High performance computing systems have made a spectacular progress over four decades. As microprocessor speeds reach the GHz region; and their bandwidth requirements are measured in hundreds of Gb/s, fundamental communication limits hinder further progress in interconnection networks performance.
Optical technology is promising to release the communication bottleneck by offering increased bandwidth and reduced latency and power. Recent breakthroughs are now leading to new opportunities in the creation of optical interconnection networks for high performance computing systems.
The optical medium, however, presents challenges which complicate the design of optical interconnection networks: buffering and processing resources, abundant in silicon electronics, are scarce or non-existent in optics. Additionally, optical switching technologies are not as reliable or as scalable as their electronic equivalents. Signal distortions in optical switches may lead to errors if physical properties are not properly considered.
In this work I present the design of two optical interconnection network architectures, in an attempt to address the requirements of HPC systems in terms of bandwidth, latency, and power while understanding and abiding by the properties and limitations of optical switching and transmission technologies.
The SPINet architecture is an architecture for an interconnection network in large scale HPC systems, as a chip-to-chip or rack-to-rack interconnect. The work here shows that the system can deliver very high transmission bandwidths at low latencies while scaling to large ports counts. The SPINet architecture is experimentally validated on a prototype network, proving its credibility.
The ICON architecture is designed as an architecture for a photonic network-on-chip (NoC). As the intra-chip communication bandwidths in multicore systems increase, so does their power consumption. Photonic NoCs can address high-bandwidth communication requirements while dissipating substantially less power, thus offering unparalleled advantages in terms of bandwidth per Watt. The work on the ICON includes a power study and a design-space exploration.
The main conclusion from this work is that optical interconnection networks offer advantages that will inevitably lead to their integration in HPC systems. This integration, when it happens, will require a paradigm shift in interconnection network design and architecture. The work presented is an example of this paradigm shift.