qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [WIP] RDMA transport for COLO


From: Hailiang Zhang
Subject: Re: [Qemu-devel] [WIP] RDMA transport for COLO
Date: Fri, 25 Dec 2015 15:06:21 +0800
User-agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:38.0) Gecko/20100101 Thunderbird/38.4.0

On 2015/12/18 4:31, Dr. David Alan Gilbert wrote:
Hi,
   I've been playing with getting an RDMA setup for COLO and
have something that mostly works, but it is very new and quite
hacky; but I thought I'd share my work so far.


Nice work, i will look at it later. :)

You can find it at:
https://github.com/orbitfp7/qemu/commits/orbit-wp4-colo-dec

What I've done is:
   a) Wire up a partner TCP connection by the side of the RDMA
     connection.
   b) Use the TCP connection just for the responses from secondary->primary
   c) Make the RDMA connection write to the colo-cache after
      the first migrate
   d) Make the RDMA connection notify the secondary when it
     sends writes, so that the secondary can know that it needs
     to flush those pages in the colo-cache.
   e) Add a shutdown function and fix some other bugs

I've had that working on both your current world (which is
what that tree is based off) and your older COLO world
from July (with a bit more hacking to make it take the newer
patches).

Looking at the speed:
   a) The CPU load on the incoming thread is much lower - maybe
only 10-11% instead of 30-40%.
   b) The performance of guest code is a little slower (~10% slower?)
     on RDMA rather than TCP (on both 10Gbps and 40Gbps links)
     I've not worked out why yet. (My guess is it could be to do with
     RDMA dynamic registration)

Things I know I need to do:
   1) Tidy it up - it's very messy!
   2) Try and get rid of the TCP connection and use an RDMA
      channel for the backwards connection
   3) Make sure the shutdown really can cope with the other host
      being dead.
   4) It only deals with the dynamic registration mode of RDMA;
      setting pin-all will probably break it.
   5) Figure out why it's slower!
   6) Test failover more.

My work on this is part of the EU Orbit project
( http://www.orbitproject.eu/ )

Dave


--
Dr. David Alan Gilbert / address@hidden / Manchester, UK

.






reply via email to

[Prev in Thread] Current Thread [Next in Thread]