[
https://issues.apache.org/jira/browse/ARROW-15229?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17477578#comment-17477578
]
Yibo Cai edited comment on ARROW-15229 at 1/18/22, 9:13 AM:
------------------------------------------------------------
I tested on two hosts (xeon gold as server, neoverse n1 as client), connected
by 100Gb network (mellanox connectx-5). FlightRPC DoGet benchmark speed is 2634
MB/s for gRPC, and 3874 MB/s for UCX/RDMA. Looks great. It's even faster than
localhost (2834 on server, 3359 MB/s on client).
Please note current code is not stable for RDMA. Most benchmark tests hang,
only very few tries happen to work. And UCX/TCP benchmark (1738 MB/s) is much
worse than gRPC. I think there are still big potentials to improve.
Find a way to stably test UCX/RDMA.
- start server
- start client, wait some time (~10 seconds), ctrl-c
- start client again, now it works okay afterwards
was (Author: yibo):
I tested on two hosts (xeon gold as server, neoverse n1 as client), connected
by 100Gb network (mellanox connectx-5). FlightRPC DoGet benchmark speed is 2634
MB/s for gRPC, and 3874 MB/s for UCX/RDMA. Looks great. It's even faster than
localhost (2834 on server, 3359 MB/s on client).
Please note current code is not stable for RDMA. Most benchmark tests hang,
only very few tries happen to work. And UCX/TCP benchmark (1738 MB/s) is much
worse than gRPC. I think there are still big potentials to improve.
> [C++][FlightRPC] Evaluate UCX/RDMA transport performance
> --------------------------------------------------------
>
> Key: ARROW-15229
> URL: https://issues.apache.org/jira/browse/ARROW-15229
> Project: Apache Arrow
> Issue Type: Task
> Components: C++, FlightRPC
> Reporter: Yibo Cai
> Assignee: Yibo Cai
> Priority: Major
>
> Evaluate FlgihtRPC UCX transport performance over 100Gb/RDMA network, based
> on https://github.com/lidavidm/arrow/tree/flight-ucx.
> cc [~lidavidm], will be great if you can give a stable branch (or commit id)
> for evaluation purpose.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
