We have continued to study possible optical interconnections for multiprocessors.
The GLORI strategy is a multiprocessor interconnect scheme designed to aid the programmer, compiler, and/or run-time scheduler in facilitating locality, e.g., in partitioning, placement, and relocation of data and processes. It also aids the hardware architecture in performing the functions useful in capturing and exploiting locality, e.g., clustering, combining, and caching. The impetus behind GLORI is to use the interconnect as an additional resource for taking advantage of multiprocessor locality and to use optics to achieve this goal. Among some of GLORI's non-optical features are hierarchical topology, clustering, combining, and minimal delay routing protocols.
Optics is a promising interconnect technology for multiprocessors.
The usefulness of optics' many interconnect features are evaluated
using GLORI as a framework.
In the proposed GLORI organization,
processor-memory elements (PMEs) are clustered together
onto shared buses at the local interconnect level,
and 
bus clusters are arranged in a hypercube
at the global interconnect level.
Reconfiguration of PME connectivity occurs
system-wide to allow arbitrary groups of PMEs to be clustered,
but the baseline shared bus-hypercube topology is assumed to remain intact.
Interconnect links are established by
fiber and free-space holographic diffractive optics.
In previous work, we found that cluster affinity among PMEs is maintained over intervals of tens to hundreds of thousands of execution cycles. Interconnect reconfiguration occurring once every millisecond is therefore sufficient to exploit this behavior. Ferroelectric liquid crystal spatial light modulators and experimental photorefractive holographic devices assumed by GLORI can operate well within this switch frequency.