ReM: Real-time Multi-core Systems
Gaining processor performance by increasing the operating frequency is diminished because of various reasons. In order to keep increasing the processor performance, the multi/many core platform has emerged. As the result, the possibility of deploying multi/many core platforms in the critical path of the real-time embedded application has increased to interact with (through sensors and actuators) and control physical processes in automotive, avionics, and various process industries (e.g. pharmaceutical, manufactory, energy, etc.). One of the most well-known multi/many core architecture is GPU.
Although the GPU may increase overall performance of the application, the current implementation of the GPU scheduling framework is not enough to deploy the GPU in the real-time domain. This is because the GPU implicitly processes its workload in sequential way and there is no proper preemption mechanism which is essential in a real-time domain.
In order to tackle this problem, we propose a run-time scheduling framework for the real-time application on the GPGPU platform. During run-time, the proposed framework estimates the current status of the application by using the execution and the timing model of the platform. At the same time, the proposed framework dynamically allocates the GPU resources to the applications based on the application information and the requirements.