Hi,

I'm starting this thread to collect issues that are likely caused by various overhead in MXNet execution so that we can focus the efforts in solving them. The focus is everything other than the operators and kvstores.

If you've seen performance issues that you found were due to the overhead in MXNet execution (e.g. ctypes/FFI call, asynchronous execution, graph construction overhead, memory release and clean-up), please either report it in comments or link relevant issue.

If you're unsure, please create a performance issue report in another github issue instead.

Be sure to include these details:

• the performance issue.
• what you tried to solve it.
• timings you've collected if any.