Telcon: 2017 06 08
Todd Gamblin edited this page Jun 8, 2017
·
11 revisions
- Todd Gamblin (LLNL)
- TBD
-
Tutorials!
-
Adam Stewart is giving a Spack tutorial at ANL on June 12:
- Title: Managing HPC Software Complexity with Spack
- Speaker: Adam J. Stewart
- Date: Monday, June 12th - Tuesday, June 13th
- Time: 1:00 PM - 4:00 PM
- Location: Argonne National Laboratory, Building 240, Rooms 1404-1405
-
Todd will give a Spack tutorial at the HPC Knowledge Meeting in San Sebastian, Spain on June 16.
-
-
Update on spack environment implementation
-
See the Environments Road Map
-
Peter Scheibel (LLNL) has implemented a "context"
-
list of abstract "user specs" combined with concrete versions of them
-
allows us to represent a spack stack in terms of:
- what the user asked for (user spec)
- what was actually installed (concrete spec)
-
Can use this to reproduce a stack in different contexts, e.g.:
- exactly the same stack, down to the architecture
- same stack, different architecture/compiler
- same stack, slightly different versions/variants
- same stack according to only the user specs
- i.e. only what the user asked for; otherwise use defaults for the reproducing machine
- etc.
-
Elizabeth: needs to be some combination of user-editability and cmdline would be nice.
- 1 file preferable for easy distribution
- Todd & Elizabeth: user single file can be hard b/c you need to bundle packages, patches, etc.
- single archive seems like it will be unwieldy if it's to be reproducible
- Seems like we could have a YAML file kind of like
Gemfile
and/orpip
'srequirements.txt
- user should edit user specs in this file or on command line.
- "building" or "installing" this file generates a single file that allows the whole thing to be reproduced.
-
Elizabeth: note on reconcretization
- option to reconcretize a spack context is important. Two options:
- add a new user spec to existing context;
- reconcretize based on what's already in the context.
- reconcretize the whole thing from the user specs.
- add a new user spec to existing context;
- option to reconcretize a spack context is important. Two options:
-
-
Greg Becker has a new PR allowing packages more control over their compiler flags
- https://github.com/LLNL/spack/pull/4421
- Unifies things like
CFLAGS
with the way Spack injects flags into the compiler wrappers.
-
We're adding a new configuration scope to Spack
-
/etc/spack
scope -- for sysadmins on machines to provide reasonable defaults - Plan to use this on our Power/GPU machines
- many different experimental compilers and MPI implementations floating around
- need very specific settings for the particular hardware
- want user compiles to "just work" on these systems.
- Spack will check in
/etc/spack
forpackages.yaml
,compilers.yaml
information- precedence would be to consider configs in this order (low to high, high overrides low):
- defaults,
/etc/spack
,$spack/etc/spack
,~/.spack
- defaults,
- precedence would be to consider configs in this order (low to high, high overrides low):
-
- Tom Merrick is playing around with
systemd
and trying to get that to work- Currently not finding
uchar.h
-
gcc
builds its ownlibstdc++
that needs a config option or a patch - Potential source of nondeterminism in our gcc build (we should route this to spack's
uchar.h
)
- Currently not finding
- Tom is also having issues with R activation
- https://github.com/LLNL/spack/issues/2951
- some R base packages are bundled with R, which is pretty standard.
- they're also in Spack so activating newer versions of them does not work well.
- Todd: we should ping glennpj
- KT wants to share a spack install
- share Trilinos package for all spack users
- allow users to leverage this but experiment with higher-level dependencies
- KT has noticed that lua dependent installs are "escaping" from Spack
- Ping @trws about this -- lua integration should be mirrorable.