On Wed, 2008-04-09 at 09:36 -0700, [EMAIL PROTECTED] wrote: > > On Wed, 2008-04-09 at 08:38 -0700, [EMAIL PROTECTED] wrote: > >> > On Wed, 2008-04-09 at 10:01 +0000, Sasha Khapyorsky wrote: > >> >> Hi Ira, > >> >> > >> >> On 16:48 Tue 08 Apr , [EMAIL PROTECTED] wrote: > >> >> > As per Hal's comments change the alternate value for [leaf] HOQ to > >> be > >> >> > "infinity" when the user specifies a value larger than "infinity". > >> >> > >> >> Actually I would prefer original version of the patch. The main > >> reason > >> >> is that infinite packet life time is really dangerous thing - in case > >> >> when a fabric is routed with credit loops (very common case with > >> default > >> >> min-hops routing) it leads to total fabric stuck and not just to some > >> >> performance degradation. > >> >> > >> >> So I think it is safer to reject invalid value and to set the default > >> >> (log an error, etc.i). As it was done in the original version of the > >> >> patch. > >> >> > >> >> Hal, do you agree? > >> > > >> > Safer yes but I think it is less to the intent of the admin who just > >> > doesn't understand the max value for this and that's why I proposed > >> this > >> > change. My preference is to max it out but it comes down to a judgment > >> > call. There's a downside either way. > >> > >> What if we set it to 0x13? This would be the maximum value that will > >> not > >> "lock" up the fabric. We could also add to the error message that the > >> admin needs to specify 0x14 if they specifically want "infinity" to be > >> set? > > > > So disallow the setting to infinity ? > > > > No, if you want infinity you have to specify 0x14 (19) in the opensm.opts > file. For example, specifying 100 will set the value to 0x13 and warn the > user that if they want infinity they will have to specify it explicitly; > ie head_of_queue_lifetime = 0x14
That's another choice but seems a little weird to me that 20 is infinite and >20 is set less than that but this is a judgment call. Not sure what others think. At this point, I have nothing more to add on this. -- Hal > Ira > > > _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
