Parallel Algorithms Using Unreliable Broadcasts

Parallel computing on networks of workstations (NOWs) allows effective use of existing hardware and software. NOWs are often connected using a broadcast medium such as an Ethernet, so it is natural to use broadcast as a communication primitive. However, existing parallel programs usually rely on point-to-point protocols for performing broadcasts since native broadcast schemes (UDP broadcast) do not guarantee delivery. Thus the potential of the broadcast medium is lost. Instead of making the broadcast reliable, we design algorithms which use an Unreliable Broadcast. Our algorithms are correct regardless of the reliability of the broadcast, but are more efficient the higher the reliability. We provide experimental results which show that UDP broadcasts are quite reliable in practice and can be used to design efficient Minimum Spanning Tree and Load Balancing algorithms for NOWs.