If you’re eager to find out who’ll supply NERSC’s next-gen supercomputer, codenamed NERSC-9, here’s a project update to tide you over until the winning bid and system details are revealed. The upcoming system is referenced several times in the recently published 2017 NERSC annual report.
NERSC-9, the successor to NERSC-8 (the Cray XC40 “Cori” supercomputer), is scheduled to arrive at Lawrence Berkeley National Laboratory’s Wang Hall in 2020. The RFP for the system went out in the third-quarter 2016 as part of the APEX procurement collaboration, but the winning bidder has not yet been publicly disclosed. That announcement is expected before year end and will be followed by a new call for NERSC Exascale Science Applications Program (NESAP) proposals shortly thereafter. It’s noted in the report that the applications that were part of the NESAP program for Cori saw speed-ups of 3X on average. NERSC-9 is expected to provide 3-4X the capability of Cori (whose 8,600 Knights Landing nodes deliver 15 Linpack petaflops) within an 18 MW peak power footprint.
Division Director Sudip Dosanjh reviews the mission need for NERSC-9 in the center’s latest annual report:
- Provide a significant increase in computational capabilities over Edison.
- Provide a system that will demonstrate exascale-era technologies and enable the user community to continue transitioning to advanced energy-efficient architectures.
- Provide support for new science initiatives and extreme data analysis from experimental facilities.
He adds, “NERSC-9 will be our first supercomputer with an all-flash filesystem to accelerate I/O. It will deploy a flexible, high-speed network to support large-scale simulation and data analysis, as well fast connectivity to experimental and observational facilities. We are very excited about the system.”
According to the APEX 2020 RFP (which outlines the technical requirements for both Crossroads/ATS-3 and NERSC-9), “the NERSC-9 system is expected to run for 4-6 years and will be housed in the Wang Hall (Building 59) at LBNL that currently houses the “Cori” system and other resources that NERSC supports. The system must integrate into the NERSC environment and provide high bandwidth access to existing data stored by continuing research projects.”
Signs from the annual report point to a GPU-accelerated NERSC-9. The NERSC staff has done a lot of work to characterize the GPU-readiness for the broad swath of codes it supports. A survey of relevant codes and literature studies found 42 percent of codes had at least some kernels that were GPU-accelerated already and 19 percent had a proxy application that was algorithmically similar. 14 percent of codes were identified as unlikely candidates for a GPU port and 25 percent were currently unclassifiable as to their GPU-readiness.
“The study also looked at a representative set of GPU-ready applications, benchmarked on GPU-accelerated platform at NVIDIA, the Swiss National Supercomputer Centre’s Piz Daint and the OLCF’s Summitdev,” NERSC reported. “The results showed that for well-optimized codes, single-node GPU-accelerated performance compared to a Cori node is typically 4 to 6 times greater for a compute-bound code and 2 to 4 times greater for a memory-bound code. However, at scale these speedups decrease depending on the respective communication characteristics of the application. From this study, NERSC concluded that it is feasible to consider GPUs in the NERSC-9 architecture.”
While we await the official NERSC-9 system announcement, pathfinding has already begun for NERSC-10, on track for the 2024/2025 timeframe. “A growing issue for supercomputing centers is the amount of power needed to operate leading-edge systems. NERSC started its transition to energy-efficient architectures with Cori, and our goal is to provide our users a consistent path to exascale technologies in the next decade,” said Dosanjh.
NERSC-11 is also on the horizon for 2028 with a proposed 5-10 exaflops of performance.
A NERSC presentation from earlier this year shows a timeline for NERSC’s leadership-class machines and facility upgrades.
Here’s a link to the full annual report (PDF). Get the download on some very interesting HPC-enabled science — and don’t miss page 67.