G-RSM benchmark test


1. MPI Wall Clock Time comparison on IBM-SP Datastar (SDSC)


GSM (T62L28 72-hr), 192*94=18048 grid cells

64CPUs
128CPUs
256CPUs
psplit
22.0sec
19.0sec
23.4sec

GSM (T248L28 72-hr)

64CPUs
128CPUs
256CPUs
psplit
177.8 sec



RSM (128x199 10km 12-hour) 128*199=25472 grid cells

128CPUs
256CPUs
psplit
93.1sec
72.8sec

RSM (512x335 10km 6-hour) 512*335=171520 grid cells

512CPUs
1024CPUs
psplit
181.9sec
338.5sec

Ratio of the number of computational operations between RSM and GSM:
25472/18048x1800/30=1.411x60=85

Ratio of computation time from above table:
91.4 *6 / 19 = 29

This means RSM is 3 times more efficient???



2. MPI Wall Clock Time comparison on COMPAS Linux cluster (SIO)


RSM (128x199 10km 12-hour) 128*199=25472 grid cells

12-hr
24-hr
36-hr
48-hr
psplit_54
473.8
475.1
458.6
468.5
(roughly 2 times slower than IBM-SP)



3. MPI Wall Clock Time comparison on IBM-SP Bluesky (NCAR)


RSM (160x199 10km)

pspl_128: 216 sec/12-hour

Compare to 116.25sec on Datastar (SDSC).



See more comparison:

4. GSM MPI Wall Clock Time Comparison on COMPAS Linux cluster (SIO) - older processors.
5. RSM on IBM Power 4 graphs.
6. RSM on IBM machines at NCAR.