An information-theoretic view of cloud workloads

Abstract

Analytics-as-a-service is emerging as a key offering for cloud systems, however in the petascale regime, data transfer bottlenecks are a limiting factor. Often information has to be transmitted to the cloud by physical transportation. Efficient information representations that leverage the functional purpose of data for the analytics service to be offered can serve to ameliorate many of these information flow bottlenecks. In this paper, we provide an information-theoretic view on optimal information representations for big data analytics in the cloud. We also provide some structural design principles for building a petascale analytics appliance.