[ofa-general] Re: OpenSM and fat tree

[Yevgeny Kliteynik]

Hi Hal,

Hal Rosenstock wrote:
> On Sat, 2008-05-10 at 21:51 +0300, Yevgeny Kliteynik wrote:
>> Hal Rosenstock wrote:
>>> Hi Yevgeny,
>>>
>>> Is it possible that OpenSM's fat tree routing somehow depends on the
>>> LIDs previously assigned ?
>> It depends only on the existence of the LIDs.
>>
>>> It seems that for a legitimate fat tree topology, the topology sometimes
>>> won't come up as a fat tree if reassigning LIDs (-r) is not used.
>> That's odd...
>>
>>> In addition to -r making fat tree work, certain routing algorithms seem
>>> to also clear this out (without using -r). For example, if lash were run
>>> and then ftree, it seems to work without doing the -r. (Haven't yet
>>> tried updn).
>>>
>>> Any ideas on this ? Should a bug be filed on this ? Thanks.
>> No ideas whatsoever. Please file a bug on this.
> 
> I filed this as bug 1031:
> https://bugs.openfabrics.org/show_bug.cgi?id=1031

Thanks

>> It would be nice if I could reproduce it in simulation.
> 
> Yes, that would be nice; but I don't have a sim case.

The problem is, I don't even know where to start.
I tested it in simulations on different topologies,
and it is used on real cluster(s) too.

I need more details, and some hint on how to reproduce it.
Can you describe the setup you used when you saw this problem?

-- Yevgeny


> -- Hal
> 
>> -- Yevgeny
>>
>>> -- Hal
>>>
>>>
>>>
> 
>