A balancing process between I/O processor groups of a non-uniform multiprocessor system enables spreading of I/O workload across multiple I/O processor groups on a group base as soon as the I/O processor group with maximum group utilization reaches a certain high limit together with other processor groups being utilized significantly lower. The additional balancing is decreased step by step again when a certain low utilization limit is reached or the workload becomes more evenly balanced between the I/O processor groups. Checking if increase or decrease of the balancing is required is done periodically, but with low frequency to not affect overall performance. The checking and balancing happens asynchronously in predefined intervals. This solves the problem that with an increasing number of I/O processors the handling of initiatives leads to increased cache traffic and contention due to shared data structures, which slows down the I/O workload handling significantly.