benchmarks for pinning to shepherds

[Dragos]

Hello all,

Are there any Qthreads benchmarks that use task pinning on qthread shepherds? I did not find any and my attempts at getting the feature to show performance improvement were also futile so far.

Thank you for your help,

[Stark, Dylan]

Hi Dragos,

I don't know of any that explicitly place tasks. You might dig into the loop constructs to see what level of "pinning" is used when distributing the workload. My guess is not much.

Dylan

--
You received this message because you are subscribed to the Google Groups "qthreads" group.
To unsubscribe from this group and stop receiving emails from it, send an email to qthreads+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Kyle Wheeler]

Hi Dragos,

Shepherd pinning in Qthreads is used relatively frequently, just not in an obvious way. In particular, all qt_loop constructs use shepherd pinning to assign consistent loop iterations to specific shepherds, and as part of task spawning logic (so that tasks can be spawned in parallel rather than launched serially). Doing this with shepherd pinning was so consistently better and faster (in particular, keeping the loop iterations assigned to each shepherd consistent) that there is no longer a way to do the same thing without shepherd pinning. The only thing that comes close is to do a self-scheduled loop (qt_loop_queue), which is intended for loops with unpredictable latencies per iteration. In practice, the former is usually faster than the latter, for lots of reasons (e.g. the latter requires synchronization to orchestrate the execution of ranges).

There are, however, no benchmarks in the Qthreads repository which attempt to highlight this, which is, I think, what Dylan was getting at.

If you're curious about studying the impact of task pinning on benchmark runtime, I recommend reading Stephen Olivier's paper from a few years ago. It won Best Student Paper at the Supercomputing conference: http://dl.acm.org/citation.cfm?doid=1988796.1988804 That paper focuses on OpenMP and some extensions to it that he designed for specifying task locality, but those extensions were implemented with Qthreads and his results are all from code that he ran on Qthreads. His results are really well explained, and the paper includes one of the coolest graphs of task behavior I've ever seen (it's one of those "wow, why didn't I think of doing that?" papers).

Hope that helps,

~Kyle

--

You received this message because you are subscribed to the Google Groups "qthreads" group.
To unsubscribe from this group and stop receiving emails from it, send an email to qthreads+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
Ahh! Arrogance and stupidity in the same package, how efficient of you!
--Londo Molari

[Kyle Wheeler]

Sorry, I got the paper citation incorrect. It's here: http://dl.acm.org/citation.cfm?id=2389085