|
|
Subscribe / Log in / New account

Moving past TCP in the data center, part 1

Moving past TCP in the data center, part 1

Posted Nov 2, 2022 4:01 UTC (Wed) by neilbrown (subscriber, #359)
Parent article: Moving past TCP in the data center, part 1

I thought that RDMA was the "new" way to get better performance in the data center.
It is already implemented in (high end) hardware, and there are already RPC implementations layered over it.

I wonder what it is missing? Low-latency for tiny message maybe?


to post comments

Moving past TCP in the data center, part 1

Posted Nov 2, 2022 9:39 UTC (Wed) by joib (subscriber, #8541) [Link]

From https://web.stanford.edu/~ouster/cgi-bin/papers/replaceTc... (which I guess is the paper this presentation is based on), section 6 deals with Infiniband.

> However, RDMA shares most of TCP’s problems. It is
> based on streams and connections (RDMA also offers unre-
> liable datagrams, but these have problems similar to those for
> UDP). It requires in-order packet delivery. Its congestion con-
> trol mechanism, based on priority flow control (PFC), is dif-
> ferent from TCP’s, but it is also problematic. And, it does not
> implement an SRPT priority mechanism.


Copyright © 2025, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds