[ofa-general] [RFC] host stack IB-to-IB router support

Michael Krause krause at cup.hp.com
Wed Mar 21 12:09:24 PDT 2007


At 10:51 AM 3/21/2007, Sean Hefty wrote:
>>Ok, lets assume Sean would finish his experiments with remote_sa, how
>>would that find its way into the commercial sm/sa versions that are
>>mostly used, how would we guarantee interoperability between all
>>implementations, .. ?
>>How would that address future routing, security, QoS, .. enhancements ?
>>can it ?
>
>The 'remote sa' as simply a proprietary UD protocol.  Whatever data two 
>'remote sa' services exchange shouldn't matter, nor should the fact that 
>each issues local SA path records.  There's nothing magical about this.
>
>If I have an app that can query its local SA, there's nothing that 
>prevents that app from sending that data to whatever peer it can connect 
>to.  It can even send the data over TCP if it wants.  Keeping the SA 
>subnet local doesn't add any real security.
>
>Coming up with a solution that doesn't work with any existing hardware, 
>targets, and SAs isn't very useful.

Just to clarify:

- Nothing in the router protocol should have an impact on existing or even 
future hardware if done right.  The basic wire protocol, i.e. the use of 
GRH, etc. should not require any modifications to operate on existing hardware.

- Whether a HCA or a TCA, there will be some level of management protocol 
changes.   This impacts the software above but not the hardware itself 
unless the implementation hard-coded / state machined its behavior in which 
case it is unlikely to work in any router environment.

- For the SA, I think most will agree that there will be implementation 
changes required to comprehend where a router exists on a subnet and how to 
respond to queries that target a router.   However, much of what I've noted 
here does not impact existing SA or SM operations - they continue to work 
as implemented.   The changes proposed would be new additional capabilities 
that would enable router communication to occur.   If people construct 
something like a DNS equivalent service to find the IB router, then this 
leverages the practices used with IP today and the more IB looks like IP 
when it comes to its operation, the easier it is to get it actually 
deployed beyond the HPC market.

None of these items breaks or changes interoperability among any components.

Again, I don't see any harm in waiting until Sonoma to discuss this 
face-to-face.   Quite true that no major breakthroughs are likely but the 
benefits of plowing ahead with an implementation that may be viewed as an 
academic or a niche experiment does not seem worthwhile.   However, people 
are free to spend their time as they desire.   My only caution is such work 
should not set precedence nor should there be any expectation that it will 
ever see commercial adoption or deployment.

Mike 





More information about the general mailing list