Content area
Full text
Abstract: The communication is predicted to pass the computation as the limiting factor of performance of complex digital circuits. The shared bus is the most common communication medium in system on chip (SoC) and buses have significantly evolved along increased requirements. One of the new properties of the buses is distributed arbitration. The study presents a novel dynamically adaptive arbitration algorithm and compares it with round-robin, priority, their combination and random algorithms, all with varying parameters. Algorithms are compared in two multiprocessor SoCs with 4 and 11 processors with IP blocks on field programmable gate array (FPGA). The algorithms are benchmarked with a complete MPEG-4 encoder. Different bus utilisation levels are considered by scaling the bus frequency with respect to speed of the processors. Results show that the arbitration algorithm may account for up to 1.6 times increase in performance and optimising the transfer lengths may yield speed-up of 4.4 times in application execution. The proposed dynamically adaptive arbitration was found to be the best overall algorithm in performance.
1 Introduction
As the requirements for on-chip bus architectures increase, the weight of the arbitration algorithms needs to be reassessed. Arbitration is a mechanism that decides the owner of a shared resource, the bus in this case. Previous studies are typically based on a centralised arbitration approach [1-7], However, contemporary buses benefit from distributed arbitration, where long request-grant wiring is avoided, thus reducing the number of global wires and enhancing the bus scalability. As a consequence, the roles of the arbitration algorithms change. Because of lack of request-grant signalling, the distributed arbitration algorithm has to predict which intellectual property (IP) block next wants to access the bus, instead of resolving a request contention. In practice, almost the same algorithms as with the centralised arbitration can be used, but now the bus may also be offered to an IP block, which does not request it. The centralised arbitration is dominating in embedded systems currently. On the other hand, Intel Itanium 2 system bus uses symmetric distributed arbitration (with fixed round-robin arbitration) [8], HewlettPackard's runaway bus uses distributed arbitration [9], and also Silicon Backplane uses distributed arbitration [10]. However, no arbitration algorithm comparison on these architectures is present.
Synthetic test cases are often used in the...