We have a group of 20 using a local 256-processor cluster for design and development activities. Over the past 12 months we've used the machine at about 65% capacity (in terms of number of processors running jobs). Of course we have months at a time where we are at 100% and other times where we are well below 50%. Is this typical for an in-house system or is this system under-utilized? Thanks Rob Kunz

My guess is that when your have a 90% utilization of your cluster on average (for a year) you probably need expansion. Your purchase department (or alternatively the department with the budget) might think otherwise. So I think 65% is reasonable.

When it is production (doing the same CFD runs over and over with small changes) you might be able to get a high utilization. When the runs differ (research) you will need time to postprocess the results and make changes to a new run. This costs time. However you still want to have a solution as fast as possible. Finacially this means: Your time is expensive, the machines' is not.

Thanks Bart: Yes, I think that most large shared systems, like a central university system or NASA or DoD system with many users, have high utilization, >90%. But that comes at the cost of competition. Users have to wait in queues, write proposals, etc... As you observe, for a local system we do not want our people to compete with eachother, we want them to get their work done, so the target size machine should be closer to maximum utilizatino need. Then we pay the price of lower average utilization.