Hi Jared, That I am not an expert. The hard part is still that this is a throughput question: Although when using CPU it is more straightforward to run multiple instances (like launching a series of jobs on a server CPU, each using a core and a given amount of RAM), is it better than running jobs in serial but each with many cores?
I think it's quite hardware-dependent. I guess the former is more scalable on server CPUs assuming you don't have that many reads and writes, and could have a better throughput than the latter. You just have to deal with the fact that you'll have to wait long before you get each batch of results, considering terramechanics problems can be large in size. This is worth testing out too. Thank you, Ruochun On Tue, Jun 4, 2024 at 7:19 AM Jared Long-Fox <[email protected]> wrote: > Hi Ruochun, > > Thank you very much for the speedy and informative response. If/when I end > up trying this, I will post any findings here. Apologies for another > question, but has there been any parallelization testing on CPUs? If so, > what were the results? Thanks in advance! > > > All the best, > Jared > > On Thu, May 30, 2024 at 9:31 PM Ruochun Zhang <[email protected]> wrote: > >> Hi, >> >> Since you mentioned the GPUs on your desktop, I assume it's about the >> GPU-based Chrono DEM tools. You should be able to simply run a couple of >> them on parallel, but truthfully, about the throughput impact, we never >> tested. >> >> My guess would be that on consumer GPUs, running 2 or 3 instances >> together may slightly improve the throughput. On data center GPUs, running >> a small number of instances may moderately improve the throughput. Again, >> this is a guess and I am nore than interested in knowing your findings >> should you test it out. >> >> Thank you, >> Ruochun >> >> On Thu, May 30, 2024, 7:16 AM Jared Long-Fox <[email protected]> >> wrote: >> >>> Hello all! I am working on a project in which I need to calibrate a >>> model to experimental data. To do this, I need to run a parameter >>> estimation algorithm such as Monte Carlo or a grid search to sweep across >>> parameter values to determine which model parameters give the best fit to >>> the data. >>> >>> However, such calibrations take hundreds to thousands of model runs. >>> There are ways to reduce this, but I was originally planning on running >>> multiple DEM models in parallel instead of in serial to speed the process >>> up. Is it possible to run more than one Chrono model at once on a single >>> machine such as my desktop workstation (assuming sufficient hardware, e.g., >>> 2x NVIDIA 3090 GPUs)? If so, how much performance is sacrificed (if any)? >>> >>> Thank you all for the time, consideration, and insights on this! >>> >>> >>> All the best, >>> Jared Long-Fox >>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "ProjectChrono" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to [email protected]. >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/projectchrono/730e41ac-b7d2-40cd-96da-a5472f9cbc1cn%40googlegroups.com >>> <https://groups.google.com/d/msgid/projectchrono/730e41ac-b7d2-40cd-96da-a5472f9cbc1cn%40googlegroups.com?utm_medium=email&utm_source=footer> >>> . >>> >> -- Ruochun Zhang Email: [email protected] Email: [email protected] <[email protected]> Tel: 832-353-5111 -- You received this message because you are subscribed to the Google Groups "ProjectChrono" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/projectchrono/CAHvQpOtcPoJ6cNP4kULHaH7A_Rm4xFV%3DNjqsGtM8EGXfSjBmGw%40mail.gmail.com.
