Dynamic Warp Scheduler Selection Policy Using Linear Regression for GPUs


GPGPU has been popular in many research fields due to its ability to execute massive operations concurrently. Each and every emerging GPU application shows different execution behavior, so it becomes very important to provide suitable execution environment for each application. Warp scheduling is one of the key factors that determine the performance of GPGPUs. Accordingly, various warp scheduling policies have been proposed. However, general warp scheduling policies target different optimization points, thereby dynamic switching of warp scheduler could yield extra performance gain.

In this work, we propose dynamic warp scheduler selection policy that dynamically selects the optimal warp scheduler for each GPU kernel by predicting the performances of given warp scheduling policies. We consider three policies: LRR, TLV, and GTO. Scheduler selection is performed based on multiple linear regression by monitoring architectural events dynamically. Our evaluation results show that proposed scheme improves performance of GPU by 3.6%, 9.8%, and 7.7%, compared to the LRR, TLV, and GTO, respectively.

International Conference on Electronics, Information, and Communication (ICEIC)
Ipoom Jeong
Ipoom Jeong
Assistant Professor

My research interests include CPU/GPU microarchitectures, memory/storage system designs, and smart-I/O devices