diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc
index 7e3f015dabdb3..7f4da99c3e5a1 100644
--- a/paddle/fluid/operators/detail/sendrecvop_utils.cc
+++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc
@@ -80,7 +80,9 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
         auto& gpu_dev_ctx =
             static_cast<const platform::CUDADeviceContext&>(ctx);
         auto copy_size = tensor.memory_size();
-        payload = memory::Alloc(cpu, copy_size);
+        // FIXME(typhoonzero): make sure data can be copied use DMA
+        // before doing the copy.
+        payload = memory::Alloc(cpu, copy_size, true);
 
         memory::Copy(cpu, payload,
                      boost::get<platform::CUDAPlace>(tensor.place()),
@@ -116,7 +118,8 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
         auto& gpu_dev_ctx =
             static_cast<const platform::CUDADeviceContext&>(ctx);
         auto copy_size = tensor->memory_size();
-        payload = memory::Alloc(cpu, copy_size);
+        // FIXME(typhoonzero): make sure data can be copied use DMA.
+        payload = memory::Alloc(cpu, copy_size, true);
         memory::Copy(cpu, payload,
                      boost::get<platform::CUDAPlace>(tensor->place()),
                      reinterpret_cast<const void*>(tensor->data<void>()),