libbem
Presented are the times for building the panel-clustering-operator and for
solving the system via GMRES. The speedup is calculated by adding both
times.
An overview of the results is given at the and of this page.
Sun E6000
| level |
time for PCop |
time for GMRES |
speedup |
memory |
| 5 |
146.6 |
49.4 |
1.0 |
412 |
| 6 |
707.9 |
241.0 |
1.0 |
1846 |
| 7 |
2604.3 |
1162.6 |
1.0 |
8185 |
| 8 |
9581.0 |
5522.4 |
1.0 |
36300 |
Sun SunFire 6800
| level |
time for PCop |
time for GMRES |
speedup |
memory |
| 5 |
43.5 |
25.1 |
2.86 |
412 |
| 6 |
215.3 |
126.2 |
2.78 |
1846 |
| 7 |
768.9 |
608.1 |
2.74 |
8185 |
| 8 |
3232.8 |
3077.3 |
2.39 |
36122 |
Sun SunFire 15k
| level |
time for PCop |
time for GMRES |
speedup |
memory |
| 5 |
44.0 |
26.3 |
2.79 |
412 |
| 6 |
219.3 |
136.6 |
2.67 |
1859 |
| 7 |
784.8 |
688.9 |
2.56 |
8185 |
| 8 |
3412.5 |
4717.9 |
1.86 |
36122 |
HP 9000 Superdome
| level |
time for PCop |
time for GMRES |
speedup |
memory |
| 5 |
50.0 |
20.1 |
2.80 |
- |
| 6 |
210.4 |
99.5 |
3.06 |
- |
| 7 |
2136.4 |
502.7 |
1.43 |
- |
| 8 |
3665.0 |
2280.8 |
2.54 |
- |
| |
- |
|
From 16k to 65k unknowns one can see a dramatic increase in the time to build
the panel-clustering operator. In contrast to this, the floating-point
performance scaled rather well, as one can see at the GMRES-times.
|
|
IBM eServer p690
| level |
time for PCop |
time for GMRES |
speedup |
memory |
| 5 |
50.7 |
13.1 |
3.07 |
412 |
| 6 |
216.3 |
65.1 |
3.37 |
1833 |
| 7 |
885.0 |
351.2 |
3.05 |
8199 |
| 8 |
3719.9 |
1770.8 |
2.75 |
36112 |
AMD Athlon MP 2000
| level |
time for PCop |
time for GMRES |
speedup |
memory |
| 5 |
26.4 |
13.5 |
4.92 |
391 |
| 6 |
113.3 |
65.4 |
5.31 |
- |
The following table shows the speedup of the systems compared to the Sun
E6000. Written with a red color is the fastest system on each level.
| level |
SF6800 |
SF15k |
HP9000 |
P690 |
Athlon |
| 5 |
2.86 |
2.79 |
2.80 |
3.07 |
4.92 |
| 6 |
2.78 |
2.67 |
3.06 |
3.37 |
5.31 |
| 7 |
2.74 |
2.56 |
1.43 |
3.05 |
- |
| 8 |
2.39 |
1.86 |
2.54 |
2.75 |
- |