Buffered Resource Constraint: Algorithms and Complexity

Abstract

The notion of buffered resource is useful in many problems. A buffer contains a finite set of items required by some activities, and changing the content of the buffer is costly. For instance, in instruction scheduling, the registers are a buffered resource and any switch of registers has a significant impact on the total runtime of the compiled code.

We first show that sequencing activities to minimize the number of switches in the buffer is NP-hard. We then introduce an algorithm which, given a set of already sequenced activities, computes a buffer assignment which minimizes the number of switches in linear time, i.e., O(nd) where n is the length of the sequence and d the number of buffered items. Next, we introduce an algorithm to achieve bound consistency on the constraint Switch, that bounds the number of changes in the buffer, in O(n ² d + n ^1.5 d ^1.5) time. Finally, we report the results of experimental evaluations that demonstrate the efficiency of this propagator.