Bridging the Computation Gap in a Future of Massive Data Fred Chong - - PowerPoint PPT Presentation

bridging the computation gap in a future of massive data
SMART_READER_LITE
LIVE PREVIEW

Bridging the Computation Gap in a Future of Massive Data Fred Chong - - PowerPoint PPT Presentation

Bridging the Computation Gap in a Future of Massive Data Fred Chong Director, Greenscale Center for Energy-Efficient Computing Director, Computer Engineering UC Santa Barbara Computation Gap Optimistic technology scaling assumptions


slide-1
SLIDE 1

Bridging the Computation Gap in a Future of Massive Data

Fred Chong

Director, Greenscale Center for Energy-Efficient Computing Director, Computer Engineering UC Santa Barbara

slide-2
SLIDE 2

Computation Gap

  • Optimistic technology scaling assumptions
  • “Internet of Things”
  • Greg Papadopoulos, keynote IGCC June 2012:

“$1 trillion market for ubiquitous sensors.”

slide-3
SLIDE 3

Outline

  • Bridging the gap
  • Environmental Costs
  • Cultural change
slide-4
SLIDE 4

Efficiency Gap

  • IEE/Kavli Roundtable
  • 40X efficiency gap in 13

years

  • No single solution

– Gains needed at many levels

  • f the system

[IEEE Design and Test 2014]

slide-5
SLIDE 5

Solutions

  • Eliminate Waste

– Turn stuff off – Avoid overprovisioning

  • Change the Rules

– New Technologies – Approximate Computing

  • Will give several examples

– About 20% savings each

slide-6
SLIDE 6

Barely-Alive Servers

0.2 0.4 0.6 0.8 1 Week day Weekend day

  • Norm. Energy

Base Somniloquy On/Off Barely Alive [JETC’12]

  • Turn off microprocessors but allow other servers to use memory
  • Decouple load variation from data variation
slide-7
SLIDE 7

Offered load Power Two-speed LogStore Low speed High speed Log is active here A B

LogStore: Extending Low-Power Disk Modes

  • Random disk writes are energy intensive (require higher speed to

meet performance needs)

  • Sequentially logging writes can defer high-speed operation

[FAST’12]

slide-8
SLIDE 8

Targeted Thermoelectric Cooling

  • Superlattice layer
  • n microprocessor
  • Acts as a Peltier

heat spreader targeted at hot spots

  • Avoids worst-case

provisioning in datacenter-level cooling

[ISCA’11]

slide-9
SLIDE 9

Heterogeneous 3D Phase-Change Memory

  • Different operating temperatures in 3D stack
  • Tailor GST mixture to operating temperature
  • 10% memory energy savings

PCM cells One Bank

Ge Te

GeTe Sb2Te3 Ge2Sb2Te5 0% 100% 0% 100% 0% 100%

slide-10
SLIDE 10

Computational Sprinting

Wenisch, U Mich

slide-11
SLIDE 11

Phase-Change Heat Sink

slide-12
SLIDE 12

Deep Memory Hierarchies

  • Motivation: Hierarchy is very non-

energy-proportional

– Existing technologies: faster flash, multi-speed & IDP disks – New byte-addressable technologies: PCM & STT-RAM – Deep hierarchy can improve energy- proportionality

  • Recent Progress:

– Predict data location instead of search – Simpler design allows compact table to be recalibrated periodically (22% energy savings)

Main memory PCM/Memristor SSDs MSD/IDP disks Regular disks

[IPDPS’14 (Best Paper)]

slide-13
SLIDE 13

Memory De-duplication

  • 32% avg memory savings on MPI apps

– 60% max

5000000 10000000 15000000 20000000 25000000 30000000 100 200 300 400 500 600 700 800 900 1000 Memory Footprint (Bytes) Millions of Memory References Default Merged

[IPDPS’11]

slide-14
SLIDE 14

Approximation

  • Approximate de-duplication
  • Approximate computation

– NPU 3.0X energy savings [Esmailzadeh 13]

  • Guided approximation with information

flow techniques

slide-15
SLIDE 15

3D Beamforming in Datacenters

  • Zheng and Zhao with Vahdat at Google
  • 60 Ghz links with 2-6 Gbps
  • Flexible BW for burst loads

[Sigcomm’12]

slide-16
SLIDE 16

Datacenter Placement

[Goiri et al, ICDCS’11] Electricity Rates Temperature Datacenter Placement Example Cost Breakdowns Cost of Green Datacenters

slide-17
SLIDE 17

Bridging the Gap

  • 40X in 13 years
  • Assuming 20% improvements can be

compounded:

– Need a new idea deployed every 7-8 months! – Probably much worse!

slide-18
SLIDE 18

Part II: Environmental Costs to Bridging the Gap

slide-19
SLIDE 19

Server = SUV

  • More precisely:

– 80 billion terawatt-hr / yr = 6 million SUVs in carbon production (10 mpg, 11K miles/yr)

=

slide-20
SLIDE 20

Warehouse Computing

  • Google at Columbia River Gorge =>

hydroelectric power

  • $30B annual energy bill worldwide
  • Energy starting to cost more than capital

expenditures =

slide-21
SLIDE 21

Resource Use in Silicon Fabrication

  • 1.6 kilowatt-hrs / cm2
  • 20 liters water / cm2
  • 3.3 billion active cell phone

subscriptions

  • (212 Billion wireless devices by 2020)

[IDC 13]

  • ~20 cm2 / phone
  • 106 billion kilowatt-hrs (recall that

datacenters use 80 billion kwh annually)

slide-22
SLIDE 22

Throughput

  • 280 Million phones sold / quarter
  • Average lifetime of a phone: 1.5-2 yrs
  • Old phones sitting in drawers, but

throughput of over 1 billion phones / yr

  • 32 billion kilowatt-hrs / yr just for uproc
slide-23
SLIDE 23

Other Impacts

  • 400 billion liters of water

– 160,000 olympic swimming pools – More than double annual global bottled water consumption

  • 400 million kg of soil to remediate just

the copper (more copper on surface than inside the earth!)

slide-24
SLIDE 24

Biodegradable Materials

  • Biodegradable plastics

– Fire retardants are bad

  • Organic LEDs / transistors
slide-25
SLIDE 25

Microprocessor Reuse?

  • Problem: obsolescence resulting from

rapid improvements

  • Solution: microprocessor food chain

[IEEE Computer ’07]

slide-26
SLIDE 26

Example Applications

500 1000 1500 2000 2500

  • Dig. Video Camera

PDA Set Top Box Cell phone Printer Automotive Nav. System Portable Game Systems Home Stereos Toys MP3 Players White Goods Home Tools BDTImark

The BDTImark2000(tm) is a summary measure of signal processing speed. For more info and scores see www.BDTI.com

BDTImark2000™

slide-27
SLIDE 27

Lifetime Energy Savings

  • Depends on die-size

– Die sizes are getting smaller

  • Depends on in-use

energy consumption

– Assume 3 hours of use per day

  • .5 W processors probably

should be re-used

  • 20 W processor, upgrade!

.5W Processor 20W Processor

Re-Use New Processor Every 2 Years New Processor Every 4 Years

1 2 3 4 5 6 7 8 9 20 40 60 80 100 120 140 Time (Years) Energy (MJ) 1 2 3 4 5 6 7 8 9 500 1000 1500 2000 2500 Time (Years) Energy (MJ)

slide-28
SLIDE 28

Technical Challenges of Re-Use

  • Form Factor

– Can’t put a Pentium in the space of an 8051

  • Battery Life

– Is adequate power consumption good enough? – Voltage scaling

  • ISA compatibility

– Some ISA are more efficient on specific workloads – May require extra cycles

  • Erode the efficiency of our re-use strategy
slide-29
SLIDE 29

Design for Reuse

  • Design for several applications and

lifetimes, not just one

  • More severe wearout
  • Added overhead to support different

applications

  • Design for easier reprogramming
  • Design for easier reclamation and re-

tasking

– form factor, wireless or serial communication

Standard building blocks

slide-30
SLIDE 30

Reclamation Costs

*Average (cell phone:$4 to $8; computer:$13 to $34) **Results from survey conducted with fifteen private US electronic recycling firms [ Bhuie et al, 2004] ***[Boon et al., 2000]

  • < $7 cell phone
  • Recycling surcharge + deposit
slide-31
SLIDE 31

Handset Reuse

  • Refurbished phones

– Only millions captured – Political issues

  • PDAs

– Learning tool / diary in elementary schools – Parking permit / navigator

  • Location beacon
  • Shipping container tracking
  • Just park benches?
slide-32
SLIDE 32

Reuse Summary

  • Silicon fabrication and disposal are

serious environmental concerns

  • Reuse is a challenging goal, but we

have to face the impact of our exponentionally-growing computing demands

slide-33
SLIDE 33

Part III: Cultural Change

slide-34
SLIDE 34

Cultural Change

  • Some sustainable

technologies and practices exist, but managers and designers unaccustomed to the tradeoffs

  • Need to develop frameworks

and educate the next generation of technical leaders

slide-35
SLIDE 35

Measuring Energy

  • Coal-fired electric plants – 35% efficient
  • Electrical transmission lines – 90%

efficient

  • Datacenter power distribution – also
  • ptimized for peak
  • Other inefficiencies

– Server power supplies – Battery charger / battery efficiency www.epa.gov/cleanenergy/energy-resources/ calculator.html

slide-36
SLIDE 36

Life-Cycle Analysis

  • Sustainable systems require a higher-

level analysis

– Energy and carbon metrics – Supply chains, end-of-life – Challenge: proprietary data – Study academic fabrication facilities – Make friends with your local industrial ecologist!

slide-37
SLIDE 37

Reacting to Policy

  • Standards and policies
  • Energy-star, SPEC power
  • Standards need knowledgeable

participants

  • Companies need to know how to

respond to legislation and standards

  • WEEE, RoHS, Energy-star
slide-38
SLIDE 38

Caveat: Jevons Paradox

  • Efficiency in coal-fired

machines led to greater demand for coal

slide-39
SLIDE 39

Jevons Paradox

  • Demand for computing could

be elastic

  • Need to measure productivity
slide-40
SLIDE 40

Closing Remarks

  • Computing for massive data poses

significant sustainability challenges

  • Good technical problems, but many

are multidisciplinary

  • We need to train the next generation
  • f multidisciplinary engineers

energy.cs.ucsb.edu

slide-41
SLIDE 41

Acknowledgements

  • Luiz Barroso, Urs Hoezle, Bill Weihl

(Google)

  • Partha Ragananthan (Google)
  • Ricardo Bianchini (Rutgers)
  • Roland Geyer (UCSB)
  • Raj Amirtharajah, Venkatesh Akella

(UC Davis)

  • John Oliver (Calpoly)