Advanced Computing in the Age of AI | Sunday, September 24, 2023

New Commercial CAE Code Scaling Record Claimed 

SGI, the old-line supercomputing vendor recently acquired by HPE, has teamed with ANSYS, the product engineering and simulation software company, to set a new world record for scaling commercial CAE code. According to SGI, the two companies broke a record set last year by running ANSYS Fluent combustion modeling software across 145,000 CPU cores, exceeding by more than 16,000 the old record.

The announcement was made this week in a blog post by Tony DeVarco, director of virtual product development manufacturing solutions at SGI. Extreme scaling of simulation code is a critical issue for manufacturers using HPC-class systems to leverage maximum performance from modeling software.

“SGI was able to run the ANSYS-provided 830 million cell gas combustor model from 1,296 to 145,152 CPU cores,” DeVarco said. “This reduces the total solver wall clock time to run a single simulation from 20 minutes for 1,296 cores to a mere 13 seconds using 145,152 cores and achieving an overall scaling efficiency of 83 percent.”

ansys-combustor-benchmarkThis is the latest in a series of CFD benchmarking projects the two companies have undertaken on a joint basis. The first took place more than a year ago, according to DeVarco, when together they demonstrated that the SGI UV in-memory computing platform and ANSYS HFSS software could solve large, high frequency electromagnetics problems, such as cosite analysis and radar cross section analysis, “as well as allow multiple frequency sweeps to be run without running out of computer system memory.”

Last November, ANSYS approached SGI to work on demonstrating ANSYS Maxwell’s Time Decomposition Method, which allows the simultaneous solving of all the steps involved in a low frequency electromagnetic problem, according to DeVarco.

He said that ANSYS 17.0, released earlier this year, drew the attention of SGI engineer because of the ANSYS Mechanical package’s ability to scale up to 1,000 cores. “Our applications engineer who works closely with our structural mechanics customers took the V17 sp-5 Model and scaled it to 1,008 cores,” DeVarco said.

SGI's Tony DeVarco

SGI's Tony DeVarco

Then last month, ANSYS and SGI application engineers worked together to achieve a new world record for scaling ANSYS Fluent on a SGI ICE XA system, one of the world’s fastest commercial distributed memory supercomputers. DeVarco said the benchmark came about when members of the SGI manufacturing team had built a new system for the National Center for Atmospheric Research. As the system went through testing the SGI engineers wanted to run a large benchmark of a commercial code on the full system. “ANSYS Fluent fit the bill,” he said.

DeVarco added that the Fluent benchmark was achieved using SGI MPI PerfBoost, which is designed to allow technical applications written for other MPI implementations to leverage SGI Message Passage Toolkit (MPI) at runtime without recompiling.

SGI’s other highlight, the SPECrate record, measures the throughput or capacity of a machine to carry out a number of simultaneous tasks. “They are based on real world application codes and are highly valued as representative workloads for large, multi-processor systems,” wrote Broner.

SPECrate FP Metric – Floating Point Application Throughput

SPECrate Integer Metric – Integer Application Throughput

JAIST was founded in October 1990 as the first independent national graduate school in Japan, to carry out graduate education based on research at the highest level in advanced science and technology. The JAIST SGI UV 3000 is a single shared memory system with 256 Intel Xeon E5-4655 v3 processors (1536 cores) and 32 terabytes of memory. The JAIST SGI UV 3000 high-performance system supports researchers and developers in various areas including large-scale simulation, machine learning, and high-speed algorithm research.