[openfabrics-ewg] RE: OFED 1.0 rc5 SDP regressions

[Matters, Todd]

I agree.


-----Original Message-----
From: openfabrics-ewg-bounces at openib.org
[mailto:openfabrics-ewg-bounces at openib.org] On Behalf Of Tziporet Koren
Sent: Wednesday, June 07, 2006 10:32 AM
To: Scott Weitzenkamp (sweitzen); openfabrics-ewg at openib.org
Subject: [openfabrics-ewg] RE: OFED 1.0 rc5 SDP regressions

I don't think we should revert to old SDP since it had basic problems,
and was written in a way that will never bring SDP to the kernel.
When we did our internal IBG2 release we tested SDP based on the old
code and found that it is in bad shape (machines were stuck all the
time). This is the reason we went for the new implementation (believe me
we don't tend to rewrite code).

Major performance enhancements (e.g. Nagle algorithm for small messages)
are not implemented yet and this is the reason for loosing small-message
latency.

We will focus on performance in the coming weeks and will release a
better SDP in few weeks. 
This is the reason SDP is only in technology preview state.

Tziporet

-----Original Message-----
From: Scott Weitzenkamp (sweitzen) [mailto:sweitzen at cisco.com] 
Sent: Wednesday, June 07, 2006 11:12 AM
To: Tziporet Koren; openfabrics-ewg at openib.org
Subject: OFED 1.0 rc5 SDP regressions

> SDP: 
> The new SDP implementation is mainly focused on stability and
> correctness and not performance. We will do a simple analysis 
> and if we
> will find things we can easily fix we will do it. However major
> performance improvements will be done in the next release.

I don't remember any warnings ahead of time that the new SDP would not
have the same performance as the old SDP, so I'm still surprised by the
situation.  I'm having a hard time with losing >1 Gbps throughput and
small-message latency now worse than IPoIB.  Maybe we should revert to
the old SDP for OFED 1.0?

I just opened a scalability bug, #109
(http://openib.org/bugzilla/show_bug.cgi?id=109), "OFED 1.0 rc5: SDP
can't sustain 800 concurrent SDP connections".

Scott



_______________________________________________
openfabrics-ewg mailing list
openfabrics-ewg at openib.org
http://openib.org/mailman/listinfo/openfabrics-ewg