Disclosed herein is a method for improving Input/Output (I/O) performance in a host system having multiple CPUs. Under this method, various interrupt affinity schemes are provided, which associate multiple processors, interrupts, and I/O channels for sending the interrupts, thereby allowing the interrupts to be almost evenly loaded among the multiple I/O channels and processors. Also, data locality (“warm cache”) can be achieved through the interrupt affinity schemes that associate each interrupt to its source processor, namely, the processor originating the I/O request that results in the interrupt.