AndyJP
Subscriber
>so it's not like the cores are going unused. nThat's a typical pattern when some auxiliary (usually data i/o) uni-threaded code is distributed by Windows.nThe fully-loaded cores perform the heavy computation. And here we see that the new code in 2020R1 and 2021R1 is NOT paralleled well. I.e. the licenses and processor power are not used effectively and for good.nnEither it is a compiler configuration bug, or a design flaw in your updated code. I know that you concentrate on paralleling the frequency sweeps and parametrics.nBut it is not the heaviest part of the job. Big anisotropic models take awfully lot of time just to reach the point of sweep, and it is performed in ONE SINGLE task. Removing parallelization there throws us 10 years back to HFSS v.14.nnnBTW, it would be fine if there are 2-core 8GHz CPUs with 128GB ECC ram support available for building a fast sweeping cluster. But there are no such. So you should improve the single-threaded parallelization.n