[ofa-general] Further 2.6.23 merge plans...
Jason Gunthorpe
jgunthorpe at obsidianresearch.com
Wed Jul 18 12:27:45 PDT 2007
On Tue, Jul 17, 2007 at 10:39:11PM -0700, Roland Dreier wrote:
> > IMHO, I still think that without some kind of SM/SA sourced
> > invalidation mechanism all client side caching (including the ipoib
> > stuff we have now) is a bad idea.
>
> But for IPoIB at least doing a path lookup for every packet is
> obviously not feasible. And ARP table aging gives a way to recover
> from stale cached data, eventually at least.
Well, aside from Michael's points about the current implementation,
even a perfect version relying only on ARP will still have annoying
failure modes. ARP in ethernet has a built in means to revoke a bad
mac, and IB also will be able to revoke a bad GID - but since the path
information in incoming APR LRH's isn't used it doesn't fix changes in
the network caused by the SM. ARP entry aging helps, but IIRC there
are cases where aging can be slowed if the right packets are Rx'd.
Also, I think I ment 'bad idea' ==> 'has annoying and subtle failure
modes' - UD ipoib definately needs to cache LRH data with ARP entries..
Jason
More information about the general
mailing list