Control and reliability of optical networks in multiprocessors

Optical communication links have great potential to improve the performance of interconnection networks within large parallel multiprocessors, but the problems of semiconductor laser drive control and reliability inhibit their wide use. These problems have been solved in the telecommunications context, but the telecommunications solutions, based on a small number of links, are often too bulky, complex, power-hungry, and expensive to be feasible for use in a multiprocessor network with thousands of optical links. The main problems with the telecommunications approaches are that they are, by definition, designed for long-distance communication and therefore deal with communications links in isolation, instead of in an overall systems context. By taking a system-level approach to solving the laser reliability problem in a multiprocessor, and by exploiting the short-distance nature of the links, one can achieve small, simple, low-power, and inexpensive solutions, practical for implementation in the thousands of optical links that might be used in a multiprocessor. Through modeling and experimentation, I demonstrate that such system-level solutions exist, and are feasible for use in a multiprocessor network. I divide semiconductor laser reliability problems into two classes: transient errors and hard failures, and develop solutions to each type of problem in the context of a large multiprocessor. I find that for transient errors, the computer system would require a very low bit-error-rate (BER), such as 10-23, if no provision were made for error control. Optical links cannot achieve such rates directly, but I find that a much more reasonable link-level BER (such as 107) would be acceptable with simple error detection coding. I then propose a feedback system that will enable lasers to achieve these error levels even when laser threshold current varies. Instead of telecommunications techniques, which require laser output power monitors, I describe a software-based feedback system using BER levels for laser drive control. I demonstrate, via a BER-based laser drive control experiment, that this method is feasible. It maintains a BER of 10-9: much better than my error control coding system would need. The feedback system can also compensate for optical medium degradation, and can help control hard failures by tracking laser wearout trends. For hard failures, a common telecommunications solution is to provide redundant spare optical links to replace failed ones. Unfortunately, this involves the inclusion of many extra, otherwise unneeded optical links, most of which will remain unused throughout the system lifetime. I present a new approach, which I call 'bandwidth fallback', which allows continued use of partially-failed channels while still accepting full-width data inputs. This provides, at a very small performance penalty, a high reliability level while needing no spare links at all. I conclude that the drive control and reliability problems of semiconductor lasers do not bar their use in large scale multiprocessors, since inexpensive system-level solutions to them are possible. Thesis Supervisor: Anant Agarwal Title: Associate Professor of Computer Science