[Openmcl-devel] CCL running on Knights Landing
brooks.martin at sympatico.ca
Mon Jan 30 12:30:40 PST 2017
A: One CCL image generating multiple threads via process-run-function.
This is application code, not an artificial test.
I use a simple scheme where a vector of computations-to-be-done is divided by the number of cores.
The individual computations do not typically run flat out because they share resources, although I have difficulty identifying what resources they might be; improving this is part of my development challenge.
Performance varies up and down, and GC may be one of the causes for reduced parallel performance.
I try to keep GC out of the way: (set-lisp-heap-gc-threshold (* 100 (lisp-heap-gc-threshold)))
I call GC manually between processing stages.
Note that I am running in 110GB of memory and have used only 27 GB.
I’m in uncharted territory, so there is certainly room for improvement.
I am totally impressed that CCL runs out of the box on Knights Landing — congratulations to Clozure Associates !
> On Jan 30, 2017, at 3:13 PM, Ron Garret <ron at flownet.com> wrote:
> Wow! Cool!
> Is this a single CCL process loading up 100+ cores using native thread, or is it 100+ separate CCL processes? If the former, are you seeing any GC-related performance problems or synchronization issues?
> On Jan 30, 2017, at 11:38 AM, martin brooks <brooks.martin at sympatico.ca <mailto:brooks.martin at sympatico.ca>> wrote:
>> This machine: http://dap.xeonphi.com/ninja-dev-platform-pedestal-ac.aspx <http://dap.xeonphi.com/ninja-dev-platform-pedestal-ac.aspx>
>> Here is my computational photography code (varilets.org <http://varilets.org/>) generating > 100 core load on 256 cores.
>> Martin Brooks
>> <100 cores.png>
>> Openmcl-devel mailing list
>> Openmcl-devel at clozure.com <mailto:Openmcl-devel at clozure.com>
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Openmcl-devel