G-RSM benchmark test Part 2


4. GSM MPI Wall Clock Time Comparison on COMPAS Linux cluster (SIO) - older processors.


Wall clock time required to run 24 hour forecasts (Global model with 28 levels in the vertical)

200 km resolution
Speed-up
1/(NPE ratio)

T62L28 16PEs
80sec
0.45
0.27

T62L28 32PEs
49sec
0.73
0.53

T62L28 60PEs
36sec
1
1

T62L28 120PEs
40sec
1.06
0.5
slowed down!
T62L28 186PEs
44sec
1.26
0.32
slowed down!
T62L28 252PEs
47sec
1.34
0.24
slowed down!

100km resolution




T126L28 60PEs
218sec (3min 38sec)
1
1

T126L28 120PEs
128sec (2min 08sec)
0.59
0.5
very good
T126L28 186PEs
116sec (1min 56sec)
0.53
0.32
poor
T126L28 252PEs
140sec (2min 20sec)
0.64
0.24
slowed down!

50km resolution




T248L28 60PEs




T248L28 120PEs
2100sec (35min 00sec)
1
1

T248L28 186PEs
1515sec (25min 15sec)
0.72
0.65
very good
T248L28 252PEs
970sec (16min 10sec)
0.46
0.48
excellent

25km resolution (obtained from running 6-hour forecasts)


T496L28 60PEs




T496L28 120PEs
28624sec (7hours 57min 04sec)
1
1

T496L28 186PEs
17272sec (4hours 47min 52sec)
0.6
0.65
excellent
T496L28 252PEs
13188sec (3hours 39min 48sec)
0.46
0.48
excellent