[Swift-user] A question about design choice

[Mihael Hategan]

We wrote swift to automatically sort out complex parallelism between
bunches of heavy work. You seem to be doing that in Matlab. Assuming
that it's not trivial or advantageous to switch to doing those in swift,
it looks like you need a mechanism to efficiently deal with those
bunches of heavy work.

I think the quickest path for you would be to use a persistent coaster
service. You would be able to call swift as you do now, but different
swift runs would re-use the same pbs jobs to run your apps.

If you think that swift is unnecessary (and a parallel for loop is
something for which I'd say swift is only marginally useful), there
might be some compromise solutions, but they may require a bit more work
for benefits that I cannot quantify well, but I suspect to not be
significant.

So I think that if you want to explore the latter, you should probably
think in terms of specific and tangible downsides with the former. So
please list those, and we can start from there if needed.

Mihael

On Mon, 2012-10-08 at 15:45 -0500, Jonathan Margoliash wrote:
> In my previous email I forgot to mention that I'm using a
> doubly-nested for loop, but I doubt that matters.
> 
> 
> Currently we are using a single cluster. I'm not sure if there are
> plans to expand to using multiple clusters at once.
> 
> On Mon, Oct 8, 2012 at 3:40 PM, Mihael Hategan <hategan at mcs.anl.gov>
> wrote:
>         On Mon, 2012-10-08 at 15:35 -0500, Jonathan Margoliash wrote:
>         > Along with a couple typedefs and an app definition, yes. Is
>         there a
>         > better way to do what I'm doing? Thanks,
>         
>         
>         Are you using a single cluster or are you planning to branch
>         out to
>         multiple clusters?
>         
>         
>