Telcon: 2016 07 21
Todd Gamblin edited this page Jul 21, 2016
·
6 revisions
- Todd Gamblin (LLNL)
- Benedikt Hegner (CERN)
- Brandon (NERSC)
- Cyrus Harrison (LLNL)
- Greg Lee (LLNL)
- Jim Galarowicz (Krell)
- Mario Melara (NERSC)
- Patrick Gartung (Fermi)
- Peter Scheibel (LLNL)
-
LLNL hackathon
-
NERSC hackathon (w/ORNL, LLNL, Kitware)
- By default we'll set
CRAYPE_LINK_TYPE = dynamic
- How to set to static if needed for some builds:
class SomePackage(Package):
def install(self, spec, prefix):
if 'platform=cray_xc' in spec:
env['CRAYPE_LINK_TYPE'] = 'static'
-
Module support good after hackathon
- need to make it easier to deal with externals
- packages.yaml can get kind of large.
- default config scope can help with this
- need to address versions in external package files
- detect different versions of builtin
-
better cray model detection
-
Currently interacting with Cray to see if we can get better clean environment.
-
Brandon: thanks to everyone for coming out to the hackathon
-
Cray defaults can benefit BG/Q machines
- know where externals and such are located by default
- Patrick: possible problems on BG/Q:
- bgclang, gcc 4.8.4, etc. don't know where their runtimes live
- How to deal with changing hashes and reinstalls: #1325
- Need better ways to maintain a production spack package repository
- a number of bugs blocking this (#1178)
- Working on fixes
- possible solutions hash issues:
1. concretizing based on what is already installed
- This is planned! but needs better constraint solving
- We've got a prototype implementation of better constraint solving from the NERSC hackathon
- thanks to Ben Boeckel from Kitware
- NOTE for @citibeth: this would definitely be optional and we'd allow clean-slate installs too. 1. Bring back "Views" (from the paper)
- project spack space onto smaller space (symlink tree) that users can rely on
- module names don't have hashes; same view can be regenerated even if hash changes
- new view might have new versions of things
- allow for seamless upgrades 1. more extensive package testing
- LLNL can do AWS testing
- Benedikt: CERN could do CentOS 7 and gcc testing
- CERN needs testing anyway
- nightly coverage of a larger portion of the Spack space.
- Todd to follow up w/Benedikt
- NERSC could also look into testing on Cray platforms.
- might have issues using enormous amounts of CPU time
- smaller set of packages could be workable.
- Brandon to investigate at NERSC.
- Thoughts?
-
OpenSSL pinging will be removed #1333
-
Peter:
- addressing some issues with properly mirroring packages that have
expand=False
set in version directive - also going to fix issues with caching tarballs downloaded from Spack create.
- Cyrus: wants more consistent optional Fortran support.
- Need a set of global variants.
- need better fortran compiler support in Spack (model as dependency)
- need variant forwarding to sub-dags (Greg Becker to work on this)
- Patrick Gartung:
- Adam Lyon did some similar work to add C++14 to dependencies
- added ways to propagate C++14 variants in DAG
- Todd: look at Adam Lyon's Spack fork
- Patrick Gartung:
- Many other PRs merged (see Pulse!)