Approximate Run Times on the Cluster
The table shows example problem sizes and the time it took to execute them using
different numbers of CPU's
Each problem had only one frequency point.
What you will notice is that there is a point where adding more CPU'S does not reduce
the execution time significantly.
This is because the network overheads increase as you add more CPU's and at one
point a new network segment is added.
So how should you use the table?
- See what size problem you have
- Make sure you have chosen enough CPU's to solve the problem (800MB per CPU*)
- The table will provide an estimate of the execution time and speedup you could achieve
should you wish to use more CPU's than the minimum required.
- The total execution time will approximately be the execution time in the table multiplied
by the number of frequency points in your problem.
*Note there are 2 CPU's per node.
|
10 CPU's (problems up to 9GB) |
8:33 min |
14:42 min |
- |
- |
- |
|
20 CPU's ( problems up to 18GB) |
5:41 min |
8:56 min |
36:17 min |
- |
- |
|
30 CPU's (problems up to 27GB) |
4:39 min |
7:31 min |
27:54 min |
- |
- |
|
40 CPU's (problems up to 36GB) |
4:34 min |
7:04 min |
25:50 min |
- |
- |
|
50 CPU's (problems up to 45GB) |
4:33 min |
7:01 min |
23:34 min |
69.23 min |
- |
|
60 CPU's (problems up to 54GB) |
4:22 min |
6:39 min |
22:26 min |
62.37 min |
- |
|
70 CPU's (problems up to 63GB) |
4:41min |
6:57 min |
22:15 min |
64.52 min |
- |
|
80 CPU's (problems up to 72GB) |
4:32 min |
6:49 min |
21:19 min |
60.19 min |
- |
|
90 CPU's (problems up to 81GB) |
4:48 min |
7:05min |
22:06 min |
60.02 min |
- |
|
98 CPU's (problems up to 88GB) |
4:44min |
6:53min |
21:38 min |
58.39 min |
140.53 min |
|