Update from the Frontier of Exascale Software Enhancement

In advance of the scheduled shipment this year of the U.S.’s very first exascale supercomputer, Frontier, at Oak Ridge National Laboratory, an worldwide crew of program builders led by a University of Delaware professor is doing the job on a plasma physics application.

An posting published yesterday in the university’s UDaily by Tracey Bryant details the work underway by UD’s Sunita Chandrasekaran, assistant professor of computing and facts sciences, alongside with her staff, which is 1 of 8 performing on apps for Frontier, an HPE-Cray system powered by AMD CPUs and GPUs.

The plasma physics software less than improvement, PIConGPU (Particle in Mobile), is developed to speedily produce simulations for subsequent-era plasma (particle) accelerators, significant to advancing radiation therapies for most cancers and to growing the use of X-rays to probe the structure of materials.

“Dr. Chandrasekaran’s PIConGPU crew is an elite team spanning quite a few geographic regions, scientific domains and backgrounds,” claimed Dr. Nicolas Malaya, technical guide from Sophisticated Micro Equipment (AMD) for the Exascale Centers of Excellence. “I fully expect this application to create critical scientific final results from this crew in computational science, supercomputing and plasma physics.”

Bryant’s post contains a Q&A with Chandrasekaran, right here are excerpts:

Q: How is the project going?

Chandrasekaran: Pretty excellent. We are thrilled to have gotten accessibility to the new AMD Instinct MI100 (information heart GPU) accelerator cards. We ran the comprehensive PIConGPU on these newly released playing cards, and in our scientific studies utilizing a solitary GPU, we observed a 1.4 instances enhance in speed in comparison to MI60. This is promising and provides us a lot to glimpse forward to, for the upcoming-era CPUs and GPUs for Frontier.

The crew is making use of accelerator cards like this from Highly developed Micro Units (AMD) to velocity the processing of plasma simulations and perform other intense calculations.

Q: In wanting at these two supercomputing titans, how do you assess Frontier’s pace to Summit’s?

Chandrasekaran: Chatting with my collaborator, Dr. Alexander Debus (head of the Center for Superior Programs Being familiar with [CASUS] at HZDR, a analysis laboratory based in Germany) served me make some observations — simulations like ours with PIConGPU that would take two months on Summit might conclude up taking one week on Frontier. This also usually means we would now be equipped to operate several 10-million time-action simulations on Frontier (just about every time step would take ~50 milliseconds). Time-phase simulations allow us to examine the procedure of the computer’s ability technique from hour-to-hour intervals, proper down to thousandths of a next.

Q: Who are your collaborators and what is it like coalescing an global crew?

Chandrasekaran: My collaborators are from ORNL, HZDR, CASUS, and the Georgia Institute of Engineering. I have not satisfied fifty percent of my team in man or woman, nevertheless it feels like we have been performing jointly for yrs. We are now a little family members. You should see this webpage for particulars.

As soon as each handful of months, we make guaranteed to focus on the team’s, as well the project’s, prevalent vision and objectives to assure the quick- and long-expression ambitions align very well with CAAR (cost investigation and resolution) deliverables. This is particularly important for an worldwide staff like ours. Most of the discussions and discussions are hashed out more than electronic mail/Slack prior to scheduling a team cellular phone contact, given that there are extra than a couple hours of time difference involving the U.S. and Germany.

Q: What is the most fascinating/rewarding aspect of the project for you?

Chandrasekaran: I feel it is the interdisciplinary element of this undertaking. It is intriguing to imagine about implementing computer system science concepts to a genuine-planet scientific software. I am also thrilled that our shut collaborations have led to this venture getting funded by Dr. Michael Bussmann (CASUS at HZDR, Germany). This is my to start with internationally funded collaborative undertaking.

Q: What are the parts exactly where Frontier is poised to have the biggest affect? Do you anticipate Frontier to enable advance long term virus analysis, for illustration?

Chandrasekaran: I feel so, particularly when we are in the section of integrating high-overall performance computing (HPC), synthetic intelligence (AI) and knowledge science. Massive-scale (and speedy) simulations that couldn’t be imagined just a couple several years in the past are now heading to turn out to be doable with the huge compute means that Frontier is heading to give. Not just virus investigation, but these kinds of compute abilities are of paramount significance to studies like finding a remedy for Alzheimer’s disorder or studying local climate improve.

Q: How are UD students contributing to the effort and hard work?

Chandrasekaran: My Ph.D. scholar, Matt Leinhauser, has been functioning on this task considering that its inception. With mentorship from myself and my CAAR workforce (especially Rene Widera, Sergei Bastrakov and former CAAR liaison Ronnie Chatterjee), Matt has been equipped to set with each other two technical files on profilers — these are tools that recognize parts in the laptop program that get the most computation time.  We have so considerably employed NVIDIA’s nvprof and Nsight profiler instruments to dive deeper into the code. HZDR also invited Matt to devote past winter (January 2020) with them, which was a gratifying possibility when he was nevertheless in his very first year of the Ph.D. software.

Q: What is on the horizon?

Chandrasekaran: With aid from the Frontier Center of Excellence team, we will be marching forward to port PIConGPU on the early access units and planning the application for Frontier, which is currently being built as we speak. As upcoming techniques, we will be functioning on optimizing PIConGPU on the early obtain programs and speeding up the simulations even even further.