The Research Idea of HAEC: Highly Adaptive Energy-Efficient - - PDF document

the research idea of haec highly adaptive energy
SMART_READER_LITE
LIVE PREVIEW

The Research Idea of HAEC: Highly Adaptive Energy-Efficient - - PDF document

Gerhard P. Fettweis Coordinator W. Lehner, W. Nagel, C. Baier, D. Plettemeier 29.05.2015 Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing 3 Collaborative Research Center 912: HAEC Highly Adaptive


slide-1
SLIDE 1

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

29.05.2015

Gerhard P. Fettweis – Coordinator

  • W. Lehner, W. Nagel, C. Baier, D. Plettemeier

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

3

The Research Idea of HAEC: Highly Adaptive Energy-Efficient Computing

slide-2
SLIDE 2

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

3

Motivation

0 % 50 % 150% 100 %

ICT

2010 2030

Energy Consumption with Reference to 2010 Grid

Current trend: 10x in 20 years

?

Current trend: 10x in 20 years 200%

Non-ICT 90% Non-ICT

ICT 10%

>50% of 2030 grid

HAEC Challenge: enabling a massive-magnitude reduction!

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

4

Energy Proportionality/Efficiency

Goal:

Minimizing Energy

by

Multi-Layer SW/HW Adaptivity

Center for Information Services and High Performance Computing (ZIH) Measurement at June 20, 2008 Percentage Time of Day

  • Rel. Energy Consumption
  • Rel. Computing Load
slide-3
SLIDE 3

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

The Communication Bottleneck

  • f Parallel Computing

5

Parallelism = 2

  • Comm. Links =

1 P = 4 C = 6 P = 8 C = 28 P = 16 C = 120 P = 32 C = 496 P = 64 C = 2016

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Highly Adaptive Energy-Efficient Computing

High-Rate Inter-Chip Communications

6

Radio Interconnect

  • n-interposer/on-package

antenna arrays analog/digital beam steering and interference minimization 100Gb/s 25 GHz channel @ 200GHz carrier 3D routing & flow management

R di I Optical Interconnect

adaptive analog/digital circuits for e/o transceiver embedded polymer waveguide packaging technologies (e.g. 3D stacking of Si/III-V hybrids) 90° coupling of laser

slide-4
SLIDE 4

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Today’s energy optimizations

  • Are considered on a component level
  • Mostly in isolation
  • Unaware of applications and contexts

Goal of project group B

  • Address energy-efficiency at all system levels in an integrated holistic manner
  • Ranging from the OS to software components to virtual machines

Energy Optimization

7

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Control

Energy-Adaptive Software Architecture

Global QoS optimization with “Energy-Utility Functions” Cross-layer adaptation: application - system - hardware

Monitoring

Energy-Optimization Control Loop

Measurement Analysis & verification Strategic regulation Adaptation

Software Architecture Energy-adaptive SW stack Run-time Optimization

Focus on energy efficiency Using context information

Hardware Architecture

New communication channels

HAEC B – Closed Energy-Control Loop

8

slide-5
SLIDE 5

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Example of Adaptivity by Software

  • Three in-memory database services
  • All are using a scan to access their data

Under-utilized cores and interconnects

9

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Example of Adaptivity by Software

1) Exchange software implementations (dbscan index scan) 2) Migrate database service 4) Turn off unnecessary interconnects 3) Migrate memory

Hardware and software adaptations

Necessary for optimal energy-efficiency/proportionality

How to optimize?

10

slide-6
SLIDE 6

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

1) Partition memory and activate additional sockets and interconnects 2) Add more computing power and activate necessary interconnects

Trade Energy for Utility/Capacity How to gain (more/less) utility from (more/less) energy ?

Example of Adaptivity

11

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Highly Adaptive Energy-Efficient Computing

12

Devices & Circuits Information Processing

Application Middleware Runtime Operating System Processor

Flexible Hardware Interconnects Flexible Software

Energy Control Loop

Simulation & Prototyping HAEC Box

slide-7
SLIDE 7

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

15

Highlights of Phase I (2011-2015)

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

HAEC A

20

HAEC Architecture HAEC Software

Butler Matrix LNA PA Mixer ADC DAC Base- band Device Modelling

Wireless Optical

LD MZM PD TIA LDD Base- band

System Integration (Packaging)

slide-8
SLIDE 8

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Selected HAEC A Highlights

21

HAEC Architecture HAEC Software

Device Modelling

Optical

LD MZM PD TIA LDD Base- band

System Integration (Packaging)

TIA with Adaptivity: Bandwidth vs DC power 50% Energy Transimpedance Amplifier (TIA) 90 Gbit/s 0,23 pJ/bit 0,16 mm2 Embedded waveguides

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

22

HAEC Architecture HAEC Software

Butler Matrix LNA PA Mixer ADC DAC Base- band Device Modelling

Wireless System Integration (Packaging)

A05 Bow-tie antenna (1mm x 0.6 mm) A05 Half clover- leave antenna (1.2mm x 0.7 mm) A01 200 GHz 18 mW LNA (low noise amplifier) A01 200 GHz 18 mW mixer A01 24GSps 3bit single core flash ADC

Selected HAEC A Highlights

slide-9
SLIDE 9

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

slide 23

slide # should not contain “slide”

Selected HAEC A Highlights: Artemis

Shunt-based instrumentation for energy measurements in a state-of-the-art server

1300 1800 2300 2800 3300 1 1,05 1,1 1,15 1,2 processor frequency (MHz) relative energy consumption join subgroup subselect leftfetchjoin projectdelta leftfetchjoinPath

Energy monitoring Energy modeling

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

Highlights Software HAECubie

slide 24 Selected HAEC B Highlights: HAECubie

HAECubie

adaptations

Resource Management (B04)

sensor data

Database (B05)

system information

Knowledge Plane

update data

System Management

Software Architecture (B01) Service Execution (B06) synopses Context Recognition (B02) es Energy-Utility Analysis (B03)

trade-offs

aggregated data coordination

(B02)

context context

slide-10
SLIDE 10

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

The Outlook: The Box in 2030+

Assume 64K processors per chip double-sided 4x4 chip-stacks on board 160x chips stacked in 3D 4x boards in a box in 10x10x10 cm3 (1 liter) 108 processors! 1.6K processor + 16K memory chips 104x performance of today!

Box

30

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

HAEC Box Embedded Everywhere

31

slide-11
SLIDE 11

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

32

Adaptivity Energy proportionality Energy efficiency Revolutionary approach to computing hardware/software Impact onto more than computing applications Chance for scalability of the box and the software approach Chance for transfer

High-Level Highlights / Conclusions

Collaborative Research Center 912: HAEC Highly Adaptive Energy-Efficient Computing

33

slide-12
SLIDE 12