Discussion Forum

Dear BP,

I think that we should be happy about the performance improvement. It was not an easy thing for us to fix.

The reason for the dramatic improvement is that we are no longer using Windows' built in memory allocator for computers with more than 8 cores. Instead, we are using a scalable memory allocator adapted for modern computers with many cores and several CPU sockets. The new memory allocator uses slightly more memory than the one included in Windows' but yields better performance, as you pointed out.

Best Regards, Ed Fontes COMSOL

Hi Ed,

I completely agree. (I was being sarcastic.) Time is money!

Cheers, Brennan

Hi Ed,

Can you comment on ways to best take advantage of this memory allocator? I'm not complaining, but I don't think I'm seeing anywhere near the "7X faster" improvement that BP noted above. But then again, I was already seeing pretty good utilization, or so I thought, of my 36 cores (18 per CPU) on a Dell T7910 Windows 10 based, 64-bit workstation with 256 GB of RAM. I tend to run a lot of large 3D RF models and typically use the PARDISO solver with linear discretization. Do you expect that this should be benefitting from the improvements you mentioned, in Comsol version 5.4? Would you possibly now recommend a different approach in regard to solver (or other) choices, in regard to optimizing speed of computation and efficient use of memory? I know the iterative solvers use less memory, but they also tend to be less stable (especially BiCGSTab) for the kinds of problems I do. When I run into memory barriers, I usually switch to GMRES and the SSOR Vector preconditioner, which is usually stable, but is also pretty slow. Thanks in advance for any suggestions.

Hi Robert,

I'm sure Ed has more to say on this, but be sure to go to preferences/ Multicore and Cluster Computing / Memory allocator set to "Optimized for Multicore".