Gordon is a flash-enabled cluster architecture for data-centric applications. It combines low-power Atom processors, flash memory and data-centric programming systems (e.g., MapReduce) to improve performance and efficiency.

Gordon provides a custom-built, specialized flash translation layer (FTL) that provides high-performance for the type of accesses that data-centric applications perform. The FTL uses multiple write points to improve inter-request parallelism and aggressive striping to maximize bandwidth.

We explore Gordon's design space to identify Pareto-optimal design points that trade-off performance and efficiency. We find that Gordon systems can outperform disk-based clusters by 1.5x and deliver 2.5x more performance per watt.

For more information, please contact Dr. Steven Swanson (swanson@cs.ucsd.edu), Adrian Caulfield (acaulfie@cs.uscd.edu) or Laura Grupp (lgrupp@cs.ucsd.edu).

Results

Results of our design space survey. The Pareto-optimal designs are on the lower-right perimeter of the cloud of points.

A scale drawing of one Gordon node.