Network

With respect to the networking hardware, we assume that eventually ATM networks will link the processing units with a very low error rate such that lost packets are a rare event. Assuming a bit error rate of 10 and a packet size of about 1kB, there will be a packet loss at most every 10 packets. For example, an additional overhead of say 1s/packet due to a more sophisticated protocol will add up to a full second of overhead per dropped packet. For this reason, we think that on low error rate networks, minimal protocol overhead is an important design goal. While we make the common case go fast, we must also ensure that the rare case executes correctly. Assuming a correct MPI application, our protocol should handle all failure modes correctly.