MAPREDUCE IMPLEMENTATION WITH MULTI-GPU

Map Reduce is a programming framework introduced by Google for large-scale data processing. Several studies have implemented Map Reduce model on graphic processing unit (GPU), however most of which are based on the single-GPU approach. Because of the scalability issue, the real-world applications require that the GPU-based MapReduce to be revised to support multiple GPUs on single machines. On the other hand, the current state-of-the-art MapReduce employs a Shuffle stage with sorting between the Map stage and the Reduce stage, and uses atomic operations among different threads. However, the sequential merge inside the Shuffle stage and serialized memory access with global atomic operations can severely impact the performance. In this paper, we present MGMR, a standalone MapReduce system that takes advantage of multi-GPU for large-scale data processing. Experimental results show the effectiveness of MGMR and performance gains over CPU-based approach.