[openib-general] IB Address Translation service

[shaharf]

> 
> The advantage of ATS is that it "just works" whether wired point to
> point, or via a switch, or whatever. It requires no central
> administration,
> works as transparently as ARP and ND, and supports IP addressing so
> applications don't have any ambiguity in how they resolve names.
> 
> If we get rid of ATS, what do we replace it with? Raw IB GID's from
> the application??
> 
> Tom.
> 
Tom, I am with you regarding that subject. Even though both IB-ARP and
ATS should be considered to be a hack, I think that IB-ARP is much more
problematic and it does not deliver a complete solution: it doesn't
contain all data required, it does not solve the SM load problem (due
the requirement for the path record query) and it is a mechanism that
contradicts IB management architecture. I would fix the ATS to include
some missing fields, and maybe define unified ATS + path query for
performance.

The SM/SA scalability problem should be solved by distributing the SA
part of it, probably using a single write/multiple reader model and a
simple cache coherency protocol to allow efficient caching by sub SA
agents or even hosts. This type of distribution is also requested
clearly in the SOW section 1.3.1.

Shahar