From: Jun Zhang (coolrainbow_at_yahoo.cn)
Date: Wed Dec 08 2010 - 05:59:53 CST
I have just compiled CUDA version NAMD 2.7 and have been testing its performence. It does speed up for at least 3.6 times than a non CUDA version on the same machine. However, I have encountered a subtle issue and I cannot explain it:
With the same input files (the same psf, pdb, conf, and force field parameters), the non CUDA version runs well, but the CUDA version, fails to converge for the SHAKE algorithm. Only after I manually set the "rigidInterations" to 2000 did work well. Since it is well known that SHAKE algorithm has very good convergence property (usually less than 20), this phenomenon seems somewhat strange. Does this matter? By this I mean will such issue generate some abnormal trajectory?
I guess this may due to the different algorithms of nonbonded interaction computation on CUDA and nonCUDA version. But I am not sure. Does anyone have some ideas on this problem? Thank you for your help.
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:56:27 CST