next up previous
Next: Conclusions Up: Empirical Results Previous: Convergence Rate

Comparing Communications Paradigms

Underlying the whole concept of active messages is the paradigm that the user takes the responsibility for handling messages as they arrive at a destination. The user writes a handler function that takes the data from a register and uses it in a calculation or assigns the data to memory. By assigning message handling responsibilities to the user, communications overhead can be significantly reduced. The effect of reduced overhead can be clearly seen in this algorithm, when performance of an active message-based algorithm is compared to performance of an algorithm with more common blocking send and receive commands. The requirement in this algorithm to broadcast the values of before the next color can proceed causes substantial amounts of communications. In the portion of the algorithm that solves for values of in the last diagonal block, the amount of communications is , and as the number of processors increases, the size of the messages for conventional message passing decreases. For traditional message passing paradigms, the cost for communications increases drastically as the number of processors increases, because each message incurs the same latency regardless of the amount of data sent. Meanwhile, with active messages, latency is greatly reduced because the user has the responsibility to process the message. This increase in the number of messages can be seen in figure 10, as the performance for solving for values in the last block eventually increases slightly as the number of processors increases. For an algorithm based on a more traditional send and receive paradigm, performance quickly becomes unacceptable in this portion of the calculations as the number of processors increases.



David P. Koester
Sun Oct 22 15:29:26 EDT 1995