STAR  Computing  GRID  Scheduler local vs NFS benchmark

The scheduler is able to assign the input file from both local filesystem and network filesystem. In this test we try to compare the performance of these two methods.

Benchmarking is complicated and expensive because it has to take into account multiple factors. This is just a simple test, and it doesn't expect to be in any way complete. Here we are not comparing the performance of local filesystems over network filesystem. In fact, the result will depend on multiple factors as the specific filesystems (NFS and Linux), the setup at RCF, how ROOT and the STAR software reads file, and what the ROOT macro itself will run. We are actually going to get a feel of what is the performance perceived by the job when running.

The test consists of a root macro running over 4000 STAR MuDST. The job is dispatched through the scheduler, the first time requiring NFS files, and the second time using local files. The NFS server at the time wasn't under any particular load.

To measure the running time of the job, the CPU time as reported by LSF is used. The number of files is taken from the script generated by the scheduler and the number of events is taken by the output of the macro.

The results of the test can be found here (NFS - local). For each process submitted to LSF, the average time per file and per event is computed. The average time per file is 14.13 ± 6.58 seconds on NFS versus 8.37 ± 2.48 seconds on local disk. The average time per event is 84.71 ± 43.13 ms on NFS versus 55.21 ± 18.53 ms on local disk.

I have also prepared two histogram that compare the distributions of the CPU per file and per event.

The timing are not extremely different, but one should note that NFS performance depends a lot on the load on the server. During this test, the server weren't overloaded, and even in these condition, NFS performance was never better than local performance. The scheduler itself is helping reduce the load on the NFS server (by correctly using LSF static resources).

It is interesting to note that NFS has in both cases two peaks, one where the peak for local is, and one higher. It is not clear to me hot to interpret that result.

Gabriele Carcassi - page was last modified