[ofa-general] RE: running opensm 3.0.3 on 4000+ node system

Hal Rosenstock hrosenstock at xsigo.com
Wed Apr 9 14:54:11 PDT 2008


On Thu, 2008-04-10 at 00:50 +0000, Sasha Khapyorsky wrote:
> On 14:17 Wed 09 Apr     , Hal Rosenstock wrote:
> > On Wed, 2008-04-09 at 15:13 -0600, Maestas, Christopher Daniel wrote:
> > > I think we may have fixed it:
> > > ---
> > >  3998 pts/0    Sl     1:47 /usr/sbin/opensm -maxsmps 15 -t 200 -f /var/log/osm.log -g 0
> > > --
> > > 
> > > I changed maxsmps to 15 (from default of 0 => unlimited) and it seems to be working now. 
> > >  That is the same value we use for the cisco host based sm.
> > 
> > Yes, an infinite value could overrun the unflow controlled VL15 buffers
> > in the switches.
> 
> Even if not - it overflows mad response matching table in vendor layer
> (there are 4k+ nodes and only 1k entries in the table). In recent
> version (master) this table size can be redefined with
> OSM_UMAD_MAX_PENDING environment variable.

Right; I forgot about that but not sure why that wouldn't have happened
on his earlier use of OpenSM though.

-- Hal

> 
> Sasha
> _______________________________________________
> general mailing list
> general at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general




More information about the general mailing list