On 5/22/08, Hal Rosenstock <[EMAIL PROTECTED]> wrote: > > Olga, > > On Thu, 2008-05-22 at 18:28 +0300, Olga Shern (Voltaire) wrote: > > Hal, > > You pointed out that we cannot rely on the assumption that on SM > > failover there is not path change. > > In the previous patch we only flush multicast. > > What Moni changed in this patch is that on SM failover (SM change > > event), we will flush not only multicast but also all paths but > > without destroying ah. > > I missed that in the patch :-( It addresses the first level of concern > in terms of the unicast paths but leaves open the path parameter changes > (rate, etc.) as the address handles are preserved as Moni stated in > other words. I agree it's in the right direction. I would like to see > the whole problem solved. Is the cost of recreating the AHs too much or > is something else leading towards preserving the AHs ? That's what's > needed to be resolved for a complete solution.
I didn't explain it well, I will try again :) On SM change event we will not destroy ah but will flush paths, therefore unicast traffic will continue without packets lost. When there will be arp probe (issued by the kernel) it will look for a path and because we have deleted it will issue path query to SM and after reply from sm it will create a new ah that will replace the old ah. Before this patch all packets were dropped till there is a new ah, this patch creating new ah at the background. I hope it is clear now. Olga -- Hal > > > Olga > >
_______________________________________________ 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
