mlcommons / mlcommons/storage

[Passing score blocker] `ParquetReader` involves a CPU time in `get_sample` 5x / 4x times longer than the I/O time

Open
#323 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug Future
Dominant language
Python
Stars
203
Forks
67
Avg merge
20m
Merged PRs (30d)
8

Description

Testing flux model on multiple CPU architectures and collecting traces with DFTRACER, the time for getting the next sample with ParquetReader is spent for approximately 80% on the deserialization part rather than in I/O.

For the case of UNET3D and NPZ reader, this was not higher than 20-30%

Also, there seems to be a considerable impact of the CPU architecture in the deserialization time.

This may impact the calculation of the Accelerator Utilization % as mainly influenced by elements not strictly related to storage performance.

This could potentially impact other models using ParquetReader

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the reported ParquetReader timing with the flux model and DFTRACER, comparing deserialization with I/O and with the UNET3D and NPZ reader cases. Done means the deserialization overhead and CPU-architecture impact are measured, with the effect on Accelerator Utilization % assessed.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.