Cusp low GPU utilization

[lesgidis]

Hello
I have just installed the Nvidia cusp library in my notebook as i need some increase in my performance in comparison to the CPU opensees solvers, however i am having some trouble in fully utilizing my GPU. In more details through a simple nonlinear bridge example of 300-400 nodes i get the following report from nvidia-smi with GPU-utilization not surpassing 16-17%. Through different programs which use my GPU i am capable of achieving up to 100% utilization of my GPU, so the low performance is only indicated in my opensees analysis runs. Finally the cusp analysis is much much slower than the sequential opensees analysis with the SuperLU solver. Could this be a solvable issue with opensees (i have tried different options with no sufficient results) or could be utilization percentage be dependent to the quality of the GPU?

Thank you for your time

| NVIDIA-SMI 382.05 Driver Version: 382.05 |
|===============================+======================+======================|
| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce GTX 950M WDDM | 0000:01:00.0 Off | N/A |
| N/A 51C P0 N/A / N/A | 61MiB / 2048MiB | 16% Default |
+-------------------------------+----------------------+----------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 17804 C ...RARIES\OpenSees Cuda\EXAMPLE\OpenSees.exe N/A |

[fmk]

the problem with using the GPU to solve is time spent transferring the data to gpu and back again from the main memory with the GPU having to wait additionally for the code on the core processor to loop over elements, do state determination and form the equation in the first place. you might actually muck about in code abit and put timers around the different parts of the code to see what % of time is actually spent in solvers versus state determination. depending on the model the time spent in the solving can actually be a small % of overall time and GPU's not going to help unless other parts of code are ported to the GPU.

[shahabshamlou]

Hi
I have this problem.
My nvidia graphic card is k4000m with 960 cuda core.
Model is a 4 span bridge with 300 nodes. Some parameters are variable.
I need to run ida analysis. But it is very slow.
Slower than a simple analysis on cpu.
In my case Cpu is 5 time more faster than gpu.
Is it normal?
Where is problem?
I guess something is wrong.
Maybe in install cuda or cusp library or ...
Please help me.

[fmk]

read the above post as to what part of the calculation is being done on the GPU.