cuFile Async APIs

Are the APIs cuFileReadAsync and cuFileWriteAsync ready to use? If not, when will they be staged? Thanks.

Is there any update on when the async API will be available?

We have added support for batch apis (still in Beta) as part of 11.6. But, if you are asking for cuFile API Reference Guide :: NVIDIA GPUDirect Storage Documentation then we do not have it today.

We would love to know more about your use case.

The batch API would also work for my use case, but when I tried it I ran into stability issues, (lots of memory corruption, malloc was not happy). Once it stabilizes it may work for me. The primary downside compared to the async API is that I need to also poll for completions and deregister batch handles. I am using 1.3.0.44, and I just noticed that the release notes for 1.3.1 include bug fixes for the batch API, so I may upgrade and try again.

I am working on a research project for GPU+flash accelerated database systems, and being able to string together reads + decompression (through nvcomp) + database operators on a stream would be ideal.