Add wait

paddle/fluid/framework/parallel_executor.cc

@@ -813,6 +813,16 @@ void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
   fetch_ops.clear();
   *member_->global_scope_->Var(fetched_var_name)->GetMutable<LoDTensorArray>() =
       fetched_data->tensors_;
+  // FIXME:
+  // It could be optimized by using multiple events in an operator.
+  // Manually sync computation during iter.
+  for (auto &s : member_->communication_streams_) {
+    s.second.ctx_->Wait();
+  }
+
+  for (auto &p : member_->places_) {
+    platform::DeviceContextPool::Instance().Get(p)->Wait();
+  }
 }
 
 void ParallelExecutor::RunOp(