By Taner Pirim

The optimized version of Amber ver.8 is now installed on redwood at /usr/local/appl/Amber8 by using Intel Math Kernel Library, MKL, version 7.2.1 as well as Amber ver. 7. Also, Amber 8 is installed on mimosa cluster. The performances of Amber 8 and Amber 7 on Redwood and Mimosa have been analyzed for various numbers of cpus together with Dr. Randy M. Wadkins. The results obtained from the benchmark sets were compared. Amber 8/7 benchmarks were ran for 2, 4, 8, 16, and 32 cpus and the wallclock times obtained from the output files.

As shown in the figure above, Amber 7 and 8 benchmarks on Redwood outperformed the benchmark results of Amber 8 on mimosa for all the cpu parallelization. Redwood finished the runs at least 2.5 times faster than mimosa. On less than 8 cpu jobs, redwood completed the runs an average of approximately five times faster than mimosa. Also, Amber 8 performed slightly better than Amber 7 on redwood for all the runs.

            When the speed-up for each run calculated by dividing the runtimes of Redwood to Mimosa runtimes, the speed-ups for each run were better when the numbers of cpus were increased as shown in above Table. Thus, when the speed-ups of 8 and 16 cpu runs were compared, the 16 cpu jobs completed the runs in half the time than 8 cpu jobs. However, 32 cpu speed-ups showed an average of 11% improvement than 16 cpu speed-ups even though the number of cpus used were doubled. Thus, for the benchmarks the best efficiency were obtained by running the benchmarks on 16 processors. If you would like more information on the versions of Amber installed on our systems, please visit our Amber website. For more questions on Amber, please contact us via assist@mcsr.olemiss.edu.