On Tue, Nov 29, 2016 at 7:22 PM, Hee Il Kim heeilkim@gmail.com wrote:
Hi Frank,
How was the performance of KNL? I had a chance to run qc0 on a KNL test machine here (single node). It was 4.6x slower than Haswell 2.3GHz (AVX). The tests were done with -O1 option to bypass the issue of Intel 17.
That's approximately what I would expect. I've been fighting with an (unrelated, luckily) segfault until a few days ago.
More on the machine:
KNL 7230 Centos 7.2 with xppsl-1.4.3 Intel compiler 17.0.1 Quadrant mode Cache/Flat modes tested with AVX512/AVX2/AVX
Vectors thorn is going to be updated for AVX512?
Yes, the respective changes are implemented, and will be available after the release freeze is lifted.
-erik
Please take a look at the attached option file.
Hee Il
2016-11-23 6:42 GMT+09:00 Frank Loeffler knarf@cct.lsu.edu:
On Tue, Nov 22, 2016 at 04:37:19PM -0500, rahul kashyap wrote:
Let me port first and then I'll follow your instructions for the tests.
I've put together an experimental simfactory entry to the new nodes. It runs, and tests pass (except the ones from Dissipation). It should be considered experimental because it does not fully implement proper logging in, and I also used a compute node to compile (because I didn't want to mess with cross-compiling on my first try). This works, but makes using the machine a bit more 'hand-work' than others.
I'll upload those soon to simfactory, marked as 'experimental'.
Frank
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users