
         self.serve()
         while not self.thread_stop:
@@ -272,7 +286,10 @@ class TaskMgr(threading.Thread):
         self.gpu_usage[subtask.worker] += subtask.vnode_info.vnode.instance.gpu
         return [True, '']
 
+    @subtask_lock
     def stop_vnode(self, subtask):
+        if not subtask.vnode_started:
+            return [True, ""]
         try:
             self.logger.info('[task_processor] Stopping vnode for task [%s] vnode [%d]' % (subtask.vnode_info.taskid, subtask.vnode_info.vnodeid))
             channel = grpc.insecure_channel('%s:%s' % (subtask.worker, self.worker_port))
@@ -484,11 +501,15 @@ class TaskMgr(threading.Thread):
 
         sub_task.status = report.subTaskStatus
         sub_task.status_reason = report.errmsg
+        sub_task.task_started = False
 
         if report.subTaskStatus == FAILED or report.subTaskStatus == TIMEOUT:
+            self.clear_sub_task(sub_task)
             sub_task.waiting_for_retry(report.errmsg)
-            self.jobmgr.report(task.username, task.id, 'retrying', report.errmsg)
+            if sub_task.status == WAITING:
+                self.jobmgr.report(task.username, task.id, 'retrying', report.errmsg)
         elif report.subTaskStatus == OUTPUTERROR:
+            self.clear_sub_task(sub_task)
             sub_task.status = FAILED
             task.status = FAILED
             task.failed_reason = report.errmsg
@@ -506,7 +527,7 @@ class TaskMgr(threading.Thread):
             self.logger.info('task %s sub_tasks %s' % (task.id, str([sub_task.status for sub_task in task.subtask_list])))
             if self.check_task_completed(task):
                 continue
-            self.logger.info("test")
+            self.logger.info('schedule task %s sub_tasks %s' % (task.id, str([sub_task.status for sub_task in task.subtask_list])))
 
             if task.at_same_time:
                 # parallel tasks
@@ -689,7 +710,7 @@ class TaskMgr(threading.Thread):
                     timeout = int(json_task['expTime'])
                 # commands are executed in all vnodes / only excuted in the first vnode
                 # if in traditional mode, commands will be executed in all vnodes
-                ) if (not 'atSameTime' in json_task.keys() or json_task['runon'] == 'all' or vnode_index == 0) else None
+                ) if (json_task['runon'] == 'all' or vnode_index == 0) else None
             } for vnode_index in range(int(json_task['vnodeCount']))])
 
         if task.at_same_time: