optimizing power and energy lei fan, martyn romanko
TRANSCRIPT
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 Power-Aware 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?