Evaluating computation/communication balance in an object recognition task

I present performance measurements for some compute- and communications-intensive image processing tasks required by an object recognition application. The application was implemented as a task farm on a network of transputers, which still are one of the cleanest and best-balanced building blocks for parallel systems. I report on relevant properties of the basic algorithms the application is comprised of; some relevant details of the task farm software; performance measurements on a range of system sizes for tasks of very different granularity and characteristics (including broadcast of global data) with measured loads on processors and communications links; and the detrimental effects on performance of hardware variants that increase communications latency and reduce available bandwidth. The measurements show that for some of the image processing tasks, the transputer system is `on the edge' with respect to available bandwidth and latency. Some possible improvements to the communications infrastructure are discussed in light of these results. Finally, current approaches to affordable parallel computing, such as networks or clusters of workstations, are put into perspective by comparing them with the transputer system, using computation/communication balance as a figure of merit for the comparison. This shows that for current microprocessors, even their pin bandwidth is not sufficient to sustain equivalent (scaled) performance for this typical image processing application; any multiprocessor system built with current networking hardware is so unbalanced compared to the transputer system that it will be severely limited by communications bandwidth and latency.