[ofa-general] [RFC] Fat-Tree upgrades

[Nicolas Morey Chaisemartin]

Yevgeny Kliteynik wrote:

> 
> Not exactly. Even on your way up (while creating downward routes)
> you can find shorter paths than what was configured by main path.
> It's hard to explain in words, so I draw some ppt slides (see the
> attachement).
> 

You're right. In fact my patches already deal with these cases. I just didn't remember it :)

> 
>> To test these algorithms we have written some scripts.
>> We basically calculate the number of routes (CN to CN) per link on the
>> whole network.
>> The better the algorithm is, the less point you should have (on
>> regular topology, theoretical points is about one per level). We
>> mesure this for each switch/link loss using ibsim/opensm and aggregate
>> all the results. Using the minhop and sorted secondary routes, the
>> behaviour of the algorithm is totally changed, in well :)
>>
>> I'm also working on some more enhancements and speed upgrades (sorting
>> costs a bit of time but I've managed to reduce the cost a lot). I'll
>> present them when they are ready.
> 
> The speed penalty was my other concern.
> Will be happy to see the comparison of the calculation time with and w/o
> the sorting once its ready.

I'll run some bench and post the results as soon s possible


Nicolas