[
https://issues.apache.org/jira/browse/RATIS-2387?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18054505#comment-18054505
]
Sergey Soldatov commented on RATIS-2387:
----------------------------------------
Here are the numbers for a simple 3-node deployment using LXC containers. Each
node is using a dedicated NVMe SSD. Tests are uploading 5/10/50Gb file using
ozone fs -copy.
Without patch:
File Size | Avg Upload
------------ | --------------------
5GB | 17.65s ( 290.10 MB/s)
10GB | 46.24s ( 221.47 MB/s)
50GB | 219.61s ( 233.14 MB/s)
With patch:
File Size | Avg Upload
------------ | --------------------
5GB | 11.37s ( 450.40 MB/s)
10GB | 29.43s ( 347.95 MB/s)
50GB | 153.44s ( 333.68 MB/s)
With a physical network in place (deployment on dedicated servers), the
difference might be less visible but still noticeable.
> Performance degradation after RATIS-2235
> ----------------------------------------
>
> Key: RATIS-2387
> URL: https://issues.apache.org/jira/browse/RATIS-2387
> Project: Ratis
> Issue Type: Improvement
> Components: common
> Affects Versions: 3.2.1
> Reporter: Sergey Soldatov
> Priority: Major
> Attachments: 1337_review.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> RATIS-2235 made append processing synchronized. That made Apache Ozone write
> operations slower (up to 30% for relatively slow drives). We haven't seen any
> issues with parallel processing, so it might be reasonable to make it
> optional.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
