Method and system for determining an optimal workload level for a data processing system. In various embodiments samples of system operating characteristics are collected. Each sample includes a workload level, a throughput level, and one or more performance metrics having values used in detection of a bottleneck condition. A bottleneck condition exists if the workload level, the throughput level, and one or more of the performance metrics in a sample have values equal to a predetermined set of values defined as a bottleneck condition. From the samples, a first workload level is identified that corresponds to the sample having a greatest throughput level. A workload level is selected from among workload levels of one or more samples as the optimal workload level, where, in each of the one or more samples the workload level is less than the first workload level, the throughput level is less than the greatest throughput level, and no bottleneck condition is indicated by the performance metrics.