Discussion contributions prepared by
During the discussion we identified four more or less connected problem areas:
o Multiple solutions are used in production (UPS, CMS environment, lcgcmake, orch+waf,...).
o Common agreement that we need not a single software stack for everybody, but tools to allow people to assemble whatever they need.
o Common agreement that source code of software build needs to be preserved/kept as copy.
o Broad consensus that the ability to patch software is a requirement. Patches may be needed for urgent fixes, adaption of packages to new platforms or compilers, or for ensuring relocatability. HSF should act as point of exchange for these patches
o Giulio presented the way of CMS. One git repo per package is used for the bookkeeping of patches. This triggered a discussion on the need to push patches upstream and whether the CMS system isn't over-engineered.
o Marco, Ben, Giulio proposed to set up a system of largely independent layers - build layer, dependency handling, packaging. With an agreed 'protocol' between these layers. This triggered a longer discussion with the following outcome:
- The build layer should allow the build of a single package in a configure+make style; this would allow existing systems to take advantage of it. CMS expressed interest in this.
- for newly starting efforts one of the existing systems could be extended/adapted to cover all three layers.
- An open question is how relocation scripts affect the decoupling of these layers
- Discussion on whether the packaging step should be a 'catch-all' or specific to the target system. Target oriented seems to be the best idea.
o Support for multiple platforms required
o Should take advantage as much as possible of the target system; allow to mix local installations and custom build packages
o Deployment should be possible entirely in user land
o Relocation needs to be supported
o Andrea Valassi presented the current status of HEP_OSSlibs, which ensures a minimum host environment on RH/Centos systems
- raised the question how it interconnects with HSF releases software
o Solutions like chroot, RH software collections, python's virtualenv were presented
o Little experience in the community on this.
o Agreed that we need to look into the potential of these solutions
o The usage of dockers was seen as a possibility to tackle multiple issues - runtime environment, deployment in the Grid, ...
o Pere Mato and others proposed the concept of layered containers - base HEP system, commonly shared packages, experiment specific container
o It can be used to run one linux flavour on a differently flavoured host or other systems using boot2docker
o Dockers does not support diamond like container setups. So only simple setups can be provided here.
o Containers are per default stateless. Needs change to be useful for local/interactive usage.
o With little experience on all sides this needs closer investigation.
Possible follow ups:
a few people volunteered to follow up with the package build 'protocol'
looking into existing solutions for virtual environments
investigate docker potential and use cases
We plan to have a follow up meeting in ~1 month from now.