site-specific documentation for summit
Using UPC++ on Summit can be challenging, because this system is a "special snowflake" in various non-meaningful but annoying ways.
We should add some site-specific docs somewhere (probably on our wiki) to clue users in to at least the following:
module use /gpfs/alpine/world-shared/csc296/summit/modulefiles
to get the installs we maintain. These ideally would be in the default module path, but sadly are not.module load pgi-cxx14/default
to get a libstdc++ for PGI compiler that is C++11 compliant (the default is horribly broken)- Some notes about NUMA-ness, multi-rail and the recommended way to spawn jobs
Comments (4)
-
-
I have begun revisions to the
site-docs.md
document in the wiki of my own fork:
https://bitbucket.org/PHHargrove/upcxx/wiki/docs/site-docs
So far I've only tweaked the intro and updated the Cori information to include PrgEnv-cray.Unfortunately there is no facility for pull requests of wiki content.
I'll indicate in this issue when the document in my fork's wiki is ready for review.
I am open to suggestions as to how the review can best be accomplished.TODOs:
- Add PGI use example to Summit section, including the
module load pgi-cxx14
incantation. - Add Summit-specific info about NUMA/multirail and any spawn instructions beyond use of
upcxx-run
. Hopefully this can be accomplished by reference to the center's examples for OpenMP (or just MPI when upc++ job is single-threaded). - Update "2019.9.0" to "2020.3.0" in every instance
- Add PGI use example to Summit section, including the
-
Pull request 189 contains the content I am proposing to resolve this issue.
-
- changed status to resolved
Resolved via pull request 189
- Log in to comment
We already have a public wiki document I generated for last quarter's tutorials:
https://bitbucket.org/berkeleylab/upcxx/wiki/docs/site-docs
That is linked from the README.md files for both NERSC tutorials in in upcxx-extras, but not yet from any other places I am aware of.
I plan to add PGI and NUMA/multi-rail information requested to that document, as opposed to creating a new one.
Additionally, I'll seek an appropriate place to like to that document from the README.md