[openfabrics-ewg] OFED procedures

[Scott Weitzenkamp (sweitzen)]

PCI-X max throughput is still about 30% below what is was before.  More
details are in http://openib.org/bugzilla/show_bug.cgi?id=108.

I have enclosed a PCI-E example of bad throughput performance, this is
on a Dell PE 1950 (2 CPUs, each with dual core) with Lion Cub SDR HCA.
The graphs are for running netperf with TCP_STREAM and TCP_RR.  IPoIB
does very well on this platform, but look at page 5 for SDP.  The
throughput peaks at 7 Gb/sec but then falls back to 5 Gb/sec.  And I
agree the small message throughput is not good, for message size 1K
TCP_STREAM shows approx 1 Gb/sec, this was 4 Gb/sec before the rewrite.

But we do see 10.5 usec SDP latency on this 1950, thank you for that :-)

Scott Weitzenkamp
SQA and Release Manager
Server Virtualization Business Unit
Cisco Systems
 

> -----Original Message-----
> From: Aviram Gutman [mailto:aviram at dev.mellanox.co.il] 
> Sent: Wednesday, October 04, 2006 12:45 AM
> To: Scott Weitzenkamp (sweitzen)
> Cc: Michael S. Tsirkin; Open Fabrics
> Subject: Re: [openfabrics-ewg] OFED procedures
> 
> Scott Weitzenkamp (sweitzen) wrote:
> >> I agree. We were less good filing bugs. We tried to make sure 
> >> that the 
> >> RCs are getting out in good shape. That is the reason you 
> see a good 
> >> improvement in SDP stability.
> >>     
> >
> > I agree OFED 1.1 SDP stability is back to where it was before the
> > rewrite in OFED 1.0 rc5.  SDP max throughput is still not as good,
> > though.
> >
> > Scott
> >   
> Yes, we worked a lot on stability. But we also worked on performance. 
> Please pay attention that we cut down he SDP latency to ~13uSec.  To 
> best of my knowledge max throughput is also better, ~950MB/s. Small 
> messages BW is better then OFED 1.0 but still is not as good as we 
> wanted it to be.
> 
> Aviram
> 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pb7.1950.sdr.pdf
Type: application/octet-stream
Size: 68186 bytes
Desc: pb7.1950.sdr.pdf
URL: <http://lists.openfabrics.org/pipermail/ewg/attachments/20061004/e3adb8b7/attachment.obj>