A modular network for dense optical interconnection of processing elements
A.B. Ruighaver · ACM SIGARCH Computer Architecture News · 1990
The efficient exploitation of parallelism in scientific applications is often limited by the insufficient communication capabilities of current multicomputer architectures.Routing hardware in the second generation of multicomputers [Athas88] has removed some of the latency of data transfer between non-neighbouring nodes, but actual data transfer delays can still be high as a result of the limited throughput of the network.Because routing hardware is only efficient for large packets of data, the efficient exploitation of small-to-medium granularity parallelism will be restricted.To extend the applicability of multicomputer systems we need to reduce this communication bottle-neck.When an application has only limited parallelism available, a multicomputer needs to be able to exploit small granularity parallelism without its data transfer delays reducing the efficiency.In our view a sparse topology, such as a hypercube or a mesh, will not provide sufficient throughput for a next generation parallel processor.A network with full parallel interconnection between all Processing Elements will certainly be the best solution to this problem.Such networks are not feasible with traditional technology.Using optical communication, however, full interconnection becomes a real alternative for medium sized parallel architectures.