Next: Performance for different grid
Up: NEMO performance
Previous: Scaling for equal grid
Contents
Single core versus dual core performance
Comparison of single node versus virtual node mode shows that the runtime
is generally faster when running in single node mode. The likely reason
for this is that single node mode will create less contention for both
memory and I/O nodes than running in virtual node mode. Table 2
shows the runtimes for 256 and 221 processors using a 16 by 16 grid
for both single and virtual node modes. It should be noted that the 221
processor run has had the land only cells removed. The effect of removing
land only cells will be examined in section 6.5.
Table 2:
Runtime comparison for 60 time steps for single node (mppnppn=1)
and virtual node (mppnppn=2). Runs were performed using the PGI compiler
jpnij |
jpni |
jpnj |
Time for 60 steps (seconds) |
|
|
|
mppnppn=1 |
mppnppn=2 |
256 |
16 |
16 |
119.353 |
146.607 |
221 |
16 |
16 |
112.542 |
136.180 |
|
From table 2 we can see that single node mode is up to
18.59% faster than virtual node mode. As the charging structure on HECToR
is per core, single node mode will cost significantly more (almost
double) AU's than virtual node mode. Thus, running NEMO in single node mode
should only be considered if it's critical to obtain a fast solution.
Next: Performance for different grid
Up: NEMO performance
Previous: Scaling for equal grid
Contents