Low cost 10 gigabit/s optical interconnects for parallel processing

Abstract

Fiber optic links offer several advantages over copper-based links including greater transmission distances, reduced cable and connector bulk and improvedelectrical isolation yet copper-based links are still the predominant choice for parallel interconnections. The overwhelming disadvantage for fiber optic links to date has been the cost of the technology. The Jitney Optical Bus was designed to push the limits of low cost optics while maintaining enough performance for high speed parallel computing. Jitney is a 20 channel (500Mb/s/ch) optical bus packaged in a plastic molded lead frame with a 'snap together' plastic optical coupler. The cables use plastic ferrules and large core fiber.Manufacturing cost estimates are competitive with copper, and significantly under $100/Gbit/sec, for full duplex, Gigabyte/sec operation. Jitney has been successfully demonstrated in IBM AS/400 and RS6000 Power Parallel systems testbeds. Parallel Interconnect EnvironmentIn today's data processing systems there has been a shift away from stand alone systems with a single processor/memory frame and only data connections to storage or LANs, to systems with multiple processor/ memory units and a large, intraframe and interframe, switched, data processing network. The use of a large interconnect fabric has resulted in more flexible systems with higher reliability and availability and lower overall system costs. At the same time, there are more demands on the interconnect fabric for lower latency, higher bandwidth, smaller physical size, lower emissions and longer distances. Processor speeds and word size effectively set the interconnect bandwidth. In products today, one can find transfer rates of 250MBytes/s with projections of 2000MB/s within a few years. The low latency requirement would suggest that each new generation of parallel systems should have a shorter and shorter interconnect but the physical size of the systems prevents that. The systems are often designed with the intention of being assembled in a single room with the maximum interconnect limited to about 20m but, increasingly, large systems with 100's of processors are assembled spanning several rooms requiring much longer connections between rooms up to 100m or more. Because of the large number of high bandwidth interconnections, the total system cost becomes sensitive to the cost of the interconnect and the overwhelming technology choice is copper wire. As interframe data transfers approach 1GByte/s, copper cabling becomes increasingly awkward to use; the cables are bulkier, elaborate signal conditioning electronics must be added to suppress distortion, skew, and crosstalk (often impacting the data transfer architecture), and EMI becomes a larger system problem. Nevertheless, even for the longer room to room interconnections the use of fiber is relatively rare due to the high cost. To combat the high cost of fiber interconnections, link designers should make performance trade-offs for low cost tailored to the specific application requirements of parallel processors (<30m...

Efficient Parallel Algorithms for Hierarchical Clustering on Arrays with Reconfigurable Optical Buses

“…Currently, the optical interconnection techniques are used to establish the reconfigurability and simultaneous switching for massively parallel processing system [36]. Finally, it should be mentioned that the Jitney Optical Bus with 20 channels (500MbÂsÂch) has been designed for high speed parallel computing and successfully demonstrated in IBM ASÂ400 and RS6000 power parallel systems testbeds [15]. Due to these new developments, the models with reconfigurable optical buses are likely to become feasible architectures in the near future.…”

“…Recently, some optical interconnections, such as time-division multiplexed (TDM) systems [36], wavelength division multiplexed (WDM) systems [5], and time slot interchangers (TSI) [14], have been designed as various methods of establishing reconfigurability and switching for systems with TerabitÂs throughput requirements. Such interconnections can be used to provide massively parallel data communication among a few ports [15,36]. Most optical switch systems are constructed by the switch matrices with crossbar or tree structures [9,28].…”

Packaging aspects of the Jitney parallel optical interconnect

“…elsewhere [9]. The results of system link test have been discussed Briefly, wrap tests were conducted which showed error free operation over all channels simultaneously, system testbeds were set up to exercise the modules and cables in a database application, and operation with graded index fiber were evaluated.…”