A Performance Model for a Thermally Adaptive Application Implemented in Reconfigurable HW
Abstract
A model is developed for estimating the performance of a thermally adaptive application. The goal of this model is to allow a user to estimate the performance impact of adjusting the thermal budget allotted an application for a given thermal condition. The model is developed by applying a multiple linear regression approach to experimentally obtained performance measurements. The thermal conditions used during the collection of performance measurements were a small subset of the large number of possible thermal conditions. The model is validated by comparing the performance predicted by the model against a set of performance measurements obtained using thermal conditions not included in the regression data set.
Table of Contents
- 1. Introduction
- 1.1 Motivation for Adaptive Thermal Management
- 1.2 Thermal Management Approach
- 1.3 Performance Model Goal
- 1.4 Outline
- 2. Related Work
- 2.1 Measuring Temperature
- 2.2 Dynamic Thermal Management (DTM)
- 2.2.1 DTM for Microprocessors
- 2.2.2 DTM for FPGAs
- 3. Platform Infrastructure Description
- 3.1 Platform Hardware
- 3.2 Thermally Adaptive Frequency Control
- 3.2.1 Overview
- 3.2.2 Implementation Details
- 4. Application Description
- 4.1 Image Correlation Application Overview
- 4.2 Services and Performance Metrics
- 4.3 Parameters and Factors
- 4.3.1 System
- 4.3.2 Workload and Operating Conditions
- 5. Performance Model
- 5.1 Assumptions
- 5.2 Avoiding Multicollinearity
- 5.3 Multiple Linear Regression
- 5.3.1 Model
- 5.3.2 Analysis of Variation
- 6. Performance Measurement Experiments
- 6.1 Experimental Setup
- 6.2 Results
- 7. Analysis
- 7.1 Model Parameter Estimation
- 7.2 Analysis of Variation
- 7.3 Model Validation
- 7.3.1 Factor Independence
- 7.3.2 Validation Against Measured Performance
- Summary
- References
- List of Acronyms
View complete report on-line
Shift-click to download the paper in Adobe Acrobat
format
List of other reports in this series
Back to Raj Jain's home page