[openib-general] [PATCH] expand local SA cache to include all path records from source

[Hal Rosenstock]

On Fri, 2006-04-21 at 12:30, Sean Hefty wrote:
> Hal Rosenstock wrote:
> > As the end node cannot know the policies the SM used for routing, etc.,
> > path records are insufficient in terms of some of the interesting
> > characteristics for filtering like path independence (as fault
> > independent as possible). 
> 
> This patch doesn't handle policy.

I'm not following you on what policy you are referring to here.

>   It only stores multiple path records.

Yes, but the path records returned from MPR queries can have "better"
properties than PR queries for things like APM, etc.

> Also, it seems like the argument can go both ways here.  The SM doesn't know the 
> policies of the applications or their traffic patterns.  A distributed 
> application may want paths from nodes A to B to have path independence from 
> paths from nodes C to D.  Can the SM handle such requests?

I think that is usually addressed in the subnet configuration (e.g.
richness of the topology) when the subnet is architected (and use of
LMC). Another possibility would be an SM feature to be configured in
some manner which may not be realizable based on the topology.

> > There's also a 1.2 erratum which enhances SA MultiPathRecord to support
> > scoping of S/DGIDs. It supports explicit, same node, same system with
> > and without high availability (separate HCAs). That could also be used
> > when supported by the SA.
> 
> It shouldn't be hard to change the path query from GetTable to GetMulti, once 
> the MAD layer supports dual-RMPP.
> 
> > The API looks fine to me. One thing to note is to keep an eye on the
> > lookups done on the local database and possibly optimizing them if
> > needed.
> 
> Tracking path usage becomes difficult, and would likely require involving the 
> CM, and possibly address handle creation (for UD QPs that didn't use SIDR). 
> When the cache is updated, checks would need to be added to see if a path is 
> currently in the cache, and invalid paths would need to be removed.  There would 
> be a lot of additional work needed to do this.

I was referring to the types of lookups used (e.g. key variety).

-- Hal

> - Sean