[ofa-general] [PATCH] opensm: Parallelize (Stripe) LFT sets across switches

Hal Rosenstock hal.rosenstock at gmail.com
Thu Jul 23 04:31:09 PDT 2009


On Wed, Jul 22, 2009 at 8:51 PM, Jason
Gunthorpe<jgunthorpe at obsidianresearch.com> wrote:
> On Wed, Jul 22, 2009 at 08:28:25PM -0400, Hal Rosenstock wrote:
>
>> > But you overload the switch the SM is connected to with processing
>> > N*limit DR SMPs rather than just 'limit' SMPs. That is what concerns
>> > me.
>>
>> As I said, the current algorithm is worse as it sends N*no limit DR
>> SMPs (where no limit means any needed blocks). Not sure that VL15
>> droppage due to this has been identified. So I think this improves on
>> what's been deployed and seemingly works in OpenSM for quite some time
>> now.
>
> Hmm, OK I didn't realize that.
>
> I've heard of reports of VL15 droppage in real networks, maybe this is why..
>
>> > I first implemented an algorithm like this for switches based on Gamla
>> > chips, and then for Anafa. If something doesn't support it, it is
>> > very uncommon.
>>
>> I'm aware of at least two very different switches where this is the case.
>
> Well, that's horrible - but again, I personally have a hard time
> caring if using LID routing gives even a 5% reduction in setup time
> with compliant devices.
>
> I suppose if you really cared it would be asy to black list certain
> devices.
>
>> Understood; that's what I meant when I wrote below that it's harder
>> and more expensive computationally. I think that it's also overly
>> pessimistic so the number might want to be made artificially higher
>> based on experience that these SMPs can be pipelined quite a bit more
>> than this would allow.
>
> That is highly device dependent - some devices have more CPU SMP
> buffering than others and this affects things greatly.

The reason I said that this was overly pessismistic is that the count
at the parent switch deals on a transaction basis rather than in
flight as it passes that switch. So to compensate the default(s) would
need careful selection (and be made artificially higher to gain back
performance lost by cranking this down too tightly.

-- Hal

>
> Jason
>



More information about the general mailing list