[mpich-discuss] Integrating parallel and serial code

[Reuti]

Hi,

Am 12.07.2010 um 23:52 schrieb Michael Morrison:

> Hi All, I hope i'm posting this in the correct forum. I'm extremely  
> new to MPI so forgive me if my question comes off as silly. I've  
> only been dabbling around with parallel computing for a few weeks  
> now but I'm wondering about the possibilities. So far with each  
> example that I've been through, I create an MPI application and then  
> run this application using the mpiexec command. What I'm wondering  
> about is if it would be possible to say have a C function that  
> executes some algorithm in a parallel manner that could be called by  
> another C function. For example suppose i had two C functions, the  
> first called function A and the second called function B. Function A  
> is just a normal function that runs serially, function B executes  
> its algorithm parallelly using MPI constucts. Is it possible to have  
> a set up like this?
>
> From the examples I've seen it appears that the MPI code runs as  
> standalone units and couldn't be integrated with  serially executing  
> code. Is this true or is there someway to make this set up work? I  
> may not have given enough information to answer the question, if  
> there's any confusion please ask and i'll clarify.

do you want to run this code just local on one machine or in a cluster  
with an installed job scheduler? I'm not aware of any scheduler, which  
allows you to specify varying resource requests for the lifetime of a  
job. E.g. I need:

- 1 hr - 1 core
- 4 hrs - 2 cores
- 1 hr - 4 cores

If you submit such a job with the highest amount i.e. 4 cores, you  
will have unused resources. You could split the job to three steps,  
which just wait for the end of the predecessor, but then you are left  
again with pure serial and pure parallel steps.

If you are just in the beginning of a project, it would be best to  
design the application in such a way, that all requested resources are  
used to their maximum and the offcut is minimized.

-- Reuti


> Thanks for your time,
>
> Mike
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss