[openib-general] Performance Degradation with OFED v. Voltaire(lustre)

[Michael S. Tsirkin]

Correct. But endnode does not know what's on the other side of the link,
and which path MTU is best.

This is SA's job (SA sees all the topology), and you will get exactly this behaviour
you ask for if you run opensm with "quirk mode" enabled (you must disable the *other*
SM you are running though for this to take effect).

This mode will be enabled by default in OFED 1.2.

Quoting r. Bernadat, Philippe <philippe_bernadat at hp.com>:
Subject: RE: [openib-general] Performance Degradation with OFED v. Voltaire(lustre)

 
Koren & Or,

I am building and testing as we speak.
But my feeling is that this issue shouldn't require user to set the
tavor_quirk param.

The stack should detect this HCA flavor at the appropriate end (active
according to Or) and should automatically adjust the MTU.

Philippe


> -----Original Message-----
> From: Or Gerlitz [mailto:ogerlitz at voltaire.com] 
> Sent: Tuesday, December 19, 2006 3:47 PM
> To: Tziporet Koren
> Cc: Michael S. Tsirkin; Bernadat, Philippe; Roland Dreier; 
> openib-general at openib.org
> Subject: Re: [openib-general] Performance Degradation with 
> OFED v. Voltaire(lustre)
> 
> Tziporet Koren wrote:
> > Hi Or,
> > Can you update OFED support page on Wiki with this issue?
> 
> Basically, yes but actually, not...
> 
> We (Michael and myself) do not agree yet on some issues here, 
> also the 
> cma tavor quirk will not work with some 3rd party SM/SA, so 
> for the time 
> being i will also put there a note on how to do it in the ULP 
> level (eg 
> as Philippe was fixing Lustre)
> 
> Or.
> 
> 

-- 
MST