[Swift-devel] Falkon and Coaster support for MPI

[Michael Wilde]

On 6/29/08 1:28 PM, Ben Clifford wrote:
> On Sun, 29 Jun 2008, Michael Wilde wrote:
> 
>> The initial discussion here was based on the assumption that a Falkon-like
>> mechanism was required in order to run workloads of many small jobs on the BGP
>> - whether that be through Swift, or directly. (Small meaning 1 to 64 CPUs each
>> and order of a few minutes of runtime each).
> 
> Is this the sort of workload that the applications that are targetted for 
> MPI use on this machine create? (what are those applications, btw?)

The app that accelerated this discussion is CHARMM for molecular 
dynamics. CHARMM on BGP with the parameters needed in the use case in 
question has a long runtime on 1 CPU (24-36 hours), and seems to peak in 
performance using MPI at 32 CPUS. Runs >6 hours are also not runnable 
today on the Argonne BGP without a reservation. So the most effective 
way to use the BGP for some specific CHARMM runs needed by Benoit Roux's 
group is to run large numbers of multi-hour 32-rank MPI jobs.

(In the meantime we're looking to run 1-CPU jobs broken into multiple 
separate time steps and then merged. Extra work, some questions on 
accuracy and equivalence, but seemingly doable).

> 
> 
>> Recently, IBM circulated info on their "HTC" mode support for the BG/P, which
>> may change the nature of the assumptions above.
> 
> That would be useful to see. Its hasn't circulated to me, though.

http://www.bgconsortium.org/documents/HTC%20WhitePaper%20V2%20050508.pdf

>