Transparent Avoidance of Redundant Data Transfer on GPU-enabled Apache Spark

Abstract

This paper presents an extension to IBMSparkGPU, which is an Apache Spark framework capable of compute- or memory-intensive tasks on a graphics processing unit (GPU). The key contribution of this extension is an automated runtime that implicitly avoids redundant CPU-GPU data transfers without code modification. To realize this transparent capability, the runtime analyzes data dependencies of the target Spark code dynamically; thus, intermediate data on GPU can be cached, reused, and replaced appropriately to achieve acceleration. Experimental results demonstrate that the proposed runtime accelerates a machine learning application by a factor of 1.3. We expect that the proposed transparent runtime will be useful for accelerating IBMSparkGPU applications, which typically include a chain of GPU-offloaded tasks.