Conference call notes 20131105

Notes on the 1st EasyBuild conference call, Tuesday Nov. 5th 2013 (3pm -3.30pm)

planned event
link to Google Hangout: https://plus.google.com/116729781669672561222/posts/3LfHWAy7dDy
note: direct link breaks easily, should follow link in "Video calls" section on EasyBuild community Google+ page instead

Alphabetical list of attendees (11):

Planned topics to be discussed:

(short) progress on support for new easyconfig format (Kenneth)
progress report on adding support for UNITE and the individual performance tools that are part of it (Bernd)
current status of 'toolchain soup' (Ward)
remaining time is for open questions on broad topics, which included:
loading various profiler/debugger modules at the same time, using them at the same time (Fotis)

Notes on topics discussed during the conference call (by Kenneth)

Actual timing: 3.10pm - 3.45pm (CET)

hard limit of 10 attendees in Google Hangout
side-note: people at same site should sit together if feasible
no easy way to obtain a reliable direct link that can be communicated shortly before the conf call to let people join
easiest way to join in is to follow the "Video calls" link on the EasyBuild Google+ community page

pull request for adding initial support for new easyconfig format open, but still WIP (see framework#693, stdweird-framework#22)
goal is to merge this in for EB v1.9.0
will not provide support for usable easyconfig files in this format yet (most likely)
(very) preliminary documentation available
(Bernd) will section markers for dependencies (e.g. [dep:PAPI > 4]) also be supported?
(Kenneth) in the end, yes, but not initially
(Ward) how will patch file specifications be handled?
(Kenneth) a default list of patch files can be specified, version or toolchain specific patches can be added in corresponding sections
(George) what about conditional constructs that are currently used in easyconfig files?
(Kenneth) these will still be possible in the Python header part, but typically indicate a missing feature that should be implemented in the framework (or the need for an easyblock)
(Andy) will section markers be hierarchical?
(Kenneth) yes, e.g. a goolf toolchain section with a recent version section below that:

[goolf]
foo=bar
[[> 2.0]]
foo=rab

UNITE is a set of performance tools packaged together, in a framework that makes it easy to install and use them (see http://apps.fz-juelich.de/unite/)
easyconfig files for about 24 of these performance tools were produced during and after the last hackathon
only for the gompi toolchain, following the 'minimal toolchain' approach (i.e. some easyconfigs only use GCC as toolchain)
also needs to be tested for goolf, ictce, etc.
another 10-15 tools can be added to this set
easyconfig files will be contributed to the community soon
(KH) already done in mean time, see easybuild.experimental - users/unite (make sure to check the README.md file!)
some rather annoying issues on openSuSE turned up w.r.t. lib64 vs libexec, e.g. for GCC, OpenMPI, Paraver
see easyblocks#283

question by Ward: You want to build something with a certain compiler but first you need to update several deps to the correct toolchain version, create new easyconfigs because they only exists for gcc and not ictce (or vice versa), ... Is this something the v2.0 format will solve?
--try-toolchain --robot doesn't work as expected yet (see framework#474)
so, currently requires 'manual' work
module for new toolchain and toolchain elements need to be provided first
exhaustive listing of all easyconfig files that need to be rebuilt using other toolchain
new easyconfig format should make this a lot less painful, but is no magic solution
if easyconfig for foo does not have support for building with toolchain bar, is will still need to be added somehow (either manually or via --try-toolchain)
wiki page for pros/cons of minimal vs full toolchains: https://github.com/hpcugent/easybuild/wiki/Minimal-vs-full-toolchains
lets try and document pros and cons of both approaches there, and then revisit the discussion once it's worked out
dummy toolchain is mainly intended for constructing compiler toolchains (e.g. for building GCC as the compiler for goolf
can in principal also be used for 'compiler-independent' software packages (e.g. GDB, Valgrind, CMake), but should be done with great care (e.g. no dynamically linked libraries provided by compiler, because loading an (older) compiler module may break the tool)

Provide feedback