[Swift-user] Swift and allocations

Mihael Hategan hategan at mcs.anl.gov
Tue Aug 12 12:37:18 CDT 2014


Hi Jonathan,

There is no specific code to attempt to bunch the invocations to fewer
blocks. Essentially the 382 cores that finish their previous 512 jobs
first will get the new jobs.

However, if any of the nodes are left without jobs, they will get shut
down.

Mihael

On Tue, 2014-08-12 at 11:53 -0500, Jonathan Ozik wrote:
> Hello all,
> 
> I had a discussion with Yadu and we thought we’d open it up to the
> list. Suppose that we are throttling the total number of concurrent
> tasks to 512 (taskThrottle=512) and by specifying that maxJobs=32 and
> tasksPerWorker=16. If there are a total of 2430 total tasks to run,
> there will be 2430 - 2048 = 382 tasks left for the last “go
> around” (assuming that each of the tasks take about the same time to
> complete). If there are 382 tasks left towards run completion and 512
> cores (16 for each of the 32 nodes), would the last few jobs, end up
> spread across the 32 Nodes, effectively holding them and burning your
> allocation, or would they get pushed to 24 nodes and de-allocating 8
> nodes?
> 
> Jonathan
> _______________________________________________
> Swift-user mailing list
> Swift-user at ci.uchicago.edu
> https://lists.ci.uchicago.edu/cgi-bin/mailman/listinfo/swift-user





More information about the Swift-user mailing list