PlanetenWachHundNetz: Instrumenting Infrastructure for Planetlab

Abstract

We are investigating distributed parallel prefix operations. Efficient implementations, like Google's MapReduce for example, utilize a reduction tree as well as involve locality-aware algorithms. In a static network reduction trees can be built once using a-priori known perfect proximity information and need never be touched again. In highly dynamic environments, such as peer-to-peer systems, this is substantially harder to achieve and maintain. From related work two types of reduction trees for dynamic systems that both build on structured overlays, also known as Distributed Hashtables (DHTs), seem to emerge. Both of these structures are not concerned with exploiting existing locality. We show how these data-structures can be augmented to be more efficient and take advantage of locality information that might exist in the underlying overlay. Distributed parallel prefix can be used to aggregate and thus reduce data from many sources. This is for example useful for statistics collection and applicationlevel monitoring. To validate our hypothesis we are building an application-level data collection system, called PlanetenWachHundNetz (PWHN-pronounced 'pawn'), which is German that loosely translates to 'PlanetaryWatchDogNet'. Our initial evaluation of it is going to be performed on the PlanetLab testbed.