energy consumption as a first-class “performance” goal for data intensive computing

6
A FIRST-CLASS “PERFORMANCE” GOAL FOR DATA INTENSIVE COMPUTING Jignesh M. Patel Credits: Student: Willis Lang Funding: NSF, Microsoft

Upload: yin

Post on 10-Feb-2016

20 views

Category:

Documents


0 download

DESCRIPTION

Energy Consumption as a First-Class “Performance” Goal for Data Intensive Computing. Credits : Student: Willis Lang Funding: NSF, Microsoft. Jignesh M. Patel. Rethinking Objectives for Data Intensive Computing. Performance. Performance. SLA Requirements. Energy Consumption. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Energy Consumption as a First-Class  “Performance” Goal  for Data Intensive Computing

ENERGY CONSUMPTION AS A FIRST-CLASS “PERFORMANCE”

GOAL FOR DATA INTENSIVE COMPUTING

Jignesh M. Patel

Credits: Student: Willis Lang

Funding: NSF, Microsoft

Page 2: Energy Consumption as a First-Class  “Performance” Goal  for Data Intensive Computing

2RETHINKING OBJECTIVES FOR DATA INTENSIVE COMPUTING

Perf

orm

ance

Energy Consumption

Performance

SLA Requirements

Page 3: Energy Consumption as a First-Class  “Performance” Goal  for Data Intensive Computing

3ENERGY EFFICIENCY

OPPORTUNITIES

Cluster Level

Node Level

Page 4: Energy Consumption as a First-Class  “Performance” Goal  for Data Intensive Computing

4ENERGY-ENHANCED DBMS

QUERY OPTIMIZATION

0.75 0.80 0.85 0.90 0.950.70

0.75

0.80

0.85

0.90

0.95

1.00

MJ, 2GB

MJ, 4GB

HJ, 2GB

HJ, 4GB

System Energy Consumption (ratio vs HJ,A)

Perf

orm

ance

(rati

o vs

HJ,

A) SLA• HJHash Join

• MJMerge Join

Commercial DB: join, 5GB WB tables, 0.01% selectivity

Page 5: Energy Consumption as a First-Class  “Performance” Goal  for Data Intensive Computing

5 SYNERGIES

Energy-based optimization/s

cheduling

Energy Efficiency /

Performance Tradeoffs

Page 6: Energy Consumption as a First-Class  “Performance” Goal  for Data Intensive Computing

6 QUESTIONS? Can the hardware and software tango together?

Does the hardware provide mechanisms that the software really needs?

Can the software specify what it needs from the hardware?

At the DC level, what are the real pain-points in energy-efficiency? Need to take a holistic end-to-end approach Are the low DC utilization observations still relevant with

techniques like spot-pricing?