Temperature-driven Thread mapping and Shadowing in Hybrid Multi-cores (SMASH)

点击翻译按钮获取中文摘要

This project aims at the investigation of novel temperature-driven runtime techniques to facilitate dependable operation of future hybrid multi-cores comprising programmable and reconfigurable cores. Previously, we mapped threads dynamically to software and hardware cores to optimize for thermal balance and gradients based on sensors sampling core temperatures. We allowed for intra-modal thread migration where we migrate threads between software cores or hardware cores, respectively. We now propose to migrate threads even across the software/hardware boundary which is denoted as trans-modal migration. Furthermore, we previously introduced thread shadowing to detect core failures. The operating system autonomously attaches shadow threads to application threads and compares their signatures of operating system calls. We now propose to also shadow memory accesses to be able to detect more types of failures. Furthermore, we will develop error injection methods which will form a basis for studying error models and to evaluate the effectiveness of our shadowing techniques. We proposed a thermal sensitivity map that captures information about the thermal stress threads pose on cores as well as the cores’ susceptibility for thermal stress in the last project phase. Beyond that, we will now apply on-line learning techniques to determine static and dynamic parameters of the thermal sensitivity map which enables us to develop prediction-based thermally-driven mapping strategies.