No abstract

The power density may limit the amount of energy a many-core system can consume. A many-core at its maximum performance may lead to safe temperature violations and, consequently, result in reliability issues. Dynamic Thermal Management (DTM) techniques have been proposed to guarantee that many-core systems run at good performance without compromising reliability. DTM techniques rely on accurate temperature information and estimation, which is a computationally complex problem. However, related works usually abstract the temperature monitoring complexity, assuming available temperature sensors. An issue related to temperature sensors is their granularity, frequently measuring the temperature of a large system area instead of a processing element (PE) area. Therefore, the first goal of this work is to propose a finegrain (PE level) temperature monitoring for many-core systems. The second one is to present a dedicated hardware accelerator to estimate the system temperature. Results show that software performance can be a limiting factor when applying an accurate model to provide temperature estimation for system management. On the other side, the hardware accelerator connected to the many-core enables the fine-grain temperature estimation at runtime without sacrificing system performance.

New technology nodes enable the integration of billions of transistors in a small silicon area by replicating identical structures, resulting in many-core systems. However, power density may limit the amount of energy the system can consume. A many-core at its maximum performance may lead to safe temperature violations and, consequently, result in reliability issues. Dynamic Thermal Management (DTM) techniques proposals guarantee that many-core systems run at good performance without compromising reliability. In this paper, we review recent DTM works, discussing their limitations, and propose new heuristics for thermal-aware application mapping and migration, using a hardware accelerator that enables temperature monitoring on systems with a large number of processing elements. Results show that using straightforward heuristics, with reactive actions based on runtime temperature monitoring, reduce the peak temperature in high workload scenarios (6.8%), and improve thermal distribution significantly on a large (8x8) many-core system.

No abstract