Title: Application aware techniques for scheduling general purpose applications on Graphics Processing Units
Speaker: Mihir Awatramani, ECpE Gradudate Student
Advisors: Diane Rover, Professor; Joseph Zambreno, Associate Professor
Abstract: In the last decade, Graphics Processing Units (GPUs) have been widely adopted as accelerators for general purpose computing. In addition to the High Performance Computing domain, owing to a rise in the number of data parallel consumer applications, more recently there is a growing interest in using GPU acceleration for commercial data centers as well. With this increasing adoption of GPU computing in different domains, more and more applications are being ported to GPUs for acceleration. Alongside, GPU architectures continue to provide an increasing amount of computational and memory access throughput with each generation. To provide such high throughputs, GPU architectures incorporate several cores, each of which have hundreds of arithmetic units and support thousands of active threads. Scheduling such a high number of threads, from an increasingly diverse set of workloads, in hardware; is a non-trivial and interesting challenge.
We envision that to continue making advancements in GPU computing, application-aware scheduling techniques are required that would make the architecture more robust to the increasingly diverse set of workloads. In this work, we leverage the fact that workloads have varied runtime characteristics, and develop techniques that monitor the hardware state to make scheduling decisions. Precisely, in the GPU computing workflow, scheduling is done at three levels – system, core and thread. This work develops novel scheduling techniques at each of the three levels. The performance impact of each technique is evaluated against the state of the art using a cycle accurate GPU simulator and open source GPU computing applications. In this talk, I would present work that has been completed to date, provide conclusions established thus far and propose work that we have identified as next steps towards tackling the GPU scheduling challenges.