@@ -93,6 +93,7 @@ void DumpQueue::DispatchOneTask(Trans::Stream* stream, size_t tensorSize, TaskPa
9393
9494Status DumpQueue::DumpOneTask (Trans::Stream* stream, size_t tensorSize, TaskPtr task)
9595{
96+ auto tp = NowTime::Now ();
9697 Detail::TaskDesc backendTaskDesc;
9798 backendTaskDesc.brief = " Cache2Backend" ;
9899 const auto nShard = task->desc .size ();
@@ -117,12 +118,14 @@ Status DumpQueue::DumpOneTask(Trans::Stream* stream, size_t tensorSize, TaskPtr
117118 Detail::Shard{shard.owner , shard.index , {shardTask.bufferHandle .Data ()}});
118119 backendTaskIndex.emplace_back (i);
119120 }
121+ auto tpMakeBuffer = NowTime::Now ();
120122 if (backendTaskIndex.empty ()) { return Status::OK (); }
121123 auto s = stream->Synchronized ();
122124 if (s.Failure ()) [[unlikely]] {
123125 UC_ERROR (" Failed({}) to sync on stream." , s);
124126 return s;
125127 }
128+ auto tpSyncStream = NowTime::Now ();
126129 for (const auto & i : backendTaskIndex) { shardTasks[i].bufferHandle .MarkReady (); }
127130 auto res = backend_->Dump (std::move (backendTaskDesc));
128131 if (!res) [[unlikely]] {
@@ -134,6 +137,9 @@ Status DumpQueue::DumpOneTask(Trans::Stream* stream, size_t tensorSize, TaskPtr
134137 shardTask.backendTaskHandle = res.Value ();
135138 dumping_.Push (std::move (shardTask));
136139 }
140+ auto tpEnd = NowTime::Now ();
141+ UC_DEBUG (" Cache task({}) mk_buf={:06f}s, sync={:06f}s, back={:06f}s." , task->id ,
142+ tpMakeBuffer - tp, tpSyncStream - tpMakeBuffer, tpEnd - tpSyncStream);
137143 return Status::OK ();
138144}
139145
0 commit comments