GPU Cache System for COMPSs: A Task-Based Distributed Computing Framework

In this paper, we propose a novel GPU cache system for COMPSs, a task-based distributed computing framework that enables the execution of parallel applications on heterogeneous clusters. GPU COMPSs tasks can exploit the computational power of NVIDIA GPUs to process large data blocks. However, the current implementation of COMPSs requires each task to write its output data to disk and the subsequent tasks to read them from disk, which introduces significant overhead. To overcome this limitation, we design and implement a GPU cache system that allows tasks to store and retrieve data from the GPU memory, avoiding unnecessary disk operations and reducing data transfer time. We conducted extensive experiments on several benchmarks and demonstrated that our GPU cache system can achieve significant speedups compared to the baseline COMPSs implementation.