Project Proposal

Presented byMichael Kazecki

Outline• Background

– Algorithms

• Goals

• Ideas

• Proposal– Introduction– Motivation– Implementation

Background

Hill-ClimbingMid-point No. of Processors (P)

Target miss

Reduce DVFS

Choose optimum DVFS for that P just above miss

Yes

Continue until miss

Continue for another P`

Better than previous Yes

No

Disregard and switch sides

Cost: L lg N No

Optimum DVFS operating point

Hill-Climbing Optimization

Mid-point No. of Processors (P)

Apply DVFS formula for that P

Continue for another P`

Optimum DVFS operating point

Better than previous Yes

Disregard and switch sides

No

Final Algorithm

Hill-climbing + DVFS optimization

Cost: alpha lg N

alpha grows slower than L

Goal

Dynamically optimizing power consumption of a parallel application

Low-overhead algorithms for dynamic optimization that produces optimal power savings.

Ideas

Improve on existing algorithm developed to overcome local optima situations, such as the use of “jitter” or momentum.

Expand research in another dimension by varying each individual processor core DVFS

Propose a prediction based algorithm using past results to calculate DVFS in real-time

Proposal Introduction

Develop a history based algorithm that is optimized for highly variable applications

Build a history table to track patterns in application’s behavior

Use a lookup registry to match current pattern to a suitable DVFS level

Proposal Motivation

Target run-time power-performance adaptation of future processors that run non-repetitive parallel applications.

Maximizing power savings while delivering a specified level of performance in real-time with low overhead.

Proposal Implementation

Hardware supportRequires cache and a register to support the

history table and recent lookupRequires real-time hardware

power/performance monitoring

Software supportAfter search phase, mechanism needed to

control DVFS and number of processors

End

Questions?