Consider a problem which can always be partitioned into P tasks.
Assume that 10% of computation cycles will require data owned by other tasks.
The latency of the IN can have a huge effect on parallel program performance.