fix a bug when device info not exists in json format (#1166) (#1176)

visualdl/component/profiler/parser/event_node.py

@@ -265,10 +265,16 @@ class ProfilerResult:
     def parse_json(self, json_data):
         self.schema_version = json_data['schemaVersion']
         self.span_idx = json_data['span_indx']
-        self.device_infos = {
-            device_info['id']: device_info
-            for device_info in json_data['deviceProperties']
-        }
+        try:
+            self.device_infos = {
+                device_info['id']: device_info
+                for device_info in json_data['deviceProperties']
+            }
+        except Exception:
+            print(
+                "paddlepaddle-gpu version is needed to get GPU device informations."
+            )
+            self.device_infos = {}
         hostnodes = []
         runtimenodes = []
         devicenodes = []

visualdl/component/profiler/profiler_data.py

@@ -1767,6 +1767,8 @@ class DistributedProfilerData:
         data = []
         for profile_data in self.profile_datas:
             device_infos = profile_data.device_infos
+            if not device_infos:
+                return data
             gpu_id = int(next(iter(profile_data.gpu_ids)))
             data.append({
                 'worker_name':

visualdl/component/profiler/profiler_reader.py

@@ -14,6 +14,7 @@
 # =======================================================================
 import os
 import re
+from threading import Lock
 from threading import Thread
 
 import packaging.version
@@ -28,6 +29,7 @@ from .run_manager import RunManager
 from visualdl.io import bfile
 
 _name_pattern = re.compile(r"(.+)_time_(.+)\.paddle_trace\.((pb)|(json))")
+_lock = Lock()
 
 
 def is_VDLProfiler_file(path):
@@ -118,8 +120,10 @@ class ProfilerReader(object):
                 self.run_managers[run] = RunManager(run)
             self.run_managers[run].set_all_filenames(filenames)
             for filename in filenames:
-                if self.run_managers[run].has_handled(filename):
-                    continue
+                with _lock:  # we add this to prevent parallel requests for handling a file multiple times
+                    if self.run_managers[run].has_handled(filename):
+                        continue
+                    self.run_managers[run].handled_filenames.add(filename)
                 self._read_data(run, filename)
         return list(self.walks.keys())
 

visualdl/component/profiler/profiler_server.py

@@ -194,6 +194,8 @@ class ProfilerApi(object):
         run_manager = self._reader.get_run_manager(run)
         distributed_profiler_data = run_manager.get_distributed_profiler_data(
             span)
+        if distributed_profiler_data is None:
+            return
         return distributed_profiler_data.get_distributed_steps()
 
     @result()
@@ -201,6 +203,8 @@ class ProfilerApi(object):
         run_manager = self._reader.get_run_manager(run)
         distributed_profiler_data = run_manager.get_distributed_profiler_data(
             span)
+        if distributed_profiler_data is None:
+            return
         return distributed_profiler_data.get_distributed_histogram(
             step, time_unit)
 

visualdl/component/profiler/run_manager.py

@@ -104,11 +104,8 @@ class RunManager:
         return
 
     def join(self):
-        if self.has_join:
-            return
         for thread in self.threads.values():
             thread.join()
-        self.has_join = True
         distributed_profiler_data = defaultdict(list)
         for worker_name, span_data in self.profiler_data.items():
             for span_idx, profiler_data in span_data.items():