--- Home --- About Us -- ATMET LLC -- Services -- Contacts Distributions -- Software -- Data Support -- Documents -- Forums -- Workshops RAMS Links -- Operational -- Research Problems or comments about the site?

ATMET MM5 Efficiency tests Following are the results from a series of MM5 efficiency tests done at ATMET. These tests were done on a PC Beowulf cluster of 7 1.4 ghz athlon dual-processor machines. Some tests are for different options and some are for different numbers of processors. All are for 1 hour of simulation time, 108 sec coarse grid timestep, disregarding first and last timesteps. All tests were done with MM5 version 3-5. Figure 1: Seconds/timestep (coarse grid timestep) vs. number of processors. Results are plotted for 2, 3, and 4 grid simulations. Also plotted are results from an 8-processor, 3-grid simulation, varying some of the model options and parameterizations. Figure 2: Close-up of Figure 1 to show relative performance of different model options in the 8-processor, 3-grid test. Plot shows seconds/timestep (coarse grid timestep) vs. number of processors. Figure 3: Efficiency vs. number of processors. Results are plotted for 2, 3, and 4 grid simulations. Efficiency is defined as (stime/ptime)/pnum, where stime is the time for a single processor simulation, ptime is the time for the multi-processor simulation, and pnum is the number of processors. Figure 4: Speed-up vs. number of processors. Results are plotted for 2, 3, and 4 grid simulations. Speed-up is defined as stime/ptime, where stime is the time for a single processor simulation and ptime is the time for the multi-processor simulatiion. Return to ATMET MM5 Information page Return to ATMET home