Slow tests sometimes
Posted: Mon Jul 29, 2013 11:56 am
I have a problem with test code sometimes running the test code testing_sgetri_gpu with increasingly larger matrixes in -c (check) mode.
When I get to a sufficiently large matrix, the test program fails with "!!!! magma_malloc failed for: d_A". That makes perfect sense as I don't have enough memory on the GPU for the matrix. But sometimes the performance of the test code is about half its normal performance, as shown below for one case. I am wondering if the crashed test program sometimes leaves the GPU in a partially cleaned up state, so it doesn't have the compute resources for later tests. The slowness problem goes away only after I reboot my machine.
Example: 285.30 GFlops/s version 123.98:
versus:
Is this a known problem?
Is there a way to clean up the GPU's state short of a reboot?
I am running on an iMac 27" 2012 (OS X 10.8.4) with its NVIDIA GeForce GTX 675MX and an Intel Core i5 at 3.2 GHz with 24GB of RAM.
I am using the CUDA 5.5 release candidate and the magma-1.4.0-beta2.
thanks.
When I get to a sufficiently large matrix, the test program fails with "!!!! magma_malloc failed for: d_A". That makes perfect sense as I don't have enough memory on the GPU for the matrix. But sometimes the performance of the test code is about half its normal performance, as shown below for one case. I am wondering if the crashed test program sometimes leaves the GPU in a partially cleaned up state, so it doesn't have the compute resources for later tests. The slowness problem goes away only after I reboot my machine.
Example: 285.30 GFlops/s version 123.98:
Code: Select all
N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F
5000 135.72 ( 1.23) 285.30 ( 0.58) -inf
Code: Select all
N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F
5000 136.23 ( 1.22) 123.98 ( 1.34) -inf
Is there a way to clean up the GPU's state short of a reboot?
I am running on an iMac 27" 2012 (OS X 10.8.4) with its NVIDIA GeForce GTX 675MX and an Intel Core i5 at 3.2 GHz with 24GB of RAM.
I am using the CUDA 5.5 release candidate and the magma-1.4.0-beta2.
thanks.