diff --git a/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h
index 0b9351a67635ed2bf49eeb60b5348fd1a4f62ade..eadd842c7f6ead56006fd0c34814b1b7bd9b62f4 100644
--- a/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h
+++ b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h
@@ -57,8 +57,6 @@ class ConcurrentSet {
   std::future<void> GetAndClear(std::vector<int64_t>* result) {
     auto task = [this, &result] {
       result->clear();
-      // FIXME(qiao): use a trick to avoid the bug of recv an selected rows
-      result->push_back(0);
       for (auto& id : set_) {
         result->push_back(id);
       }
@@ -125,8 +123,7 @@ class AsyncSparseParamUpdateRecorder {
       fs.push_back(set->Update(update_rows));
     }
     for (auto& f : fs) {
-      // no need to wait here because GetAndClear will wait.
-      // f.wait();
+      f.wait();
     }
   }
 
diff --git a/paddle/fluid/operators/distributed/parameter_recv.cc b/paddle/fluid/operators/distributed/parameter_recv.cc
index e9d34b173942b1ae27be3515cd9862f0b5984af0..a387ee5ee193122dcd4e3f42eafbbd078bca4b8d 100644
--- a/paddle/fluid/operators/distributed/parameter_recv.cc
+++ b/paddle/fluid/operators/distributed/parameter_recv.cc
@@ -109,8 +109,7 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
                   << sstream.str();
         }
 
-        // FIXME(qiao): use a trick to avoid the bug of recv an selected rows
-        for (auto i = 1; i < recv_slr.rows().size(); ++i) {
+        for (auto i = 0; i < recv_slr.rows().size(); ++i) {
           auto row_id = recv_slr.rows()[i] + row_offset;
           PADDLE_ENFORCE_LT(row_id, recv_dims[0]);
           memcpy(recv_tensor->data<T>() + row_id * width,