Tree Arbiter with Nearest-Neighbour Scheduling

A tree arbiter designed to minimize the average delay between consecutive allocations of a contentious resource is described and evaluated. The idea is to divide the competing users into nested clusters corresponding to sub-trees of varying size, and to keep the resource within the smallest cluster currently containing requests. This design reduces the number of steps in the arbitration procedure and hence increases the overall throughput of requests. The performance of the new arbiter is analysed and is compared to that of an algorithm which maintains rst-in-rst-out order among requests. The trade-oos between the nearest-neighbour and FIFO designs are examined numerically over wide ranges of parameter values. It is shown that when the system becomes large and/or heavily loaded, the beneets of the the nearest-neighbour arbiter become greater.