[Users] meeting minutes for 2017-05-15

[Roland Haas]

Present: Andreas , Bhavesh, Eliu, Qian, Erik, Frank, Roberto, 
Vassili, Roland, Steve, Ian

Hangout limit:
* reached again, will try alternatives at increased reate
** could try webex but no Linux client and using a phone is not
convenient (eg one needs to use specific headset)
** will look into zoom, other options

New thorns:

LORENE:
* needs to be reviews

Failing tests:
* Roland has not looked into SphericalHarmoncRecon test
* Frank has not looked into GRHydro
** older tests by Roland indicate that the NaNs go away and the test
passed if one disables the assoicative math option of -Ofast
** right now the exact file / code that is affected by this is not
known now whether this is a compiler bug or not

KNL performance:
* is highly depended on MPI and OpenMP settings
* best situation is often pure MPI since our MPI parallelization is
  more efficient than our OpenMP parallelization
* comparison per node on NERSC Cori Erik finds that Haswell nodes are
  5 times as fast as KNL. This is due to our code generator being
  optimized for Haswell. For just a kernel Erik got 5% of peak on
  Haswell and 10% on KNL. So we should be able to better but it will
  require changes to the code generator
* the crucial change seems to be to introduce tile based thread
  parallelization. LoopControl does not quite work yet, Kranc already
  does this tiling but its overhead is too large.
* plus local stuff like vectorization, avoid divisions, sqrt

Vectorization in hydro code:
* the hydro code in CactusExamples performs better than GRHydro
* better able to use vectorization
* Erik would like to give a presentation on this in the near future

KNL on stampede:
* will only be able to provide simfactory files for KNL machines but
  not optimized code
* currently works on Cori and Marconi but not Stampede
* removing tacc_affinity helps a bit but now it crashes in
  TwoPunctures after interpolation
* Ian will provide instructions via email on how to obtain a core dump
  file

new user tutorial:
* add mention in tutorial that this file is tuned for single
  workstation and won't work on cluster with many nodes due to
  resolution being too low
* add some instructions on how to use ET as a black box
** diagnostics output in example parfiles
** env variables to set
** how to set up ET on new machine

RNSID:
* proposal to introduce RNSID into ET
* https://bitbucket.org/GravityPR/prthorns.git
* ask for inclusion process to be started just after the next release
  which should be about a month from now
* guidelines for contribution are:
http://einsteintoolkit.org/contribute.html

Yours,
Roland
-- 
My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from http://keys.gnupg.net.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
Url : http://lists.einsteintoolkit.org/pipermail/users/attachments/20170515/6954af98/attachment.bin