-
Notifications
You must be signed in to change notification settings - Fork 144
Conference call notes 20170802
Davide Vanzo edited this page Aug 2, 2017
·
1 revision
(back to Conference calls)
Notes on the 82nd EasyBuild conference call, Wednesday August 2nd 2017 (5pm - 6.pm CEST)
Alphabetical list of attendees (3):
- Damian Alvarez (JSC, Germany)
- Davide Vanzo (Vanderbilt University)
- Fotis Georgatos (Illumina, UK)
- Move packages from main R and Python easyconfigs to bundles
- Allow users to install locally with EB
- Optimizing code for multiple architectures
- Fotis: For R minimal installation with few modules and all additional packages organized in a single big bundle. Only a small subset of common packages are included in the R main module if at all.
- Davide: The main need of taking packages out of the main module is to better manage minimal toolchain dependencies and simplify builds.
- Davide: Is it possible with a HNMS to allow users to use EB to install in their home directories?
- Damian: Alan was working on something like this since it required changes on the framework. Need to be careful though on how you set the MODULEPATH.
- Fotis: Big supporter of having a flat NMS in parallel with HNMS to allow users to access it but not as default. This would allow users to use the naming scheme that best fits their need.
- Fotis: Allowing users to build software locally with EB would significantly complicate admins work when it comes to identify the source of a software issue.
- Davide: To transparently provide software stacks optimized for different CPU architectures within the same cluster it may be possible to use as software installation path a symlink that points to an NFS mounted directory that contains the stack optimized for that specific architecture. Or in alternative use single stack with binaries built with function multipathing to support the different architectures. Suggestions?
- Damian: Depending on the binary, dynamic function dispatcher may impact the performance because of added overhead, especially if significant number of loops with function calls inside are present.
- Fotis: At the moment we build for the oldest architecture that represents the minimal common denominator addressing the users that have IO bound applications. You have to be careful that some software may not build correctly on certain architectures.
- Damian: Same problem with compiling OpenBLAS on KNL and Haswell. The CPU architecture auto-detection fails when building OpenBLAS (supposed to be fixed in the last release).