Magdalena Balazinska, Hari Balakrishnan, and Mike Stonebraker

This paper focuses on load management in loosely-coupled federated distributed systems. We present a distributed mechanism for moving load between autonomous participants using bilateral contracts that are negotiated offline and that set bounded prices for moving load. We show that our mechanism has good incentive properties, efficiently redistributes excess load, and has a low overhead in practice.

Our load management mechanism is especially well-suited for distributed stream-processing applications, an emerging class of data-intensive applications that employ a ``continuous query processing'' model. In this model, streams of data are processed and composed continuously as they arrive rather than after they are indexed and stored. We have implemented the mechanism in the {\em Medusa} distributed stream processing system, and we demonstrate its properties using simulations and experiments.

1st Symposium on Networked Systems Design and Implementation (NSDI)
San Francisco, CA, March 2004.

[PostScript] [PDF]