With the advent of big data era, it has become even more important to make data processing systems optimized for data-intensive workloads. However, creating such a system is very challenging due to (a) the diverging gap between computation speed and data access speed, and (b) heterogeneity of computation platforms specialized for certain data-intensive workloads.

 

Our proposal has two main components. First, we aim to explore heterogeneous hardware architectures for efficient processing of data-intensive workloads. Second, we aim to devise a good software abstraction model that provides an easy way to describe complex data analytics workloads yet being able to aggressively optimize their implementations in a target-(in)dependent manner.