More cores always better?

[firesim1990]

Hello

I am looking to buy a working station that is capable of calculating a maximum of 16 meshes with 10 million cells (in a reasonable amount of time). If I understand correctly, a cluster is not a good option, since the overall performance is lower (due to the slower communication between several computers instead of cores).

Would a CPU with 16 cores be enough or could additional cores (for example, AMD Threadrippers with 32 or 64 cores) bring more performance? Of course, taking into account that at least 2 GB of RAM for each core is ensured. Is it possible to employ several cores to calculate one mesh? What would be the best strategy when it comes to hardware for FDS?

Thanks.

[Kevin McGrattan]

You do not understand it correctly. The best platform for FDS computations is a linux compute cluster. Each node in the cluster will certainly have multiple cores, say 8, 12, or 16, but as the number of cores per CPU socket increases, performance tends to decrease because you hit a bottleneck accessing RAM. 

If you only have funds for a single computer, then you should maximize the number of sockets and cores per socket. We have found hyper-threading to be ineffective.

[firesim1990]

Thanks. In that case, if I am to buy one single computer only, I should aim for more cores. For example, if I have 64 cores and 16 meshes, I would employ 4 cores to each mesh? Is it really necessary to go to 32 or 64 cores, or 8 or 16 cores would work as well? Is there a decrease in the performance related to the number of cores?

[Kevin McGrattan]

It's not just a question of cores, but rather cores and sockets. I would read up on the architectures mentioned in this figure.

[firesim1990]

I am not even close to these maximum values I see in the table. My questions was more related to the difference between lets say a computer with 16 cores and 32 GB RAM and 64 cores and 128 GB of RAM, if I want to run the model consisting of 16 meshes. What is the advantage of having so many more cores if the number of meshes is only 16? Pr is it possible (and desirable) to assign more processes to the same mesh? Thanks.

[firesim1990]

I tried running few cases in FDS and I guess one process can be dedicated to one or more meshes, but one mesh can not be assigned to more processes. Again, if more cores are available, would be wise to divide the model into smaller meshes so each mesh could be calculated using separate process?

[dr_jfloyd]

More memory means you can have more grid cells. 32 Gb may not always be enough for 10 million cells depending on the specifics of the input.

If you had 16 cores and 10,000,000 grid cells, you would have you would have 625,000 grid cells per core. If you had 64 cores, then you could take each mesh and divide it in 4 giving you 156,250 cells per core and run your case four times as fast. You could also run four different 16 core jobs. You could also keep 16 meshes and use 4 OPEN MP threads per FDS process but this would buy you no more than a factor of two improvement (i.e. the better option is just having more meshes).

[Rob]

I have heard from people who work at other firms, that they bought second hand servers cheap and used them for FDS, they were pleased with the performance, especially for the price. But they did have to put them in another room as they were very loud.

Also, I did a very quick search and found the following company (Europe). The starting prices are comparable to what I paid for a high performance workstation, but with a lot more cores, albeit not so fast per core but still probably faster. Send them an email.

Happyware HPC clusters.

 I bought some desktops with 5 ghz i7s (8 cores) with all sockets full of RAM, but in retrospect, I regret not investigating the options above. Of course you would have to do your homework on server performance and what you need, run a few benchmarks etc.

[firesim1990]

Thank you for your answers. Since I plan to invest in one workstation, I will try to get the most powerful one for the money. What is the difference between running a workstation on Windows or Linux? I am more familiar with Windows platform.

[Glenn Forney]

scripting is better with Linux, ie the ability to specify and control what programs you run and when.  This will be more important if you plan on accessing this workstation remotely.  Note, if you choose one OS and decide you don't like it for whatever reason, you can always wipe the slate clean and load the other OS.

--
You received this message because you are subscribed to the Google Groups "FDS and Smokeview Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to fds-smv+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/fds-smv/5719ab05-5732-453d-a650-ed43067fc374n%40googlegroups.com.

--

Glenn Forney

[Bryan Klein]

There is also this option for handling a lot of compute load that you can start jobs on directly from PyroSim (if you use that).

https://cloudhpc.cloud/

You get 300 CPU Hours for free to see if it works for you or not.

There is also Sabalcore, but we haven't yet been able to send jobs to it from PyroSim. https://www.sabalcore.com/vertical-markets/engineering/fds-smv/

Cloud computing is a great option when you have large and infrequent spikes in demand.  It also provides capacity for sensitivity studies with multiple large jobs running at the same time.

Best Regards,

Bryan Klein