[Swift-devel] Re: Delay when coasters are shutting down

[Mihael Hategan]

On Fri, 2010-09-03 at 10:45 -0600, wilde at mcs.anl.gov wrote:
> Mihael,
> 
> There seems to be a brief (~ 2-5 second) delay at the end of a run
> when swift seems to be shutting down the coaster service. (See example
> below)
> 
> This delay degrades performance for the Swift R interface, where an R
> script will be running many separate Swift scripts, one for each call
> to parallel apply().
> 
> Can this delay be readily eliminated? If its not present for
> persistent coasters, then that's not a problem. I'll look into it when
> I get a chance, but a pointer to where in the code the delay occurs
> would be helpful.

The delay is not there with persistent coasters. It is also not there
if, by some unknown mechanism, the sub-swift JVM is reused.

> 
> Eventually we'll have an R script make multiple calls to the same
> Swift JVM for successive apply() calls. That in fact would be a better
> place to focus than eliminating this delay.
> 
> In the R interface, we use a pair of FIFOs to communicate pass calls
> from a Swift app() script to persistent R server processes, one per
> Swift worker slot. We could use the same approach (and/or sockets) to
> provide a client interface for running multiple Swift calls from the
> same JVM.

Can't the sub-swift be a swift function?

Mihael