Enabling Legacy Applications on Heterogeneous Platforms

In this paper we make the case for a runtime technique to seamlessly execute legacy applications on heterogeneous platforms consisting of CPUs and accelerators. We consider discrete as well as integrated heterogeneous platforms. In the former, CPU and accelerators have different memory systems; in the latter, accelerators share physical memory with the CPU. Our proposed runtime does not require any code changes to be made to the application. It automatically schedules compute-intensive routines found in legacy code on suitable computing resources, while reducing data transfer overhead between the CPU and accelerators. To reduce data movement, our runtime defers data transfers between different memory systems, and attempts to move computations to data instead of vice-versa. This could create multiple copies of the data - one on the CPU, and the others on the accelerators - leading to coherence issues. To address this problem, we propose adding an operating system module that maintains coherence by intercepting accesses to shared data and forcing synchronization. Thus, by exploiting existing mechanisms found in system software, we architect a non-intrusive technique to enable legacy applications take advantage of heterogeneous platforms. With neither software changes nor additional hardware support, the proposed system provides a unified compute and memory view to the application programmer.