Why no one cares about HPC!

Mike Croucher

Product Manager for Cloud and ML

Numerical Algorithms Group (NAG)

My HPC career

• Used HPC (Sheffield)
• Supported HPC users (Manchester)
• Developed HPC services (Co-founded Sheffield RSE group)
• Head of HPC centre (University of Leeds)
• Commercial HPC and cloud (NAG)

Why am I here?

• I believe that HPC has the promise to fundamentally change how we do science
• I think that traditional HPC models don't deliver on the promise
• I believe that NAG can help cloud vendors deliver the promise

Why the title?

HPC = Geek Top Gear

Any questions?

• Audience member 1: What’s a core?
• Audience member 2: Why does it run my R script slower than my laptop?
• Audience member 3: Do you have Excel installed on it?

What we like

What they need

Traditional HPC is irrelevant to most researchers!

...and most of the rest use it badly.

Hannay et al (2009)

• Online survey of 1972 international researchers
• ~80% never use a supercomputer

Prabhu et al (2011)

• Interviewed 114 researchers at Princeton
• ~40% never use a supercomputer

Prabhu et al (2011)

• "Despite enormous wait times, many scientists run their programs only on desktops"
• "About a third of researchers did not use any form of parallelism in their research at all"
• “Currently, many researchers fit their scientific models to only a subset of available parameters for faster program runs.”

and yet....

Prabhu et al (2011)

“Across disciplines, an order of magnitude performance improvement was cited as a requirement for significant changes in research quality”

and yet....

Postgrad sleeping in computing lab to guard 3 machines instead of using the HPC centre across the road!

why?

Potential HPC users are changing

What Software?

More people need HPC than ever

But it's so hard to use!

My mobile back then

My HPC back then

My mobile now

My HPC now

What users are used to

What we give them

The queue

We can do better

New Geek Top Trumps?

Our HPC is the easiest to use in the world!

What is HPC?

Croucher (2020): You are doing HPC as soon as you CARE about speed, memory or storage.

Many people will disagree with this

Remember this?

“Currently, many researchers fit their scientific models to only a subset of available parameters for faster program runs.”

Supporting the long tail

End to end support

• Idiomatic programming / application use
• Faster language / software
• Algorithmic improvements
• Care about hardware (threads, AVX, GPU, FPGA)
• Scale up (Cloud)
• Deploy to users (Cloud)

What happens when you provide this support?

They use more of everything!

ooominds speed-up

Toy data set: 15,000 events. Matrix size 8000 x 8000

No parallelisation

Original code took 5725 seconds

New 32 core sparse CPU version: 131 seconds

V100 GPU version: 49 seconds

The response?

Full data set with 11 million events (733x bigger)

Matrix size 43,000 x 43,000 (29x bigger)

Working on deployment solution in the cloud to allow other linguists to upload data

Want to do cost comparison. What is the cheapest way to do this analysis?

The HPC dream

HPC that's easy to use

Support partnership that ensure optimal use of HPC

Together we can revolutionise science