Initial changes to support cufile stream I/O. #259
Conversation
|
Pull requests from external contributors require approval from a |
|
/ok to test |
madsbk
added
improvement
There was a problem hiding this comment.
Thanks @tell-rebanta. Overall, I think it looks good, but I am wondering if it makes sense to de-couple the CUDA stream from StreamHandle?
As it is now, you are forced to create a stream per file, but I think a common use case is to read/write to multiple files using the same stream.
What about implementing read_async() and write_async() in the existing class FileHandle instead? read_async() and write_async() could then take a CUstream as argument.
Side note:
To check and fix the check-style run:
pre-commit run --all-file
cpp/examples/basic_io.cpp
Outdated
|
|
||
| kvikio::StreamHandle s_handle_wr("/tmp/test-file", "w", kvikio::StreamHandle::m644, true); | ||
| check(cudaMemcpy(a_dev, a, SIZE, cudaMemcpyHostToDevice) == cudaSuccess); | ||
| kvikio::buffer_register(a_dev, SIZE); |
There was a problem hiding this comment.
Do we need to register the a buffer in order to use it with stream IO?
- Moved async I/O APIs to file_handle and made stream_handle to be a derived class of file_handle. - Async I/O can be invoked either through file_handle or stream_handle. - Addeed basic I/O examples for both the ways. - Removed unnecessary logs in CMakeLists.txt.
There was a problem hiding this comment.
Thanks @tell-rebanta, I like that the stream is now an explicit argument for read_async and write_async!
We need to guard the stream API using KVIKIO_CUFILE_STREAM_API_FOUND and I suggest that we exclude StreamHandle from this PR.
Let's discuss how to design the stream registration in a follow-up PR (I think we need to de-couple it from the file handle).
|
/ok to test |
Will be added as part of a separate check-in.
|
/ok to test |
| * @param size Size in bytes to read. | ||
| * @param file_offset Offset in the file to read from. | ||
| * @param devPtr_offset Offset relative to the `devPtr_base` pointer to read into. | ||
| * This parameter should be used only with registered buffers. |
There was a problem hiding this comment.
Is there a way to distinguish registered buffers in the type system at the moment? That would be preferable over this void * C-like interface.
There was a problem hiding this comment.
No, not ATM but it would be a cool addition: #266
|
/ok to test |
Co-authored-by: Lawrence Mitchell <[email protected]>
|
/ok to test |
|
Thanks @wence-, added you suggestions. |
|
/ok to test |
|
/merge |
Hi there, Thanks for this great repository! I want to use the cuFile async IO in my research project and noticed this kvikio repo. However, the initial support has been done in #259 and tracked in #204, but the Python interface hasn't been done yet. So I exported the write_async and read_async to the CuFile Python class and added test case. This will be very helpful for my project where I want to do the PyTorch training computation and simultaneously load tensors from the SSDs. I created this PR because hopefully, it could be helpful for your repository as well as keeping the Python interface current. Please let me know your thoughts. Thank you. Best Regards, Kun Authors: - Kun Wu (https://github.com/K-Wu) - Mads R. B. Kristensen (https://github.com/madsbk) Approvers: - Mads R. B. Kristensen (https://github.com/madsbk) URL: #376
No description provided.