> Here's some anecdotal evidence :)
> http://lists.openfabrics.org/pipermail/general/2007-May/035758.html
Right, but then we went on to say that we probably want to use
multiple vectors to separate out multiple HCA ports rather than
send/sreceive on the same port. And the current IPoIB implementation
of having that second CQ seems suboptimal anyway, since it seems to
leave us susceptible to the interrupt overload that NAPI was supposed
to solve.
At a higher level, I'm left wondering why nobody talked about multiple
EQs during the last months of the 2.6.22 process and now all of a
sudden it becomes urgent in the last few days of the 2.6.23 merge
window. That's not really how I like to merge features....
- R.
Quoting Roland Dreier [off-list ref]:
Subject: Re: [PATCH 01/10] IB/ehca: Support for multiple event queues
> Here's some anecdotal evidence :)
> http://lists.openfabrics.org/pipermail/general/2007-May/035758.html
Right, but then we went on to say that we probably want to use
multiple vectors to separate out multiple HCA ports rather than
send/sreceive on the same port. And the current IPoIB implementation
of having that second CQ seems suboptimal anyway, since it seems to
leave us susceptible to the interrupt overload that NAPI was supposed
to solve.
Sure, the ipoib patch is just a proof of concept anyway.
And I'm actually working on merging send/recv CQs now,
to address the livelocks.
At a higher level, I'm left wondering why nobody talked about multiple
EQs during the last months of the 2.6.22 process and now all of a
sudden it becomes urgent in the last few days of the 2.6.23 merge
window.
I don't see any emergency in merging the IPoIB hack either.
I just hoped that once we merge the core changes people will start
experimenting with multiple vectors. This did not seem to have happened.
Could this be because there's no low level driver support upstream yet?
So I wonder whether merging the mthca patch [that was patch 2 of the series]
in 2.6.23 will finally get the
ball rolling, get people to experiment with multiple vectors
in userspace, and that will hopefully teach us something.
That's not really how I like to merge features....
If you look just at the mthca patch in isolation,
do you still see a problem?
--
MST
Roland Dreier [off-list ref] wrote on 17.07.2007 19:52:55:
At a higher level, I'm left wondering why nobody talked about multiple
EQs during the last months of the 2.6.22 process and now all of a
sudden it becomes urgent in the last few days of the 2.6.23 merge
window. That's not really how I like to merge features....
OK, let keep multiple eqs for next release with a more stable
verbs def.
For the other patch to support MR with large pages we'll resend
it (without deps on multiple eqs patch) to you soon.
Regards
Nam