Software thermal monitoring and management for high-performance microprocessors
Jun Yang, Ling Ling Jin · 2006
The evolution of microprocessors has been hindered by the increasing power consumption and the rate at which heat is dissipated on die. Excessive amount of heat creates reliability problems and reduces lifetime of a processor. While heat can be removed by the cooling package, designing such package for the worst-case temperature is not cost-effective as the worst-cases are rare and the cost increases super-linearly. Therefore, a more reasonable solution is to use a less expensive package with dynamic thermal managements (DTM) that throttles the processor performance when cooling is inadequate. To ensure an effective control of the chip temperature, it is imperative to be able to monitor the temperature variations across the die timely and accurately. Most current techniques rely on on-chip thermal sensors. Unfortunately, the significant variation in chip temperature both spatially and temporally exposes the limitation of the sensors since hot spots migrate with workloads. This work seeks an alternative approach to tracking chip temperature through an OS resident software module that generates live power and thermal spectral distributions of the processor. A software thermal sensor (STS) is developed in a Linux system with a Pentium 4 Northwood core. The software thermal sensor offers detailed power and temperature breakdowns for each functional unit at runtime. As a demonstration of its benefits, this work proposes a software level thermal-aware job scheduling mechanism to reduce the performance loss due to the thermal pressure. The method leverages the natural discrepancies in thermal behavior among different workloads, and schedule them to keep the chip temperature within the cooling limit so as to minimize the amount of throttling. This work also demonstrates the need of including topology-information into OS scheduling algorithm in multi-core architecture to improve the performance further under thermal constraints.