GPUSync: Architecture-Aware Management of GPUs for Predictable Multi-GPU Real-Time Systems

The integration of graphics processing units (GPUs) into real-time systems has recently become an active area of research. However, prior research on this topic has failed to produce real-time GPU allocation methods that fully exploit the available parallelism in GPU-enabled systems. In this paper, a GPU management framework called GPUSync is described that was designed with the goal of increasing parallelism in mind. GPUSync can be applied in multi-GPU real-time systems, is cognizant of the system bus architecture and affinity among computational tasks and GPUs, and fully exposes the parallelism offered by modern GPUs, even when closed-source GPU drivers are used. In empirical evaluations presented herein involving real-world applications, GPUSync improved real-time response times by three times or more, on average, making previously unschedulable workloads schedulable.