[ofw] [RFC] Generate IBAT path records in IPoIB

Tzachi Dar tzachid at mellanox.co.il
Wed Aug 6 14:04:14 PDT 2008


Hi Fab,

Please note that one of our goal is to have ipoib reach bw that is
higher than 10 GbE.

Xalex has already started working in this direction (and there is still
a lot to do).

I understand that you have a complicated task to solve in a short time,
but please try to allow the exiting mechanism to be used as they are
today. (for example by using a registry key).

Thanks
Tzachi

> -----Original Message-----
> From: ofw-bounces at lists.openfabrics.org 
> [mailto:ofw-bounces at lists.openfabrics.org] On Behalf Of Fab Tillier
> Sent: Wednesday, August 06, 2008 7:23 PM
> To: Alex Estrin; ofw at lists.openfabrics.org
> Subject: RE: [ofw] [RFC] Generate IBAT path records in IPoIB
> 
> Hi Alex,
> 
> > I think this should work fine along with your previous 
> patch. Couple 
> > questions though. Why broadcast broup record values could 
> be different 
> > from path query record?
> 
> I don't think there's anything in the IB spec that requires 
> them to be the same.  The packet rate can for sure be 
> different if you have mixed rate ports in the broadcast 
> group.  The broadcast group could be comprised of SDR, DDR, 
> and QDR HCAs in which case it would have a SDR packet rate.  
> This wouldn't give optimal performance between two DDR or QDR 
> HCAs. Same goes for the MTU, if you have a mix of 2K and 4K 
> MTU HCAs in the fabric.  The packet lifetime might be too big 
> for a RC connection too, but won't be too small.
> 
> I think for local subnet operations the traffic class, flow 
> label, and hop limit are likely ignored, so probably not a 
> big deal since there are likely bigger issues related to 
> routing than this (like needing the LID pair on both the 
> source and destination subnet).  The service level might also 
> need to be different for MPI traffic vs. IPoIB traffic, but 
> this needs to be handled locally to the node anyway.
> 
> > If so, maybe path query could be left as an option for Customer?
> 
> Callers can still call the existing MAC to GID IOCTL in IPoIB 
> and then do a path query to the SA.  Alternatively, clients 
> could extract the GID pairs from the returned path and 
> perform an SA query.  I either case, clients have the freedom 
> to use the returned path or do an SA query. The difference 
> between the MAC to GID IOCTL and the MAC to PATH IOCTL is the 
> output buffer size, 32 vs. 64  bytes respectively.
> 
> > Pkey is set to default - why not to get it from broadcast record as 
> > well?
> 
> Good point, thanks!
> 
> -Fab
> _______________________________________________
> ofw mailing list
> ofw at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
> 



More information about the ofw mailing list