Defense Event

Energy-Efficient Algorithms on Mesh-Connected Systems With Additional Communication Links

Patrick Poon

 
Thursday, September 05, 2013
2:00pm - 4:00pm
3725 BBB

 

About the Event

Energy consumption has become a critical factor constraining the design of massively parallel computers, necessitating the development of new models and energy-efficient algorithms. In this work we take a fundamental abstract model of massive parallelism, the mesh-connected computer, and extend it with additional communication links motivated by recent advances in on-chip photonic interconnects. This new means of communication with optical signals rather than electrical signals can reduce the energy and/or time of calculations by providing faster communication between distant processing elements. Processors are arranged in a two-dimensional grid with wire connections between adjacent neighbors and an additional one or two layers of noncrossing optical connections. Varying constraints on the layout of optics affect how powerful the model can be. In this dissertation, three optical interconnection layouts are defined: the optical mesh, the optical mesh of trees, and the optical pyramid. For each layout, algorithms for solving important problems are presented. Since energy usage is an important factor, running times are given in terms of a peak-power constraint, where peak power is the maximum number of processors active at any one time. These results demonstrate advantages of optics in terms of improved time and energy usage over the standard mesh computer without optics. One of the most significant results shows an optimal nonlinear time/ peak-power tradeoff for sorting on the optical pyramid. This work shows asymptotic theoretical limits of computation and energy usage on an abstract model which takes physical constraints and developing interconnection technology into account.

Additional Information

Sponsor: Quentin F. Stout

Open to: Public