diff --git a/src/master/taskmgr.py b/src/master/taskmgr.py
index 9c3587b..aea04da 100644
--- a/src/master/taskmgr.py
+++ b/src/master/taskmgr.py
@@ -59,6 +59,7 @@ class TaskMgr(threading.Thread):
         self.nodemgr = nodemgr
         self.monitor_fetcher = monitor_fetcher
         self.cpu_usage = {}
+        self.gpu_usage = {}
         # self.all_nodes = None
         # self.last_nodes_info_update_time = 0
         # self.nodes_info_update_interval = 30 # (s)
@@ -119,6 +120,7 @@ class TaskMgr(threading.Thread):
 
         if instance['status'] == RUNNING and report.instanceStatus != RUNNING:
             self.cpu_usage[instance['worker']] -= task.info.cluster.instance.cpu
+            self.gpu_usage[instance['worker']] -= task.info.cluster.instance.gpu
 
         instance['status'] = report.instanceStatus
         instance['error_msg'] = report.errmsg
@@ -197,6 +199,7 @@ class TaskMgr(threading.Thread):
         instance['worker'] = worker_ip
 
         self.cpu_usage[worker_ip] += task.info.cluster.instance.cpu
+        self.gpu_usage[worker_ip] += task.info.cluster.instance.gpu
         username = task.info.username
         container_name = task.info.username + '-batch-' + task.info.id + '-' + str(instance_id) + '-' + task.info.token
         if not username in self.user_containers.keys():
@@ -251,6 +254,7 @@ class TaskMgr(threading.Thread):
                         instance['status'] = FAILED
                         instance['token'] = ''
                         self.cpu_usage[instance['worker']] -= task.info.cluster.instance.cpu
+                        self.gpu_usage[instance['worker']] -= task.info.cluster.instance.gpu
 
                         self.logger.warning('[task_scheduler] worker dead, retry task [%s] instance [%d]' % (task.info.id, index))
                         if worker is not None:
@@ -282,7 +286,7 @@ class TaskMgr(threading.Thread):
             # try not to assign non-gpu task to a worker with gpu
             if task.info.cluster.instance.gpu == 0 and worker_info['gpu'] > 0:
                 continue
-            if task.info.cluster.instance.gpu > worker_info['gpu']:
+            if task.info.cluster.instance.gpu + self.get_gpu_usage(worker_ip) > worker_info['gpu']:
                 continue
             return worker_ip
         return None
@@ -322,6 +326,14 @@ class TaskMgr(threading.Thread):
             return 0
 
 
+    def get_gpu_usage(self, worker_ip):
+        try:
+            return self.gpu_usage[worker_ip]
+        except:
+            self.gpu_usage[worker_ip] = 0
+            return 0
+
+
     def set_jobmgr(self, jobmgr):
         self.jobmgr = jobmgr