[
https://issues.apache.org/jira/browse/RATIS-2259?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17937628#comment-17937628
]
Tsz-wo Sze commented on RATIS-2259:
-----------------------------------
[~andywu], sorry, I forgot answer you question.
{quote}... will increasing the number of clients for a specific raft group help
with the latency?
{quote}
The RaftClient current requires protobuf serialization and also buffer copying
(if you are using gRPC). We do see that performance impact. Using multiple
clients could help.
{quote}Could you try benchmarking the server (bypass the RaftClient, call the
server methods directly)?
We put a timer on the Server Side, and the p999 for the server is ~1 ms. Sever
side seems fine.
{quote}
It is still worth to run the followin benchmarks:
# Without RaftClient (calling the server methods directly)
# With RaftClient in a local machine (eliminate network performance)
These are baselines to compare.
> Improve the read latency of RaftClient
> --------------------------------------
>
> Key: RATIS-2259
> URL: https://issues.apache.org/jira/browse/RATIS-2259
> Project: Ratis
> Issue Type: Improvement
> Components: client
> Reporter: Andy Wu
> Priority: Major
>
> Hey Folks,
> We are using ratis library to perform around ~2K/s read operations. Here is
> the API we are using:
> {code:java}
> RaftClient#sendReadOnlyAsync ==> it is OrderedAsync{code}
> I understand the above API is ordered, it will be filed sequentially. The
> read latency is keeping going up for each cid:
> e.g.
> duration: 15 ms, cid=2
> duration: 20 ms, cid=3
> …..
> duration: 100 ms, cid=100
> For the read operations, the order is not important to us. How do we improve
> the latency while keeping the throughput high? I tried *UnorderedAsync,* it
> has similar behaviors. Anything I missed?
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
