Question about cudaMemcpyAsync measurement in nsys and async meaning

[sakaia]

I have a question about cudaMemcpyAsync measurement and its naming.

1. For transfering D to H and H to D, a host is blocked by cudaMemcpyAsync. The measurement value by nsys is real occupied time on the host. Is my understanding correct?

2. What is the meaning of async for cudaMemcpyAsync? (the API document says misnormer)

Reference

https://docs.nvidia.com/cuda/cuda-runtime-api/api-sync-behavior.html#api-sync-behavior

Questions about when using cudaMemcpyAsync(), the host is blocked

Dear all, I have met a problem when I am trying to use multiple different streams in different thread (in the same process). When I test the processing time of cudaMemcpyAsync(), I found that sometimes the host is blocked with this functions. Why could this situation happens? And how can I avoid this situation? Thanks a lot.