[
https://issues.apache.org/jira/browse/HDDS-10863?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17853758#comment-17853758
]
Wei-Chiu Chuang edited comment on HDDS-10863 at 6/10/24 5:53 PM:
-----------------------------------------------------------------
Hsync throughput microbenchmark (haven't looked at latency yet)
run freon dfsg, 32 clients, 1 byte each hsync, 100 bytes long files.
{noformat}
dfsg --buffer=1 --copy-buffer=1 -n 100000 --path
ofs://ozone1715984004/vol1/bucket1/dfsg -s 100 -t 32 --sync HSYNC -p abcde
{noformat}
For hdds.ratis.raft.grpc.server.leader.outstanding.appends.max =
1: mean rate = 28.83 calls/second
4: mean rate = 45.70 calls/second
8: mean rate = 57.16 calls/second
16: mean rate = 48.97 calls/second
32: mean rate = 55.12 calls/second
64: mean rate = 51.31 calls/second
128: mean rate = 37.89 calls/second
Conclusion: as far as hsync is concerned, the default value seems okay.
was (Author: jojochuang):
Hsync throughput microbenchmark (haven't looked at latency yet)
run freon dfsg, 32 clients, 1 byte each hsync, 100 bytes long files.
{noformat}
dfsg --buffer=1 --copy-buffer=1 -n 100000 --path
ofs://ozone1715984004/vol1/bucket1/dfsg -s 100 -t 32 --sync HSYNC -p abcde
{noformat}
For hdds.ratis.raft.grpc.server.leader.outstanding.appends.max =
1: mean rate = 28.83 calls/second
4: mean rate = 45.70 calls/second
8: mean rate = 57.16 calls/second
16: mean rate = 48.97 calls/second
32: mean rate = 55.12 calls/second
64: mean rate = 51.31 calls/second
128: mean rate = 37.89 calls/second
> Changed default value of raft.grpc.server.leader.outstanding.appends.max may
> have performance implication in Ozone
> ------------------------------------------------------------------------------------------------------------------
>
> Key: HDDS-10863
> URL: https://issues.apache.org/jira/browse/HDDS-10863
> Project: Apache Ozone
> Issue Type: Bug
> Reporter: Duong
> Assignee: Duong
> Priority: Major
>
> RATIS-1804 changes the default number of
> raft.grpc.server.leader.outstanding.appends.max from 128 to 8. This seems to
> imply an impact on Ozone performance because:
> # More smaller appendEntries requests are sent over the network.
> # The flush batch also becomes smaller because of fewer available log
> entries at the same time.
> I tested with it and saw it lower write throughput significantly (~20%).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
