A new 44-petaflops (theoretical peak) supercomputer is under construction at the Department of Energy’s Argonne National Laboratory. Called Polaris, this new supercomputing star has been selected to light the way to exascale and to Aurora, the exascale-class Intel-HPE system that’s had its delivery pushed to next year (2022).
Ahead of today’s official unveiling announcement, HPCwire spoke with Kalyan Kumaran, senior computer scientist and the director of technology at Argonne’s Leadership Computing Facility (ALCF) about how the laboratory will be using the system as a stepping stone to Aurora and beyond.
Built by HPE and powered by AMD CPUs and Nvidia GPUs, Polaris will enable researchers and developers to test and optimize software codes and applications to address a range of AI, engineering, and scientific projects planned for the forthcoming Aurora supercomputer, a joint collaboration between Argonne, Intel and HPE.
The installation currently underway spans 280 HPE Apollo Gen10 Plus systems across 40 racks, aggregating a total of 560 AMD Epyc Rome CPUs and 2,240 Nvidia 40GB A100 GPUs with HPE’s Slingshot networking. As part of a planned upgrade, the second-gen Epyc Rome CPUs (32-core 7532 SKU) will be swapped out in March 2022 for third-gen Epyc Milans (the 32-core 7543 part). At the same time, Polaris will transition from the Slingshot 10 to Slingshot 11 fabric (the same as Aurora will use). The system uses air-cooled HGX “Redstone” boards with liquid-cooling employed by rear-door heat exchangers.
At 44-petaflops (double-precision, peak) Polaris would rank among the world’s top 15-or-so fastest computers. The system’s theoretical AI performance tops out at nearly 1.4 exaflops, based on mixed-precision compute capabilities, according to HPE and Nvidia.
Polaris will tie into ALCF’s two 100 PB globally accessible Lustre filesystems named Grand and Eagle, which are backed by HPE’s Cray ClusterStor E1000 platform. Installed in January of this year, each storage array controls 8,480 disk drives with a sustained transfer speed of 650 Gbps, according to the ALCF documentation.
The choice of the Apollo Gen10 Plus rather than the HPE Cray EX architecture was deliberate, owing to Gen10’s flexibility to support additional configurations. “Each of these chassis actually fits two (single-socket) nodes, and they do support other accelerators,” said Kumaran. “So in the future, we could have new Apollo chassis added to this configuration, which would support, say, Nvidia GPUs on one side and maybe some other GPUs on the other side. And in the future, they could be supporting other AI accelerators.”
This is one of the ways that Polaris may continue to be an avenue for future research work even after Aurora – a Cray EX design – is deployed, said Kumaran. The Argonne Lab has been something of a hot spot for exploring emerging AI hardware. Its AI test bed currently includes a Cerebras CS-1 system, a Graphcore Colossus GC2 system, a SambaNova DataScale machine and (coming in 2021) Groq accelerator hardware.
Polaris will provide roughly four times as much compute power as Argonne’s 7-petaflops Linpack (11.7-petaflops peak) HPE/Cray XC40 Theta system, which was installed in late 2016 to be a companion and ramp machine for an earlier unrealized conception of Aurora (aka A18). At the beginning of this year, thanks to CARES Act funding, the lab added 24 Nvidia DGX A100 nodes to Theta, significantly boosting its capabilities.
With its heterogeneous CPU-GPU architecture (in a 1:4 ratio), Polaris is helping Argonne make the transition to the Intel-HPE Aurora system, which slipped from 2021 to 2022 on account of Intel roadmap delays (impacting Sapphire Rapids and Ponte Vecchio). Polaris will be used by researchers within the DOE’s Exascale Computing Project and the ALCF’s Aurora Early Science Program to start prepping their codes for Aurora.
“We looked at many possible solutions with Aurora in the back of our mind,” said Kumaran of the technology selection process. “We wanted something with multi-GPU node support. And we wanted something that would support some of the key programming models on Aurora, which is MPI, OpenMP, and also SYCL in DPC++ (the SYCL 2020 variant from Intel). We wanted these programming models supported, and Polaris offered that solution.
“It has multi GPU nodes. It supports the programming models. It’s got the same Slingshot interconnect that will be on Aurora. And our Early Science Program has applications in normally the traditional HPC simulation space, but also the data and learning space. So we wanted a number of optimized frameworks, optimized Python support, and things like that, that will be available on Aurora for these applications to make progress. And that’s available with the Nvidia and HPE solutions.”
Projects highlighted by Argonne include:
Advancing cancer treatment by accelerating research in understanding the role of biological variables in a tumor cell’s path by advancing the use of data science to drive analysis of extreme-scale fluid-structure-interaction simulations; and predicting drug response to tumor cells by enabling billions of virtual drugs to be screened from single to numerous combinations, while predicting their effects on tumorous cells.
Advancing the nation’s energy security, while minimizing climate impact with biochemical research through the NWChemEx project, funded by the DOE’s Office of Science Biological and Environmental Research. Researchers are solving the molecular problems in biofuel production by developing models that optimize feedstock to produce biomass and analyze the process of converting biomass materials into biofuels.
Expanding the boundaries of physics with particle collision research in the ATLAS experiment, which uses the Large Hadron Collider (LHC), the world’s most powerful particle accelerator, sited at CERN, near Geneva Switzerland. Scientists study the complex products from particle collisions in very large detectors to deepen our understanding of the fundamental constituents of matter, including the search for evidence of dark matter.
“Polaris is well equipped to help move the ALCF into the exascale era of computational science by accelerating the application of AI capabilities to the growing data and simulation demands of our users,” said Michael E. Papka, director at the ALCF. “Beyond getting us ready for Aurora, Polaris will further provide a platform to experiment with the integration of supercomputers and large-scale experiment facilities, like the Advanced Photon Source, making HPC available to more scientific communities. Polaris will also provide a broader opportunity to help prototype and test the integration of HPC with real-time experiments and sensor networks.”
The lab already has some experience with HPE systems infrastructure, including Slingshot and HPE Performance Cluster Manager (HPCM). A testbed rack called Crux includes AMD Rome processors, Slingshot technology and HPCM. “In that sense, Polaris is another testbed to continue testing HPCM at scale and getting ready for Aurora’s arrival,” said Kumaran, “not just on the applications side, but also being able to test the system software and Slingshot.”
A wider goal, long-sought and steadily inching forward, is cross-platform code portability. Argonne has researchers working with NERSC (Berkeley Lab) and Codeplay (prominent SYCL supporter) to port SYCL and DCP++ to the A100 GPU. “If people are porting code to Aurora using SYCL or DCP++, they will be able to continue to support that programming model and not have to rewrite to OpenMP or MPI or CUDA to use on Polaris,” said Kumaran. “And similarly, we’ve also explored supporting HIP on this platform (Polaris), and so if you have CUDA support, and you are developing with CUDA on Summit, or for future AMD-based platforms, with Frontier, then you can use that. And finally, we are also exploring SYCL and DCP++ for AMD GPUs [in collaboration with Oak Ridge and Codeplay]. And so if you’re looking for an alternate solution to CUDA and HIP on AMD GPUs and you want to run your DCP++ code, we have a proof-of-concept working on that.”
Polaris has been delivered and is currently being installed. Deployment for early science work related to exascale-readiness is expected in the first quarter of next year.