Many consumers immediately have fashions that run nicely on smaller core counts, or possibly license sure to a selected core rely. Because of the great amount of CPUS obtainable in a single system (56 or extra), shoppers have expressed curiosity in “stacking” a number of smaller jobs on a single server to extend throughput.
Fashionable HPC system constructing blocks primarily based on third Technology Intel Xeon Scalable Processors have excessive reminiscence bandwidth per core, and improved AVX512 and Turbo choices which have made stacking much less of a efficiency penalty than up to now.
To see the results of stacking on the most recent third Technology Intel Xeon Scalable Processors, we used the Abaqus e13 benchmark on an Intel 6346 primarily based node that has 32 complete cores per server. The baseline run was at 16 cores Completely (out of 32, which advantages from turbo results), after which “stacked” two jobs on the identical node, consuming all 32 cores (two 16-core jobs).
From the outcomes under, you possibly can see that one can stack two an identical jobs with a modest drop in general single node job time, however at the next throughput. For those who want absolutely the quickest job time for a hard and fast token price, then working unique is the most effective, however stacking jobs offers you extra throughput general for small jobs.
By default, the TotalCAE Platform disables job stacking and runs jobs in “Unique” mode, which suggests even in case you use 1 CPU, you get the entire machine to your self. This ensures your job runs at peak efficiency, and that different jobs don’t intrude together with your job in some damaging method, for instance working the machine out of reminiscence or in any other case impacting your job. To disable this default, the person can simply uncheck the Unique field on job submission or use the -x possibility in our command line tsubmit, to permit jobs to “stack” on the identical node and experiment with the way it works on your specific mannequin.
For extra data or to see a demo, contact TotalCAE at https://www.totalcae.com/contact.php
Copyright © 2022 IDG Communications, Inc.