NVIDIA / NVIDIA/cudf

[PERF]: Establish a default for max_concurrent_io_tasks on local scans

Open
#23,864 0 comments 0 reactions 0 assignees View on GitHub
cudf-polars Performance
Dominant language
C++
Stars
9.8k
Forks
1.1k
Avg merge
3d 6m
Merged PRs (30d)
278

Description

`max_concurrent_io_tasks` defaults to 2 for all scan types. PR #23847 adds separate defaults for local and remote paths but does not establish what the local value should be.

We need a sweep at SF300–SF1000 with data on NVMe, varying `max_concurrent_io_tasks` across 2, 4, and 8. Naively, more concurrency should help or at least be neutral, but we need to confirm this and understand the penalty if it is not. Any query that regresses should be profiled with nsys so we can identify where the contention is coming from.

## Related

- PR #23847
- #23810

Contributor guide

Open the contributing guide

Research direction

Start with PR #23847 and issue #23810 to understand the separate local and remote defaults. Run the SF300–SF1000 sweep on NVMe with max_concurrent_io_tasks set to 2, 4, and 8, then use nsys to profile any regressions and identify contention. Done means the local default is supported by the benchmark data and regressions are explained.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
performance
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.