Towards a scalable, low-power all-optical architecture for networks-on-chip
ACM Transactions on Embedded Computing Systems (TECS) - Special Issue on Design Challenges for Many-Core Processors, Special Section on ESTIMedia'13 and Regular Papers
| Hi-index | 0.01 |
Many have advocated silicon-based nanophotonics for future many-core data communication. It is my goal to show its applicability to other many-core needs, namely many-core coordination. The same optical properties that benefit data communication may benefit many-core coordination: low cross-chip latencies, high bandwidth, and low power consumption. However, to take full advantage of these benefits, it will be necessary to accommodate the technology's limitations. One limiter is repeater inefficiency. Converting an optical signal to and from the electrical domain is relatively slow and also power-inefficient. Thus, it is in the interest of performance to keep the signal in optical form from source to destination. I propose optical tokens as a way to manage coordination while coping with repeater inefficiencies. In most cases, our optical tokens can be passed from any interested party to any other interested party with no repeats, thus allowing for flexible communication at the speed-of-light propagation. When a token is viewed as usage rights for a resource, it becomes a fast traveling mutex. I show two applications of optical tokens. The first application employs optical tokens in the arbiters of an optical data interconnect. The data interconnect is fast and requires fast arbitration for high utilization. Also, like the optical tokens, it avoids repeating data transmissions. These complimentary repeater-less designs allow a token to maintain pace with its associated bandwidth allocation and achieve very high utilization. I further extend the proposal to support flow control and fairness. The second application employs optical tokens to address cache coherence protocol complexity. Cache coherence protocol complexity arises from the presence of concurrency races, which though rare, must be detected and resolved. My proposal, called Atomic Coherence, simplifies protocols by serializing potential races before they issue to the interconnect. Without races, sophisticated protocols are substantially easier to design, debug, and validate. Race serialization comes at some cost to performance, but is kept to a minimum with low-latency optical tokens.