From past experience we know that the speed and frequency of garbage collection is a vital part of maximizing high user load throughput, so despite documentation warnings we thought it worth devoting some experimentation time to swapping JVMs and studying the collection behavior.