6 Myths of High-Performance Computing

  • CategoryEngineering

  • View385

Report
  • 1. Advances have made tasks that were previously solvable only by a supercomputer tractable for SMALLER COMPUTERS. CFD simulations can clearly scale up to higher core counts than structural mechanics or electromagnetics simulations because of the nature of the underlying numerical algorithms. But this does not mean that structural and electromagnetics simulations are not amenable to HPC solutions. In a recent ANSYS survey, about ONE THIRD OF THE ALMOST 3,000 RESPONDENTS said that they limit the size or amount of detail for nearly every model they run due to compute capacity and/or turnaround time limitations. An additional 57 PERCENT OF RESPONDENTS said they impose these limits on at least some of their models. Twenty years ago, HIGH-PERFORMANCE COMPUTING (HPC) was primarily used by the larger enterprise organizations for engineering simulations. HPC is now gradually becoming more common in smaller and mid-sized companies as well. Nevertheless, many misconceptions still surround HPC, preventing its adoption and/or growth in product development scenarios where it could clearly provide a higher return on investment. Years ago, only SUPERCOMPUTERS had enough computing power and capacity to perform even routine simulations. HIGH-PERFORMANCE COMPUTING IS AVAILABLE ON SUPERCOMPUTERS ONLY myth #1 SOLUTION: ANSYS has teamed up with HPC strategic partners to make specification and deployment of HPC easier for you. PROBLEM: Lack of expertise and time to specify hardware configurations is a leading barrier. WITHOUT INTERNAL IT SUPPORT, HPC CLUSTER ADOPTION IS UNDOABLE myth #4 AVERAGE OF 10 DIFFERENT ANSYS MECHANICAL BENCHMARKS HPC IS USEFUL ONLY FOR CFD SIMULATIONS myth #2 I DON’T NEED HPC— MY JOB IS RUNNING FAST ENOUGH myth #3 MORE COMPUTING CORES PER CPU INTEGRATED I/O ON A PROCESSOR DIE (YIELDING HIGHER MEMORY BANDWIDTH) MORE AND FASTER MEMORY CHANNELS LARGER L3 CACHE SIZE, FASTER DISK STORAGE FASTER INTERCONNECTS AVX SUPPORT SOLVERSPEEDUP 2 cores (1 node) 16 cores (1 node) 32 cores (2 nodes) 64 cores (4 nodes) 128 cores (8 nodes) Continually Improving Core Solver Rating To 128 Cores FREQUENCY OF LIMITING SIZE/DETAIL IN SIMULATION MODELS DUE TO COMPUTER INFRASTRUCTURE OR TURNAROUND TIME LIMITATIONS 57% 34% 9% For some models Nearly every model Almost never At ANSYS, an intense focus on HPC software development has produced capabilities that set us apart and enabled breakthrough productivity on current and emerging hardware solutions. If your CFD software can only scale down to about 100,000 cells per core, then a typical two million cell model can only run on 20 cores efficiently. Having more cores available is useless in this case because they do not increase the simulation speed. However, if the CFD software can scale to about 5,500 cells per core (like ANSYS Fluent) the same two million cell model can then run on 360 cores, resulting in an almost 18-fold faster turnaround! PARALLEL SCALABILITY IS ALL ABOUT THE SAME, RIGHT? myth #5 The key is that the cost of HPC is matched to value — and HPC offers one of the greatest returns on investment possible. ANSYS HPC software licensing is designed on pricing models that ensure the highest value for engineering simulation workloads while allowing ANSYS to continue our HPC software developments. HPC SOFTWARE AND HARDWARE ARE RELATIVELY EXPENSIVE myth #6 WRONG AVX VALUE PROPOSITION OF HPC ENHANCES ENGINEERING PRODUCTIVITY BY ACCELERATING SIMULATION THROUGHOUT ENABLES ENGINEERS TO CONSIDER MORE DESIGN IDEAS AND MAKE EFFICIENT PRODUCT DEVELOPMENT DECISIONS ALLOWS ENGINEERS TO SIMULATE LARGER, MORE COMPLEX MODELS SO THAT MORE ACCURATE DESIGN DECISIONS CAN BE MADE CUMULATIVETIMESAVINGSUSING MULTIPLEMULTI-COREJOBS LICENSEFEEPERMULTIPLE MULTI-COREJOB ABSOLUTETIMESAVINGSPERJOB LICENSEFEEPERCOREPERJOB NUMBER OF CORES NUMBER OF CORES NUMBER OF CORES NUMBER OF CORES Idealscalability HPC Pack HPC Workgroup to read the full white paper visit ansys.com/hpc-myths Brought to you by Demonstrated scalability of ANSYS Fluent above 80 percent efficiency with as low as 5,500 cells per compute core Speedup Ideal 11.1K cells/core 7.4K cells/core 5.5K cells/core 4,096 8,192 12,288 16,384 16,000 14,000 12,000 10,000 8,000 6,000 4,000 2,000 0
  • Description
    1. Advances have made tasks that were previously solvable only by a supercomputer tractable for SMALLER COMPUTERS. CFD simulations can clearly scale up to higher core counts…