Next: Globus performance on a
Up: 568K mesh
Previous: Influence of the "0"
Shown in Table 16 are the performances
for different combinations of MecaGRID clusters using the 568K mesh
relative to the nina times with the -O1 and -O3 options.
Also shown for comparison are the performances for the 262K mesh.
Table 16:
Globus performance summary: 262K versus 568K
|
262K |
568K |
Name of cluster(s) |
-O1 |
-O1 |
-O3 |
nina |
1.0 |
1.0 |
1.0 |
iusti |
1.0 |
0.8 |
0.9 |
pf |
1.7 |
1.4 |
1.9 |
cemef |
2.3 |
1.9 |
1.9 |
nina-pf |
1.2 |
1.3 |
1.6 |
inter cluster |
|
|
|
nina-iusti |
1.9 |
2.2 |
5.8 |
nina-cemef |
1.7 |
2.4 |
3.7 |
pf-iusti |
2.5 |
2.4 |
4.0 |
pf-cemef |
2.2 |
2.6 |
7.3 |
iusti-cemef |
2.9 |
3.8 |
6.8 |
|
Regarding the shock-bubble test case with the 568K vertices mesh,
the following observations are made:
- The IUSTI cluster's performed 20 percent faster than the INRIA-nina cluster.
- The INRIA-pf cluster gives a slightly better performance than the CEMEF cluster (1.4 and 1.9),
showing better a than expected performance relative to the INRIA-nina cluster.
- Inter-cluster Grid computations involving the CEMEF cluster tend to be the least efficient.
- The inter-cluster performances relative to nina using the -O1 compile option are
better than with the -O3 compile option. This occurs due to a larger Work/processor when compiled
with the -O1 option and therefore a smaller Communication/Work time ratio. A way of decreasing
this ratio using the -O3 compile option is to increase the mesh size holding the number of CPUs fixed.
Next: Globus performance on a
Up: 568K mesh
Previous: Influence of the "0"
Stephen Wornom
2004-09-10