[
https://issues.apache.org/jira/browse/ARROW-8774?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17105524#comment-17105524
]
Adam Lippai edited comment on ARROW-8774 at 5/12/20, 3:29 PM:
--------------------------------------------------------------
A thing to consider on the long run (but a single default threadpool now helps
a lot!):
For an optimal result reading files should have less priority than the
computation part later.
This is what I've found on the topic:
[https://users.rust-lang.org/t/dealing-with-work-priority-and-rayon/30954/2]
Rayon doesn't support setting priority for the tasks, but as a workaround we
could create two threadpools, eg one with <=10 threads for file reading and
CPU_NUM threads for the computation. If you need fine tuning the workload (S3,
HDFS, NFS behave differently, local HDD or SSD is a different topic too) you
could either configure the threadpool sizes (even down to 1 thread) or setting
"nice" for the threadpool threads.
Should I add this to the design doc or is this out of scope for a while?
was (Author: alippai):
A thing to consider on the long run (but a single default threadpool now helps
a lot!):
For an optimal result reading files should have less priority than the
computation part later.
This is what I've found on the topic:
[https://users.rust-lang.org/t/dealing-with-work-priority-and-rayon/30954/2]
Rayon doesn't support setting priority for the tasks, but as a workaround we
could create two threadpools, eg one with <=10 threads for file reading and
CPU_NUM threads for the computation. If you need fine tuning the workload (S3,
HDFS, NFS behave differently, local HDD or SSD is a different topic too) you
could either configure the threadpool sizes (even down to 1 thread) or setting
"nice" for the threadpool threads.
> [Rust] [DataFusion] Improve threading model
> -------------------------------------------
>
> Key: ARROW-8774
> URL: https://issues.apache.org/jira/browse/ARROW-8774
> Project: Apache Arrow
> Issue Type: Improvement
> Components: Rust, Rust - DataFusion
> Reporter: Andy Grove
> Assignee: Andy Grove
> Priority: Major
> Fix For: 1.0.0
>
>
> DataFusion currently spawns one thread per partition and this results in poor
> performance if there are more partitions than available cores/threads. It
> would be better to have a thread-pool that defaults to number of available
> cores.
> Here is a Google doc where we can collaborate on a design discussion.
> https://docs.google.com/document/d/1_wc6diy3YrRgEIhVIGzrO5AK8yhwfjWlmKtGnvbsrrY/edit?usp=sharing
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
