Soren: Adaptive MapReduce for Programmable GPUs

In recent years the MapReduce programming model has been widely used for developing parallel data-intensive applications. As a result of its popularity, there exist many implementations of the MapReduce model on different parallel architectures including on massively parallel programmable GPUs. A basic challenge in implementing a MapReduce runtime system is the wide diversity of applications developed based on the model. That means a fixed implementation of the MapReduce runtime system may become suboptimal for some classes of applications. In this paper, we propose an adaptive framework for MapReduce on GPUs which is capable of monitoring key characteristics of applications and dynamically executing them efficiently in one of the three variations of the MapReduce engine it implements. Our preliminary results show that our adaptive method can significantly improve performance for many MapReduce applications (including a 11x performance speedup in one case) compared to a state-of-the-art MapReduce implementation on GPUs.