OPENRADIOSS PARALELLING

[CinekPoland]

I did some simulations of bird strike https://openradioss.atlassian.net/wiki/spaces/OPENRADIOSS/pages/8847361/Bird+Strike+On+Windshield
using 2 , 8, 16 and also 32 cores. scalability wasn't very good. Using 16 cores I achieved about 4.5 speedup są it is not satisfactory. Doeas anyone know it is possible to improve scalability ? Maybe scalability could be better if simulations will be bigger ? Doeas anyone knows where I can find sample of simulation which contain 300 000 - 1000 000 elements ?

[laurent-altr]

Hi @CinekPoland,
A couple of remarks:

• With OpenRadioss we usually measure the speed-up by computer node, and not by core. When you go from 2 to 16 cores on the same machine, you have not multiplied the resources by 8. For instance, it is very much possible that cores are sharing some cache memory and memory bandwidth, that are critical for OpenRadioss.
• Sph implementations known to have a limited scalability with OpenRadioss.

I'll let others (@PaulAltair, @MarianBulla ) reply on the other models you are looking for.

Thanks

[CinekPoland]

Thank you for your answer. Could you explain exactly what you mean "cores are sharing some cache memory and memory bandwidth, that are critical for OpenRadioss."

So it is not good when cores sharing cache memory and bandwidth ?

[laurent-altr]

I mean that OpenRadioss is memory bound.

For example, let's say that we have a CPU with two cores sharing the L2 cache. When you are running on 1 core you are already using all the L2 available. In that condition, you cannot expect a 2x speedup going from one to two cores.

[CinekPoland]

Do you think better scalability could be gained if each processor core is assigned a fixed value of cache and ram ?

I thought that if you set, for example, 16 cores to perform a given task, then each core will be assigned a specific value of the amount of cache memory, and not that one core takes up the entire cache value.