SFNN: Sparse Flat Neighborhood Networks

What is a Sparse FNN?

A Sparse FNN is a network that guarantees single-switch latency and full link bandwidth per PE pair on a selected set of parallel communication patterns. A SFNN scales from a few tens of nodes, to tens of thousands of nodes, and higher as the design tools improve.

Publications On SFNNs

Timothy I. Mattox, Henry G. Dietz and William R. Dieter, "Sparse Flat Neighborhood Networks (SFNNs): Scalable Guaranteed Pairwise Bandwidth and Unit Latency," in the Proceedings of the Fifth Workshop on Massively Parallel Processing (WMPP'05) held in conjunction with the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS 2005), Denver, CO, USA, April, 2005. Preprints are available as 182KB PDF and 1MB PS versions for personal use only. Slides from the talk are available as 106KB PDF and 140KB PPT versions.

Tools for Creating SFNN Designs

The tools for making use of Sparse FNN technology are still under construction. We have developed a very fast heuristic for finding Sparse FNNs, but the user interface and surrounding tools are still in very early stages. When we have these tools in better shape, they will be released here.

The Sparse Flat Neighborhood Concept

The main problem with Universal FNNs is their limited scalability. To get the latency, bandwidth, and cost benefits of FNNs for larger clusters, one needs to relax the FNN design constraints. Instead of guaranteeing all PE pairs have single switch latency (and link bandwidth), we only need to guarantee these FNN properties for PE pairs that will communicate in significant quantities and/or at high frequency. See the above publications for more detail.

KASY0, the first machine with a SFNN

In the summer of 2003, we built the first Sparse FNN as part of the KASY0 machine. See the KASY0 home page for an overview, and see the KASY0 Technical Information page for more details on it's SFNN.

The only thing set in stone is our name.