Trends and Context for Environmental HPC
Presentation: pdf (7 MB)
This was the kickoff talk to a NERC “town hall meeting” workshop envisaged as part of the establishment of a formal NERC HPC strategy. After my talk, we had a three component workshop: some international talks, some breakout discussions, and some local talks. The main aim of the meeting was
To begin a process of establishing and maintaining a NERC strategy for high performance computing which reflects the scientific goals of the UK environmental science community, government objectives, and the need for international competitiveness.
The breakout groups covered four themes:
- How does our science face a future where computers do not run faster? Are there known algorithmic routes to improve time to solution for fixed FLOPs (but maybe enhanced parallelism) that we are not yet exploring? What else can we do?
- Which HPC related international collaborations do we depend upon, and how can we sustain these, intellectually and financially?
- What are the implications of the new UKRI funding routes (e.g. Industrial Strategy Challenge Fund) for environmental HPC resource requirements (software, hardware, people)?
- How do we gather information about the impact of our HPC related research? What are the routes to demonstrating economic impact?
… and my talk was attempting to provide an introduction and context for these objectives and breakouts. The key point of course is that free lunch of easy (and relatively cheap) computing is over. Advancing along our traditional axes of resolution, complexity, ensembles etc, will be complicated by harder to use and (relatively) more expensive computing and storage.
I finished with six key points:
- The NERC HPC recurrent budget is currently fixed, but there is usage growth via both growth within the existing communities, and the addition of new communities.
- Future computing will not be any faster than current computing, we will only get to do-more/go-faster by expending energy (electricy) and by being smarter about using parallelism, by algorithms, or software, or both.
- HPC platforms are changing, with more heterogeneity and customisation, and that’s happening just as we have to take on machine learning as a new tool.
- NERC needs plans and evidence to justify HPC infrastructure, now and into the future, especially if we want to grow it!
- We have ambitious international scientific competitors — both a threat (if we can’t compete computationally) and an opportunity (for collaboration)
- All of which is why NERC needs a strategy …
(Incidentally, the preference for lots of post-it notes in this presentation is a homage to town-hall meetings past and their flip-charts and “stick your comment on these posters” sesssions! EPSRC, I’m looking at you :-) )