It is confusing. They meant the benchmark used a 500ms "workload" (sleep) per request handler -- x isn't really an axis.