Run-time adaptation in river

Abstract

We present the design, implementation, and evaluation of run-time adaptation within the River dataflow programming environment. The goal of the River system is to provide adaptive mechanisms that allow database query-processing applications to cope with performance variations that are common in cluster platforms. We describe the system and its basic mechanisms, and carefully evaluate those mechanisms and their effectiveness. In our analysis, we answer four previously unanswered and important questions. Are the core run-time adaptive mechanisms effective, especially as compared to the ideal? What are the keys to making them work well? Can applications easily use these primitives? And finally, are there situations in which run-time adaptation is not sufficient? In performing our study, we utilize a three-pronged approach, comparing results from idealized models of system behavior, targeted simulations, and a prototype implementation. As well as providing insight on the positives and negatives of run-time adaptation both specifically in River and in a broader context, we also comment on the interplay of modeling, simulation, and implementation in system design.