NERSC Finalizes Contract for Next-Generation Supercomputer


Might 5, 2020 — The Nationwide Vitality Analysis Scientific Computing Heart (NERSC), the mission high-performance computing facility for the U.S. Division of Vitality’s Workplace of Science, has moved one other step nearer to creating Perlmutter — its next-generation GPU-accelerated supercomputer — accessible to the science group in 2020.

Picture courtesy of NERSC.

In mid-April, NERSC finalized its contract with Cray — which was acquired by Hewlett Packard Enterprise (HPE) in September 2019 — for the brand new system, a Cray Shasta supercomputer that may characteristic 24 cupboards and supply 3-Four occasions the potential of NERSC’s present supercomputer, Cori. Perlmutter might be deployed at NERSC in two phases: the primary set of 12 cupboards, that includes GPU-accelerated nodes, will arrive in late 2020; the second set, that includes CPU-only nodes, will arrive in mid-2021. A 35-petabyte all-flash Lustre-based file system utilizing HPE’s ClusterStor E1000 {hardware} can even be deployed in late 2020.

Since saying Perlmutter in October 2018, NERSC has been working to fine-tune science purposes for GPU applied sciences and put together customers for the greater than 6,000 next-generation NVIDIA GPU processors that may energy Perlmutter alongside the heterogeneous system’s AMD CPUs. Practically half of the workload presently working at NERSC is poised to make the most of GPU acceleration, and NERSC has performed a key function in serving to the broader scientific group leverage GPU capabilities for his or her simulation, knowledge processing, and machine studying workloads.

On the core of those efforts is the NERSC Exascale Science Functions Program (NESAP). NESAP partnerships enable tasks to collaborate with NERSC and HPC distributors by offering entry to early {hardware}, prototype software program instruments for efficiency evaluation and optimization, and particular coaching. During the last 18 months, NESAP groups have been working with NERSC workers and NVIDIA and Cray engineers to speed up as many codes as attainable and be certain that the scientific group can hit the bottom working when Perlmutter comes on-line.

For instance, utilizing the NVIDIA Volta GPU processors presently accessible in Cori, NERSC has been serving to customers add GPU acceleration to numerous purposes and optimize GPU-accelerated code the place it already exists, famous Jack Deslippe, who leads NERSC’s Utility Efficiency Group.

“We’re excited in regards to the progress our purposes groups are making optimizing their codes for present and upcoming GPUs,” Deslippe stated. “Throughout all of our science areas we’re seeing purposes the place a V100 GPU on Cori is outperforming a CPU Cori node by 5x or larger. These efficiency positive aspects are the results of work being performed by tightly coupled groups of engineers from the purposes, NERSC, Cray, and NVIDIA. The keenness for GPUs we’re seeing from these groups is encouraging and contagious.”

As a part of NESAP, in February 2019 NERSC and Cray additionally started internet hosting a collection of GPU hackathons to assist these groups acquire data and experience about GPU programming and apply that data as they port their scientific purposes to GPUs. The fifth of 12 scheduled GPU hackathons was held in March at Berkeley Lab.

“These hands-on occasions are serving to be certain that NESAP codes and the broader NERSC workload might be able to make the most of the GPUs when Perlmutter arrives,” stated Brian Friesen, an Utility Efficiency Specialist at NERSC who leads the hackathons. “In some circumstances, NESAP groups have achieved vital speedups to their purposes or key kernels by taking part in a hackathon. In different circumstances, groups have developed proof-of-concept GPU programming strategies that may allow them to port their full purposes to GPUs.”

In the meantime, NERSC and NVIDIA are collaborating on revolutionary software program instruments for Perlmutter’s GPU processors, with early variations being examined on the Volta GPUs in Cori:

  • Roofline Evaluation: The Roofline Mannequin, developed by Berkeley Lab researchers, helps supercomputer customers assess the efficiency of their purposes by combining knowledge locality, bandwidth and parallelization paradigms right into a single determine that exposes efficiency bottlenecks and potential optimization alternatives. NERSC has been working with NVIDIA to create a technique for Roofline knowledge assortment on NVIDIA GPUs, and a set of efficiency metrics and {hardware} counters have been recognized from the profiling instruments, nvprof and Nsight Compute, to assemble a hierarchical Roofline. This helps customers acquire a holistic view of their software and establish probably the most rapid and worthwhile optimizations. The methodology has been validated with purposes from varied domains, together with materials science, mesh and particles-based codes, picture classification and segmentation, and pure language processing.
  • OpenMP Offload PGI compiler: Since early 2019, NERSC workers have been collaborating with NVIDIA engineers to boost NVIDIA’s PGI C, C++, and Fortran compilers to allow OpenMP purposes to run on NVIDIA GPUs and assist customers effectively port appropriate purposes to focus on GPU {hardware} in Perlmutter.
  • Python-based knowledge analytics: NERSC and NVIDIA are creating a set of GPU-based excessive efficiency knowledge analytic instruments utilizing Python, the first language for knowledge analytics at NERSC and a strong platform for machine studying and deep studying libraries.
  • RAPIDS: Utilizing RAPIDS, a set of open-source software program libraries for working knowledge science and analytics pipelines on GPUs, NERSC and NVIDIA engineers are working to know the sorts of points NERSC knowledge customers encounter with GPUs in Python, optimize key parts of the info analytics stack, and train NERSC customers tips on how to use RAPIDS to optimize their workflows on Perlmutter.

“Giving our customers entry to the very newest in GPU-accelerated know-how this 12 months is a vital step in the direction of making certain that our customers stay productive and are capable of make the most of the techniques to arrange for the Exascale period. Our efforts in getting our various person base accustomed to the brand new know-how has been very encouraging and we sit up for Perlmutter delivering a extremely succesful person useful resource for workloads in simulation, studying and knowledge evaluation,” stated Sudip Dosanjh, NERSC Director.

Situated at Lawrence Berkeley Nationwide Laboratory, NERSC is a DOE Workplace of Science person facility.

Source link

Leave a Reply

Your email address will not be published.

Previous Post

A script that detects anomalies at a local level

Next Post

Pay Per Click Ppc Advertising Market Explore Future Growth 2020-2022 by Global Top Players – Cole Reports

Related Posts