Optimizing Power and Energy

Download Report

Transcript Optimizing Power and Energy

Optimizing Power and
Energy
Lei Fan, Martyn Romanko
Motivation

31% of TCO attributed to power and cooling

Intermittent power constraints

Renewable energy

Grid balancing

20% - 30% utilization on average

Green: good for the environment

Green: saves money
Themes


Hybrid (hardware/software) optimizations

Dynamic DRAM refresh rates (Flikker)

Dynamic voltage/frequency scaling (MemScale)

Distributed UPS management

Power cycling (Blink)
Software optimizations

Dynamic adaptation (PowerDial)
Flikker: Saving DRAM Refresh-power
through Critical Data Partitioning

Partitioning of data into critical vs. non-critical

Partitioning of DRAM into normal vs. low refresh rates

Programming language construct


Allows marking of critical/non-critical sections
Primarily software with suggested hardware optimizations

OS and run-time support

Refresh rate optimizations
Flikker
MemScale: Active Low-Power Modes for
Main Memory

Modern DRAM devices allow for static scaling

MemScale adds:


DVFS for MC; DFS for memory channels and DRAM devices
Policy based on power consumption and performance slack
MemScale
Managing Distributed UPS Energy for
Effective Power Capping in Data Centers

Use of distributed UPSs to sustain peak power loads

Based on existing distributed UPS models

Larger batteries needed for longer peak spikes

Allows for more servers to be provisioned

Analysis of effect on battery lifetime

Argued benefit outweighed cost of extra batteries

Lacked detailed analysis on cooling costs
Blink: Managing Server Clusters on
Intermittent Power

Reducing energy footprint of data centers

Power-driven vs. workload driven


Blink: power-driven technique
Metered transitions between

High power active states

Low power inactive states
Blink


Three policies

Synchronous: optimizes for fairness

Activation: optimizes for hit rate

Load-proportional: both
Unknown effects of power cycling on component lifetime
PowerDial: Dynamic Knobs for PowerAware Computing


When is this applicable for a program?

QoS (accuracy) vs. power/performance tradeoff

Subject to system fluctuations
Dynamic tuning of program parameters

Adaptable to fluctuations in power/load

Determines control variables

Application Heartbeats framework provides feedback

Automatic insertion of API calls
PowerDial
Discussion, Questions?