[AMBER] Parallel pmemd.cuda: local bandwidth question. Are dual host PCI-E cards make any sense?

From: Sasha Buzko <obuzko.ucla.edu>
Date: Wed, 04 Aug 2010 20:57:37 -0700

Ross, Scott,
thank you for the feedback. Clearly, any future mpi version should go
over QDR IB.

Can you help with the issue of local bandwidth? Nvidia sells dual host
PCI-E adapter card that effectively connects 4 GPUs in an S2050 to a
single x16 slot. When pmemd.cuda is run locally (parallel or four serial
processes), how much would this impact performance?
In other words, should we even consider these cards for a host system
intended to run pmemd.cuda?

Thanks again

Sasha


Thomas Zeiser wrote:
> On Sat, Jul 31, 2010 at 10:12:29AM -0700, Scott Le Grand wrote:
>
>> I'd definitely go for QDR between nodes.
>>
>> What's up in the air ATM is whether it's best to spread the
>> C2050s across as many nodes as possible or whether 2 or 4 C2050s
>> per node is the optimum configuration.
>>
>
> for QDR to work at full speed you need PCIe2.0 8x (not only
> mechanically but also electrically). Are there any boards which have
> four 16x slots and additionally more than only 4x electrically?
>
> At least in the typical Intel-based boards you have two 16x and one
> 4x slot when there is one chipset on the mainboard; or four 16x and
> (several) 4x slots (4x electrically although their mechanics width
> is 8x) if there are two chipsets.
>
> Thus, with these boards, the maximum feasable is full DDR speed
> using a DDR-IB card supporting 5.0 GT/s; (slightly cheaper DDR
> cards with 2.5 GT/s only operate with an effective speed similar to
> SDR ...)
>
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Aug 04 2010 - 21:00:05 PDT
Custom Search