GPU Factorisation Solve

[tyronerees]

We were able to run SSIDS (v2.0) in GPU in the analyze phase. But for factorization phase and for solve phase, we were not able to use GPU yet. Please, how can we enable SSIDS to perform the factorization phase and the solve phase also in GPU?

(submitted via hsl@stfc.ac.uk)

[tyronerees]

Closing the issue, as this behaviour was expected due to the size of the problem and the (default) options set.

[Poofee]

What does it mean? Can we use GPU in factorization phase and for solve phase?

[tyronerees]

Yes, this should work. If you're not finding that it does, please reopen the issue and we can investigate further. The code is set up so that the GPU is only used if it's worthwhile to do so, so if you run on small problems the GPU may not be used, as the overhead of sending to the GPU is too expensive - this is (we think) what the user was experiencing. This can be controlled somewhat (but not completely) by changing the options min_gpu_work and gpu_perf_coeff.

[Poofee]

Ok, I will have a try. If I have some problem, I will leave it here.

[flipflapflop]

If you want more information about the options controlling the use of the GPU you can take a look at the following section in the documentation ssids_options.

Essentially, before the factorization phase, the workload is split between the CPUs and the GPU using the parameter gpu_perf_coeff which corresponds to of how much faster the GPU runs compare a CPU (including all the cores). Then the parts of the workload that are supposed to be run on the GPU are only processed on the GPU if the corresponding workload exceed min_gpu_work.

[Poofee]

Thesedays, I was using the ssids. However, I am still confused about the GPU setting. How can I set GPU threads in details?