clusterTime Results

Results of Fibonacci Experiment

the results are best illustrated with the following graph. It shows the percentage deviation in execution time taken to compute the test function relative to the mean execution time. The data is split by compute node (color), and by whether the job occupies a multiple of 4 CPUs on the node or not.

Take away:

1. The mean time per run was 32 minutes, times ranged from 30 to 34 minutes
2. There is considerable variation in computation speed across compute nodes
3. it seems obvious what is causing this: here is the number of CPUs per compute node allocated. You can see that the slow nodes are the ones where not all CPUs are reserved for the experiment. (All machines except node075 have 4 CPUs, node075 has 12)

      node num.of.CPUs
1: node051           3
2: node056           4
3: node041           4
4: node038           4
5: node028           4
6: node029           3
7: node066           4
8: node064           3
9: node075          11

This shows how important it is to obtain an entire compute node (i.e. all CPUs on a machine).