Telcon: 2017 03 23
Mario Melara edited this page Mar 24, 2017
·
15 revisions
- Todd Gamblin (LLNL)
- Elsa Gonsiorowski (LLNL)
- Greg Becker (LLNL)
- Greg Lee (LLNL)
- Mario Melara (NERSC)
- Mike Collette (LLNL)
- Patrick Gartung (Fermi)
- Peter Scheibel (LLNL)
- Steve Leak(NERSC)
- Elizabeth Fischer (NASA)
- Kiel Friedt (LLNL)
- Mike Pozulp (LLNL)
-
Kiel Friedt is finishing PR #2917
- Demo results on CDash
- Finalizing file format + spec matching semantics
- Now using BuildBot to farm out builds to AWS (CentOS/Ubuntu)
- Currently farms out 1 VM per OS; all version/compiler combos are done on that VM.
- Next steps:
- Finer-grained builds: farm out compiler/OS pairs to single VM (parallelize by compiler)
- Even finer-grained builds:
- Implement parallel install DAG
- allow VMs to access shared NFS mount and build DAG bottom-up.
- We'll be using this for an xSDK release in April
-
Todd: finishing re-working to use
sort(key=...)
instead ofcmp
- Allows Python 3 compatibility
- Code much easier to understand.
- also slightly faster (reduces number of sort calls, and number of expensive compares)
- tests from 1m36s -> 1m28s.
- also slightly faster (reduces number of sort calls, and number of expensive compares)
- Next step: implement SAT rules with
pycosat
, speed up concretizer more.
-
Mario Melara
- hosted Spackathon at NERSC last week
- developed NERSC Spack deployment plan
- forked Spack and integrated PRs that NERSC wants
- NERSC plans to keep a fork so they can stay ahead of
develop
- we'll try to upstream their work
- NERSC plans to keep a fork so they can stay ahead of
- Steve (NERSC):
- setting up spack is currently more tedious than NERSC would like (packages.yaml configuration)
- Elizabeth:
- PRs could be merged faster.
- core PRs not getting merged as fast as we'd like
- package PRs being merged relatively quickly.
- suggestion:
- name core evaluators and allow them to sign off on core PRs
- Todd: easier to do when we have package testing -- core PRs will go faster then.
- Peter:
- just concretizing exercises a lot, but not as good as building
- Mike Pozulp:
-
ncurses
-P
flag caused intel builds fail -- example of a thing that would be caught by builds.
-
-
Mario Melara
- working on CUG paper about getting Spack working on Cray
-
Greg Becker
- helping LLNL code teams with Spack
- organizing a Spack tutorial at LLNL (on April 6)
-
Peter Scheibel
- build concretization in backlog -- merging that soon.
- recent PR 3530 fixes recurring problem with failed install directories lingering and breaking later builds
- fixing issues with resource staging
- boost build on Cray (Xeon Phi)
- Boost builds a build tool that it uses in the build -- breaks if the build tool is built for Phi.
-
Thanks to Adam Stewart, Denis Davydov, Elizabeth Fischer, and others for handling so many package PRs!
- This has helped tremendously with keeping things running smoothly
- Elizabeth:
-
CONTRIBUTING.md
contribution didn't go so well (too detailed -- included static/shared stuff) - Need multi-valued variants PR to do this properly.
- working on a set of guidelines
-
-
Elizabeth: spack server
- pie in the sky issue posted to github
- idea: building software is stateless -- separate concretization (stateful) from installation of spack
- server could take build requests and install packages on behalf of users.
-
Todd was interviewed for the Request for Commits podcast
- talked about OSS in government & Spack
- podcast will be part of RFC season 2 (coming out later this Spring)
-
Patrick Gartung:
- HSF packaging group meeting last week
-
https://indico.cern.ch/event/623070/
- using Spack to build LCG
- using Spack to build on macOS Sierra
- see CERN GitLab
- cf. LLNL's Lustre/ZFS/Spack build bot
- on github (URL TBD)
-
Interesting discussion on run dependencies vs. link dependencies for hashing (#3513)
-
Lack of merging of core Spack PRs is causing real problems:
-
spack load
has been broken long-term; the fix for it has been forgotten. - Enum variants still not merged; prevents fixing-up of build types (shared/static and debug/release).
-
-
Spack Server + Spack Environments: Commitment and status?
https://github.com/kielfriedt/spack-buildbot-config
- TBD