Why doesn’t xsched support preemption for the CUDA default stream?

[KouweiLee]

Dear author, I noticed that kernels submitted to the CUDA default stream are submitted directly, instead of being buffered into the Xqueue like kernels from other streams.

Could you please explain the rationale behind this special handling of the default stream? Is there a technical or design reason why preemption is not supported for the CUDA default stream in xsched?

[Naplesoul]

Hi!

I appreciate your interest! Supporting CUDA default stream within XSched is technically feasible, but necessitates meticulous additional handling.

Because XSched requires internal streams (such as the op_stream_ within the Instrumentation module) to operate concurrently with user-defined streams, the implicit synchronization semantics of the default stream may induce unexpected deadlocks. Thus, for simplicity, our current prototype does not yet support scheduling of the default stream. But we have incorporate support for the default stream in our roadmap.