[openib-general] [PATCH 2/4] osm: LMC > 0 is not supported by fat-tree routing.

[Yevgeny Kliteynik]

Hal Rosenstock wrote:
> On Tue, 2007-01-02 at 09:58, Yevgeny Kliteynik wrote:
>> Hal Rosenstock wrote:
>>> On Mon, 2007-01-01 at 08:31, Yevgeny Kliteynik wrote:
>>>> LMC > 0 is not supported by fat-tree routing.
>>> Might this be different in the future ? Can LMC > 0 be supported with
>>> fat tree routing ?
>> Sure. We just have to figure out two things:
>> 1. Figure out what should communication pattern look like 
>> when LMC is not 0.
> 
>> 2. How to make fat-tree routing choose different paths for 
>> different lids of the same CA, because otherwise using LMC>0 
>> is pointless.
> 
> Not sure what you mean by this. How is this different for fat tree
> routing than any other routing algorithm ? Maybe this is an issue for
> all of them. 

I agree - this is an issue for any routing.
I guess it wouldn't take too much to add a naive support for LMC>0,
(in which case routes to LIDs of the same port will always be routed
through the same path).
But then we come back to the first question that we need to think about:
this routing algorithm is optimized for shift pattern. How will the 
communication pattern look with LMC>0?

--Yevgeny
 
> Doesn't the PR/MPR request preselect the LID or the response determines
> the LID to use ? The only issue I see is whether there needs to be a
> separate GID for each possible LID.
> 
> -- Hal
> 
>> --Yevgeny.
>>
>>>> Removing all the related code and adding check to
>>>> inform the user in case LMC is set.
>>>>
>>>> Signed-off-by: Yevgeny Kliteynik <kliteyn at dev.mellanox.co.il>
>>> Thanks. Applied.
>>>
>>> I'll also add a note to this effect to the documentation shortly.
>>>
>>> -- Hal
>>>
>