A High-Performance Scheduling Algorithm for Mode Transition in PIM
Seungyong Lee, Sang-Hyun Lee, Minseok Seo, Chunmyung Park, Hyuk‐Jae Lee, Woojae Shin, Hyun Kim · 2021
Processing in memory (PIM) has emerged due to the lack of memory bandwidth. To add new commands in PIM, various architectures make use of memory modes. By changing the memory mode, memory can recognize PIM commands in existing system. However, this method would incur mode conflicts when the current mode of memory and the mode required by a request are different. Mode transition overhead caused by the mode conflict would be catastrophic in the environment where memory and PIM requests are randomly mixed. To resolve this problem, in this paper, we propose a scheduling algorithm named gather & issue (G&I). It reduces mode conflicts by gathering PIM requests in a queue and postponing until the number of PIM requests exceeds threshold. We evaluate it and other polices using inputs with varying memory/PIM ratios. G&I achieves the most improvement in performance showing closer fairness to other polices.