From: Joakim Swedberg (joakim_swedberg_at_hotmail.com)
Date: Thu Nov 12 2009 - 20:07:23 CST
NAMD 2.7b1 Performance loss with CUDA 2.3
I'm currently running NAMD 2.7b1 compiled from source (net-linux-x86_64-g++ --with-cuda) and CUDA 2.3 in Linux (Sues 11.1) with an i7 (4GHz) and 2 GTX295 GPU cards. However, I am constantly getting 20% performance loss in NAMD when using CUDA as compared with running it on only the CPU (using the apoa1 benchmark). Reading all other related posts in this mailing list, I've tried all possible CPU GPU ratios as well as increasing the cutoff and pairlistdist to 20 and 21.5 respectively to put more load on the GPUs, but with the same results. NAMD says that CUDA execute on the proper devices and the corresponding GPU temperature increases about 15C when NAMD is running with CUDA. All CUDA tests have been performed and it seems to be installed properly as other CUDA applications perform well. I execute NAMD as example:
./charmrun ++local +p4 ./namd2 +device 0,1,2,3 apoa1/apoa1.namd
But as I said, I've tried many different configurations as +p2 and +device 1,3, +p8 and +device 0,1,2,3 etc. Interestingly +p8 and +device 0,1,2,3 has the least performance loss (octa threaded quad core?) but its marginal and always around 20%.
Have anybody had similar experiences? I would be greatful for any suggestions.
Institute of Health and Biomedical Innovation (IHBI)
60 Musk Avenue
Phone +61 (0)7 3138 6222
Mobile +61 (0)413 165 805
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:53:29 CST