Page 1 of 1

how CULA uses GPU?

PostPosted: Mon May 20, 2013 4:37 am
by goryales
I use culaCgesv in my project, and performance isn't so good how i expected (solving of complex system 6000X6000 takes 14sec.). So I launched application in Compute Visual Profiler, and got next results:
Kernel time = 4.31 % of total GPU time
Memory copy time = 1.3 % of total GPU time
But i know that solving of the system (culaCgesv) takes half of time of execution.
How it can be?

Re: how CULA uses GPU?

PostPosted: Mon May 20, 2013 6:17 am
by john
What is your GPU? a 6000 cgesv should take nowhere near 14 seconds.

Are you certain you've initialized your CUDA runtime before calling CULA? If you haven't, then you'll pay the startup cost in the first CUDA call.

With the profiler, keep in mind that it will reduce your performance further.

It's hard to comment further without specifics.

Re: how CULA uses GPU?

PostPosted: Mon May 20, 2013 7:07 pm
by goryales
GPU -Tesla C2050
Are you mean culaInitialize()?
I use CULA like in gesv examples:
culaInitialize()
culaCgesv(...)
culaShutdown()
But in my project I use culaCgesv 144 times.
And I got next results for different size fo systems:
Size of system solv time
104 0.0-0.3 s
504 0.1-0.3 s
992 0.1-0.8 s
1496 0.1-2.1s
3000 0.6-5.3s
6008 2.9-17.8s

Re: how CULA uses GPU?

PostPosted: Mon May 20, 2013 7:21 pm
by goryales
Also I have version of application with Intel MKL.
And results for sequential (non-parallel) LAPACKE_cgesv from MKL on Xeon e5606:
ize of system solv time
104 less than 0.1
504 0.0-0.1 s
992 0.2-0.3 s
1496 0.6-0.7s
3000 5.1-5.3s
6008 38.5-39s

Re: how CULA uses GPU?

PostPosted: Tue May 21, 2013 6:51 am
by john
Can you post a reproducer code?

Re: how CULA uses GPU?

PostPosted: Tue May 21, 2013 8:48 pm
by goryales
No, unfortunately I can not