In this deck from the UK HPC Conference, Alan Real from Durham University presents: Convergence for Scientific Method - HPC, AI, Simulation and Experiment. "This talk will discuss how advances in instrumentation have caused many new areas to embrace HPC and AI in order to successfully conduct and understand their experiments." Watch the video: https://wp.me/p3RLHQ-l0B Learn more: http://hpcadvisorycouncil.com/events/2019/uk-conference/agenda.php Sign up for our insideHPC Newsletter: http://insidehpc.com/newsletter

1. Convergence for scientific
method – HPC, AI, Simulation
and Experiment
HPC-AI Advisory Council
Presented by Alan Real, Durham University
17th September 2019

2. ∂
• A City in the North East of England
• Steeped in Heritage and Culture
• UNESCO world heritage site & other heritage assets
• 3rd oldest University in UK, 1837 Royal Charter
• 7th Century history: Venerable Bede.
• 16 Colleges, Faculties of Sciences, Social Sciences
and Arts and Humanities

3. ∂
• 18,000 students, 3000 staff
• HPC, high priority for science faculty.
• Most highly-cited papers are computational
• Created dedicated research computing
support Directorate, within Research function

4. ∂
ARC@Durham
Advanced
Research
Computing
Research
&
Innovation
CIS / IT
Services
Applied
Research
Computer
Science
Enabling research to be at the forefront of computational
and data-intensive practice.
Key components
• People & Skills
• Research Aspirations
• eInfrastructure: Hardware & Software
Interdisciplinarity
• Collaborative culture
• Bridge between Academic Practice and operational
services
• Networks of cross-cutting methods, underpinning
disciplinary themes.
• Business development help maximise opportunity.
Team
• Research Software Engineering (3 FTE)
• Platforms (5 FTE)
• Business development
Values criticality and contribution of technical expertise.

5. ∂
The research computing platform
Data
storage &
Networking
Cloud
HPC
Visualization
Data
repository
Database
Lab
equipment
Data sets
Advice
Software
Research
question Value

6. ∂
HPC @Durham
18 Years of frankly heroic effort by 2 individuals running two facilities:
Lydia Heck – COSMA – Institute of computational cosmology
Henk Slim – Hamilton – general purpose, mainly science Faculty
Purpose built room: AHDC-1
Upgrades to power and cooling ~540kW capacity (2001-4)
Generator, cooling and UPS refurbishment (2009)
Second purpose built room: AHDC-2 (adjacent & Now Lydia Heck DC)
1MW electrical infrastructure, 800kW cooling load.
Have housed 7 generations of COSMA and Hamilton facilities

7. ∂
COSMA-3 (2006)

8. ∂
COSMA-4 (2010)

9. ∂
COSMA-5 (2012)

10. ∂
COSMA6&7 (2017-18)
COSMA6 relocated from Hartree centre & combined with COSMA5 ~11k cores

11. ∂
Main driver for HPC strategy…
The scientific method requires it!
Key metric of success:
Were the objectives of the science case met?

12. ∂
Virgo consortium outputs
Simulations run for months
Millennium:
3rd most cited paper in
Nature
Eagle: 4096 cores, 47 days
Most highly cited paper
in space science in
2015
5 years from inception to publication of the first paper
10-15 people designing and executing the simulations, exploited by 100s

13. ∂
System stability
Understanding the formation of the Universe in unprecedented detail
Begins here…

14. ∂
The knowledge continuum
Plumbing
Researchquestion
Application
Software
ComputerHardware
Domainknowledge
Systemdesign
Computercomponents
Workflow
Numerics
Operatingenvironment
Require building of intellectual capability across whole spectrum
Interplay of components is important
Ensure flow of knowledge across spectrum to enable research.
Algorithms

15. ∂
Conveying computational challenges
What does successful research
computing look like?
Lots of conversations & conferences
Some internal business cases.
Not much cite-able literature.
If it matters:
Can we publish a peer-reviewed
publication?
Can we do this within the research
domain?
Encourage other members of this
community to do the same!

16. ∂
N8 Research Partnership is the 8 most research intensive universities in
the North of England – Durham, Lancaster, Leeds, Liverpool,
Manchester, Newcastle, Sheffield and York
N8 Centre of Excellence for Computationally Intensive Research (N8 CIR):
– Established to build on this partnership to expand the art of the possible
within research themes by leveraging computational- and data-intensive
practice.
Derived from N8 HPC which offered access to a shared compute platform
for all 8 universities for 6 years.

17. ∂
NICE19
Conversation around future of AI & accelerated computing
GPU card memory regarded as limiting factor – explore coherence.
Quantification of inferencing error through model uncertainty.
Conversation with the Bioscientists
Want to get the most out of lab experiment data
Identical challenges within advanced materials
Bring experiment and simulation data together to inform each other.
- Requires ability to drive the network.
- Combine observation, AI, models & simulation to advance
understanding.
- In AI, only the machine learns to predict.
- We build models to understand & use simulation to verify.

18. ∂
Experiment & AI workflow
Experiment Observation
Data
Analysis
AI Prediction
Where is the hypothesis?
Where is the Understanding?

19. ∂
Convergence…
Data
Observation
PredictionUnderstanding
Experiment
ComputationTheory
Instrumentation
SimulationModelling

20. ∂
Final Remarks
HPC (& broader research computing) is part of scientific method.
Be willing to do something different!!
It is getting increasingly difficult to undertake modern experiment without
HPC.
Convergence, for me is not having a single platform to conquer them all:
It is about the need to explore phenomena from different perspectives.
Only through using simulations of models in conjunction with AI and data,
will we understand the models our machines learn.

21. ∂
Acknowledgements
ICC @ Durham
Lydia Heck, Carlos Frenk, Richard Bower
DiRAC consortium
Mark Wilkinson, Clare Jenner, Jeremy Yates
Hamilton HPC
Stewart Clark, Henk Slim