[Beowulf] 2.6.11 is out; with InfiBand support
kus at free.net
Sat Mar 5 04:59:19 PST 2005
In message from Mark Hahn <hahn at physics.mcmaster.ca> (Wed, 2 Mar 2005
18:09:09 -0500 (EST)):
>> Arima and Iwill have mobos with IB LOM (Landed on Motherboard).
>given the choice between a $150 pcie IB nic and having it onboard,
>I'd choose the separate card. I know the IB salesdroids always
>say that getting onto the MB will change everything, but this
>doesn't make sense. IB is completely different from onboard gigabit,
>for instance, because there is no ubiquitous IB infrastructure
>ready, waiting to be exploited.
>the problem with "if you build it onboard, they will come" is also
>the marginal cost. onboard gigabit is nearly the same cost as
>onboard 100bT, very low, and you pretty much always want it.
>onboard IB is noticably higher than onboard GBE, noticable in
>absolute terms, and you definitely have no possible use for it
>on many systems.
>remember, most people don't even saturate GBE yet,
Yes, I agree. But we are developing some quantum-chemical application
which speedup at parallelization is bandwith-limited. And we obtain
that speedup on 6 processors w/IB 4x interconnect is about 34% percent
higher than for Myrinet.
Zelinsky Institute of Organic Chemistry
> and GBE
>ports are damned cheap. GBE nics are free, and switch ports
>are now down to $US 23/port:
>fundamentally, IB is still facing most of the same problems it always
>- requires fairly expensive, unique infrastructure
>- not the greatest physical layer: it's easy to wind up with
> literally tons of IB cables.
>- not clearly superior in performance vs alternatives.
>- apparently designed by people who disliked existing technique
> or were ignorant of it.
>- not a drop-in replacement for alternatives.
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit
More information about the Beowulf