msr-fiddle / ds-analyzer Goto Github PK

Great work on the data stalls paper!
Reading the paper, I was wondering what is the advantage of using a simple non-evicting cache (minio) vs pinning part of the dataset in-memory.
I didn't find any mention of this alternative in the paper.

Hello, I couldn't find the right version of torchaudio compatible with PyTorch 1.1.0 which is installed in base_image. So, could you show me M5 model and its training code with PyTorch? Thanks a lot!

Hi, in the paper it is mentioned in section 6.1 :

"Once the cache capacity is reached, MinIO will not evict any items in the cache; instead, the requests to other data items default
to storage accesses. The items in the MinIO cache survive across epochs until the end of the training job"

Going through the code, I was not able to find this implementation.

My understanding is that you implemented "MinIO" as a software cache, i.e., the data items are stored in the Anonymous Memory of the torch program, and when this capacity reaches a specified threshold, you use the O_DIRECT flag to bypass the OS page cache.

Could you point me to the code that implements this, or, if my understanding is incorrect, can you describe how this is achieved?

Thanks