On Thu, 2007-03-08 at 08:45, Sasha Khapyorsky wrote: > Mostly it implements the "switch only" optimization idea and affects > the min hops matrices building phase (for both up/down and default > builders). > > The main trick is to keep the min hop tables _ONLY_ for switches base > LIDs and don't bother with CAs, routers LIDs and any secondary LIDs in > case when LMC > 0 - this saves a lot of memory and cpu cycles needed > for calculation and storing the huge whole fabric matrices. > > For CA and router ports we will refer its neighbor switch's min hop > vectors. And for LMC > 0 case we will use base LID's min hop vectors > for any secondary LIDs (for CAs and routers it will neighbor switch's > base LID). > > Preliminary testing shows 3-4x speedup in the min-hop generation phase > and yet another 2x for up/down.
More nice work! Thanks, Sasha. -- Hal > Sasha _______________________________________________ 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
