A shared input/output (IO) resource is managed in a decentralized manner. Each of multiple hosts having IO access to the shared resource, computes an average latency value that is normalized with respect to average IO request sizes, and stores the computed normalized latency value for later use.