Hpc trends-20150924
2 likes389 views
The document summarizes observations from two conferences on high-performance computing trends. It notes that the improvements in x86-64 performance are ending, and that heterogeneous computing with GPUs and other accelerators will become standard. Exascale computing is being taken seriously worldwide but efforts in the UK appear to be lagging. There is also a perceived disconnect between the fields of computer science and high-performance computing. It suggests that the HPC-SIG group can help address issues like software challenges, career paths for research software developers, and planning for exascale in the UK.
![24 Sept 2015
What do we do? [HPC-SIG big role here!]
Scope needed University HPC infrastructure for 2016
Maximise research return for researcher time invested
Try to capture requirements in 3-5 year horizon as well
Press your Universities for adequate capital and recurrent
spending for your next purchase
Software support and training are must haves
Talk to your local Computer Science Departments!
There may be a language barrier, but persevere
HPC-SIG push for broader UK exascale planning
Hartree (energy) and EPCC are leading UK institutions – Others?
Support efforts for Research Software Engineers / Developers](https://image.slidesharecdn.com/m2inux9atvov0m6ixm3j-signature-d880125333832b18328f0258bbe428118fcb19b57fe85f7c584531dd4ae6e366-poli-150929101501-lva1-app6892/85/Hpc-trends-20150924-7-320.jpg)
Hpc trends-20150924
- 1. Cliff Addison Observations on HPC Trends (HPC-SIG / CASC and ISC'15)
- 2. 24 Sept 2015 Outline HPC-SIG / CASC meeting Attendees largely HPC support and development people Many common experiences and interests Sustainability major undertone ● Who is going to do the work? ● New users not used to the command line. ● Move to exascale infrastructure – by whom for whom? ISC'15 Reasonably well attended – many top people from US Implications of, and what is needed for, exascale major themes Exposition not exciting; really enjoyed the technical presentations New approach, solutions needed – business as usual is ending
- 3. 24 Sept 2015 Observed trends – after the meetings The marvellous improvements in x86-64 performance are ending. Haswell nodes seem to saturate memory circa 16-20 cores. Indications 24 cores can overload an InfiniBand connection. Return of the RISCs? (ARM64, Power8, SPARC64 XI etc.) Heterogeneous computing will soon be standard on nodes Where useful today? GPUs making inroads; Machine Learning could be exemplar. How ramp local / regional / national HPC to support? Software caught in a perfect storm. Who will be doing co-design / tuning with hardware?
- 4. 24 Sept 2015 More observed trends Exascale treated seriously in EU, Japan, China and US UK efforts / thinking / planning appear to be lagging Increase in compute per watt particularly relevant (Hartree) Big Data and HPC blending together Push to compute / analyse in place rather than move data Linked experimental work and modelling Deep Learning deploys massive neural nets on GPU-HPC What about the exascale software and algorithms??? Software in an even deeper hole. On other end of the scale – many new users working from GUIs OpenStack / Docker sort of environments underpinning???
- 5. 24 Sept 2015 Exascale is not just hype! ISC'15 Consensus – circa 10 years for exaflop Race for 100 PF installation now Fujitsu FX-100 – SPARC64 XIfx with 3D memory (2016) Chinese Tianhe-2a – Matrix2000 2.4 TF GDDSP (2016) US CORAL – Intel Knight's Hill with 3D memory (2017) Critical relevance for HPC-SIG – what will £1M buy? Petascale energy projected 20kW; so 70kW about 3.5 PF Not enough happening / being invested in UK! Need full spectrum support – look at 3, 5, 10 year impact What should we look for in 2016 University HPC infrastructure?
- 6. 24 Sept 2015 Perceived disconnect between CompSci / HPC Extremely interesting / heated discussion at HPC-SIG / CASC Some reinforcement of this perspective at ISC In a nut-shell: HPC needs research level help dealing with topics as diverse as: Data analysis, visualisation, computational steering, Better user interfaces Resource management Programming environments / tools / algs for parallelism BUT much of this work has little academic kudos within Computer Science – interesting bits require Mathematical skills too Career paths for research software developers?
- 7. 24 Sept 2015 What do we do? [HPC-SIG big role here!] Scope needed University HPC infrastructure for 2016 Maximise research return for researcher time invested Try to capture requirements in 3-5 year horizon as well Press your Universities for adequate capital and recurrent spending for your next purchase Software support and training are must haves Talk to your local Computer Science Departments! There may be a language barrier, but persevere HPC-SIG push for broader UK exascale planning Hartree (energy) and EPCC are leading UK institutions – Others? Support efforts for Research Software Engineers / Developers
- 8. 24 Sept 2015 Final thoughts on ISC'15 Not all doom and gloom! Lots of exciting things happening. Conference succeeded in highlighting the software issues. Interesting work taking place; question of scale and timing Top US researchers made a big impact Some of these (e.g. Ed Seidel) also at HPC SIG / CASC EU HPC summit Prague, 9-13 May 2016 – not just Horizon 2020. Virtualized, readily deployed side of cloud computing is big. Machine learning, particularly deep learning, is really taking off. But need to identify ways to spot the “unknown unknowns” Quantum computing is still circa 10 years away