paper 5.2 Consistent Runtime Thermal Prediction and Control Through Workload Phase Detection

ABSTRACT
Elevated temperatures impact the performance, power consumption, and reliability of processors, which rely on integrated thermal sensors to measure runtime thermal behavior.
1. INTRODUCTION

Dynamic thermal management (DTM) techniques allow processors to optimize performance while avoiding thermal violations. The most well-known DTM techniques include clock gating, dynamic
voltage and frequency scaling (DVFS), and thread migration/scheduling [14, 18, 6, 5].

2. RELATED WORK AND MOTIVATION

3. PROPOSED PHASE-AWARE THERMAL PREDICTION METHODOLOGY

At the highest level, our phase-aware thermal prediction approach
takes raw performance counter data that is periodically measured
for each core during workload operation and translates this data
into a temperature projection for some interval into the future using
the concept of workload phases.

In order to define workload phases and capture temperature dynamics
within them in a computationally efficient manner, we propose
the methodology that is illustrated by Figure 2.

3.1 Offline Thermal Phase Analysis

In order to avoid excessive runtime overhead, global phase analysis
and within-phase temperature modeling are performed offline
using data generated for a set of representative workloads.

(page 3, col 2)

page 4
3.2 Runtime Thermal Prediction and Control

4. EXPERIMENTAL RESULTS
A. Experimental Infrastructure.
(page 5)

5. CONCLUSIONS