Discussion Summaries¶
Benchmarking¶
Extended Summary¶
General appetite for more information on how to benchmark and profile and what tools are available to help?
Very useful for RSEs and beyond working on HPC
Benchmark reproducibility and how to measure, how to set it up? Techniques and strategies
- What should be in a common repository?
- How do you moderate content?
- How do you maintain/update benchmarks?
- What should be included, e.g. ML
- What is a fair comparison
Scope for pulling together a UK benchmarking interest group
- Potential impact:
- Procurement, unified/verifiable acceptance approach across sector
- Beneficial for purchasers and vendors
- Ops longitudinal performance/impact of upgrades
- Advising and helping users to use resources efficiently
- Which resource to use?
- How to use it, e.g. scaling, internal paramaeterisation
- Licensing:
- Material included must be free of license conditions, e.g. inputs for some codes fall under license. Included performance data must be openly available e.g. free of NDA
- HPC-UK Benchmarking is a fantastic resource and huge effort from Andy Turner, Jeffrey Salmond.
- Additional features wish list:
- Profiling to better understand performance characteristics
- Strong and weak scaling, understand codes, and system dependence.
- Expanded range of codes and test cases, better coverage
Community¶
Extended Summary¶
- HPC Champions is wanted
- Concluded that we want to continue to grow the community - What is HPC Champions for?
- Ensure it is not just an annual event and folk who cannot attend events are also included
- Co-locate with partner organisations - RSE, HPC-SIG
- Promote collaboration with industry
- Wide range of committee membership: RSE, Ops, Users, Industry?
- Encourage tiered Champions community events not just National HPC Champions
- More than events, a slack channel (#Champions @ ukrse.slack.com)
Engagement¶
Extended Summary¶
Diverse ways on how we get / increase engagements
Good to compare notes / approaches between sites
Q: How do we find the users who aren’t engaged with HPC?
- Issues:
- Academics are pressed for time
- Some departments/users/groups are more accessible that others
- Users not aware of potential value, code works means it’s ok
- Possible Solutions:
- Training,
- Coffee mornings,
- Advertising on screens - sometimes an option,
- Awareness of who’s available,
- Liaising with CDT
- 2 days of training is good, fast is good
- How to engage HPC users?
- Driving tests, periodic account expiry
- Continued training
- Symposiums (e.g. Bath)
Impact¶
Extended Summary¶
- New positions - we get flexibility but little guidance on what success means.
- RSE role/collaborators/projects can be so varied that no one size fits all.
- In the design of metrics, who are we trying to convince?
- Ideas:
- Publications
- Talks
- Workshop organisation
- Customer statements
- Training
- Quantification of benefits inc staff effort
- Funding proposals (Success rates, total income?)
Training¶
Extended Summary¶
- Yes. Support as much (good) training as possible.
- There is a distinction to be made between training to introduce you to a specific resource (eg Tier 3) and general skills courses like HPC carpentry - how to create roadmap?
- HPC Champions should contribute to and publicise good training materials (e.g. HPC Carpentry) & help with obtaining funding to hold training sessions?
- It’s tough to train with site-specific information in such a way that you make it clear what isn’t site-specific – this is especially important if people come from other institutions to shared training – which is something that would be ideal if HPC Carpentry training sessions became common.
- More modular training better - 4-5-day courses on e.g. all of Unix, Python, Git, HPC, OpenMP, MPI will be exhausting for the learner.
- But then you have challenges around prerequisites.