Eurolab4HPC supporting HPC innovation.
Expert view: Martin Schulz
- Power bound computing – use more nodes, etc. and find best configuration (makes more sense with the slides)
- Network contention: harder to predict execution times but cheaper in terms of hardware.
- Manage complex workflows
- Monitor the application and hardware stack at all points. [Seems more comprehensive than many do, but very interesting].
- Grafana
- Malleability
- Dynamic processing management in MPI not really used. It needs to be more adaptive. C.f. PVM (too dynamic). Bubbles in MPI 4. Could enable on-the-fly adjustability, workflow support and other things, ability to remove a group without global impact or duplicate them (cf. Apache Spark?). Applications need to understand how to work with this unless use very high-level abstractions.
- Questions:
- Heterogeneous systems? Code tightly or more abstract so more flexible? (cf. Chapel, X-25 which is relatively high abstraction). This exists a bit but needs to be done more, e.g. to leverage accelerators where they exist and so on.
Expert view: John D. Davis
- LOCA: HW/SW co-design for IoT to HPC
- Using ARM a decade ago
- European Processor Initiative (RISC-V)
- Questions:
- Co-design of what?
- Reconfigurable hardware?
- BLAS is still the base of very many calculations so given this perhaps there is less variability in workloads that might initially seem to be the case.
- Patent risk?
- Open source: some things are hard to open source.
- Co-design of what?
Discussion
- See slides
- New HPC applications?
- Libraries, frameworks.
- ML for simulation?
- Scaling – where speed needed. But some things scale weakly. Can use ensemble.
- New algorithms: multi-model, multi-scale, coupled simulations.
- Climate modelling – how much processing power required? 100 Exa FLOPS?
- Programming complexity?
- Neuromorphic?
- MPI, OpenMP likely to continue. But these change over time.