Digital On-Demand Computing Organism - DodOrgprojects.aifb.kit.edu/effalg/oc/inhalte/FilesColloquiumAugsburg/Karl_DodOrg.pdfScalability Application API Application Monitoring Hardware

KIT – The cooperation of Forschungszentrum Karlsruhe GmbH and Universität Karlsruhe (TH)

Digital On-Demand Computing Organism - DodOrg

SPP OC KolloquiumDFG SPP 1183 “Organic Computing”Augsburg, September 21/22, 2009

Talk Overview

►Project Motivation and Overview►DodOrg: Plasticity and Dynamics ►Results of third year:

Organic MonitoringOrganic MiddlewareOrganic Low Power ManagementOrganic Hardware

►Conclusion Phase II►Research Goals for Phase III

StabilityRobustness

►Phase III Summary

21.09.20092 SPP 1183 OC Kolloquim – Augsburg, 21.-22. September 2009

DodOrg Motivation

Classic Scenario:►Only those scenarios can be handled

that were considered in advance,where the cause can be detected,where the corresponding reaction had been explicitly programmed.

►Lack of adaptation leads to insufficient reactions (e.g. shutdown …)

DodOrg Scenario:►System reaction based on indications (higher level of abstraction)

e.g. CRC/bit error rate, network bottleneck, environmental change or change on application level

►Proper reaction possible even ifScenario was not considered in advance.Cause was not detected, Reaction was not explicitly programmed.

►Flexible response to changed environmental situation

Scenario detection: recognize that something is differentAdapt to changed requirements either by known path or gradual process of rearrangement (optimization, healing)Plasticity: Stabilization but not “petrification’’


Phase II: Refined Layer Model

Brain

Lev

elOr

gan

Leve

lCe

ll Lev

el

Myo-cardial

Cell

Nervous System

Application API

Application Monitoring

HardwareMonitoring

MiddlewareMonitoring,Feedback

Application Testbed(all groups)

Organic Middleware(Brinkschulte)

Organic Processing Cells (Becker)

Distributed Low Power Management (Henkel)

Biol

ogica

l Con

sider

atio

ns (B

ränd

le)

Orga

nic M

onito

ring

Syst

em (K

arl)

HeartHormone Level Computation

Dynamic Power ManagementReal-time considerations

Temperature,Local Traffic

Proa

ctiv

e In

telli

gent

Dat

a A

naly

sis

Self-AdaptationSelf-Optimization

Self-Healing

Stable HormoneInteraction

Stable EnergyDistribution

OPC Interaction,Metrics

Plasticity Aspects


Phase II: Project Objectives

PlasticityNarrowing the dynamic properties.The ability of the system to leave a stable state in order to adapt to larger environmental changes.

DynamicsThe ability of the system to react according different parameter changes (external, internal).

+ Decreased power consumption+ Oscillation avoidance+ Overhead reduction + Reduced complexity- Delayed reaction

+ Improved performance+ Immediate reaction- Danger of oscillation- Unstable behavior


Example of Plasticity

Scenario►Link Failure Packet Deadlock►Recovery Broadcast to deliver stuck packet►Implications

Hormon cycleNetwork topology because of broken link neighbor relationship changes

►MW redistribution of task►Different Link Workload►Power Management Regulation of new Power Situation


OPC-Swap

Scenario

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC

OPC

OPC

OPC

OPC OPC OPC OPC OPC

OPCOPCOPC OPC OPC

OPC

OPC

OPC

OPC

OPC

OPC

OPC OPC OPC OPCOPCOPC

OPC

OPC

OPC

OPC

OPC

OPC

OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC

OPC

OPC

OPC

OPC OPC OPC OPC OPC

OPCOPCOPC OPC OPC

OPC

OPC

OPC

OPC

OPC

OPC


OPC

OPC

OPC

OPC

OPC

OPC

OPC

No failure

Aver

age

pack

et la

tenc

y (b

lue

orga

n)

Optimization

time

Related tast group= organ

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC OPC OPC OPC

OPC

OPC

OPC

OPC

OPC OPC OPC OPC OPC

OPCOPCOPC OPC OPC

OPC

OPC

OPC

OPC

OPC

OPC


OPC

OPC

OPC

OPC

OPC

OPC

OPC

Detour Routing

Link error

Recovery

OPC

OPC

OPC


Organic Monitoring: Overview(Prof. Karl)

►AimEnable and support Self-X capabilitiesFocus on increased self-awareness

►RequirementsSustained system monitoringReal-time analysis and evaluation

Correlation of (many) eventsIdentification of problems/causes

Semantic data compressionProcessing at the source of dataGeneration of meta-data

Adaptivity (reconfiguration)InterfacingScalability

Application API

Application Monitoring

HardwareMonitoring

MiddlewareMonitoring,Feedback



Organic Processing Cells(Becker)

Distributed Low Power Management (Henkel)

Biological Considerations (Brändle)

Organic Monitoring System (Karl)

Hormone Level Computation

Dynamic Power ManagementReal-time considerations

Temperature,Local Traffic

Proactive Intelligent Data A

nalysis

Self-Adaptation

Self-Optimization

Self-Healing

Stable Hormone

Interaction

Stable Energy

Distribution

OPC Interaction,

Metrics

Plasticity Aspects

Application Hardware

Monitoring

Raw

& C

ooked

Data

Feedback

Configuration

Requirem

ents

Status

Status

Middleware

Low-Power Planning


Organic Monitoring: Event Monitoring and Aggregation(Prof. Karl) Hormone-Inspired

Event Specification and Transmission

Meta-Events

►Decentral Event Monitoring and AggregationBased on the hormone concept, Organic Monitoring Modules (MM) monitor the hormone concentration in each OPCEach hormone is treated as an individual eventEvents are aggregated in an Associative Counter Array (ACA)

Association of events to countersCache principleEvent transmission upon overflow or replacement

Events transmitted are forming so-called Second Messenger or Meta-EventsMeta-Events are then transmitted to High-Level instances for data analysis

Event Counter Modulo

Hormones

EventAggregation


Organic Monitoring: Intelligent Data Analysis Techniques(Prof. Karl)

►State Classification in High-Level Monitoring

Several High-Level Monitoring Units (HLM) operate on Meta-Events generated by several Organic Monitoring modulesIncoming Meta-Events are stored in so-called Event Lists which are sorted by hormone typeState Classification algorithms are using one or more Event Lists to classify the state of the (sub-) system as good, neutral or badIf necessary, appropriate actions are triggered by HLM instances

►Application Scenario – Power ConsumptionThe power consumption of OPCs is monitored by dedicated hardware sensorsScenario: encoding of a wav-file using lameA hormone is generated each time 10 mW are consumed by the CPUState Classification every 16 Meta-EventsSimulation time: ~ 4 Billion Cycles

~ 300 Mio. HormonesTraining Phase: 1 000 000 CyclesResults: 282,292 state classifications

Used for Training: 76Good: 242,590 Neutral: 35,096Bad: 4,530


Organic Middleware: Re-Intruduction of the Artificial Hormone System(Prof. Brinkschulte)►Aim:

Mapping tasks on Organic Processing Elements (OPC)Providing the system with self-X properties on the middleware layer:

Self-ConfigurationSelf-HealingSelf-Optimization

Achieving a good mapping inregards to

Requirements of each taskRelationships of the tasksCondition of each cell and it’s neighborhood

Reacting and Adapting to changes (plasticity)

e.g. increased bit-rate errorsReaching stable mapping conditions

R

OPCR

OPC

R

OPCR

OPC

R

OPCR

OPC

R

OPCR

OPC

Organic Middleware: New Challenges of the Second Phase(Prof. Brinkschulte)

Hormone Concept - Evaluation and Refinement:Analyzing the network load in the systemGenerating hormone configurations with evolutionary algorithms

Stability:Stability of the system of individual OPCs with Hormone Cycles

a bounded input (configuration + measured data) results in an bounded outputno oscillation of tasks and no unnecessary task allocations

Examination of Stability & Robustness: Finding limits of systems:

How many tasks/jobs are suitable for a given number of OPCsCalculating minimum requirements of the system:

How many OPCs will be needed for a given scenario (at a minimum)


Organic Middleware: Phases of the Configuration(Prof. Brinkschulte)

Phase 1:Strong Eagervalues, weak Suppressors and weak Accelerators every cell wants to execute tasksSuppressors rise when tasks are mapped

Phase 2:Strong Suppressors, Eagervalues to zero System will remain in a stable stateReconfiguration and Self-Optimization (fluctuation of the hormones)


Organic Low Power Management: DodOrg Interfaces(Prof. Henkel)

Cost Function

Organic MiddlewareInfluencing

Hormone Expression

Power / RTManager

Organic M

onitoring

Consume

Fade

Trade &Negotiate

Policy

Energy BudgetManager

Local EnergyBudget P4

P3

P2

P1

Fill

EnergyInput

EfficiencyRT criteria

Temperature

Local traffic

Power sourceP

eers

(in

neig

hbor

ed O

PC

s)

OPC

Voltage / frequencysetting

PowerStates

AssignedTasks

ScheduledTasks

data / informationactions

Legend:

Future energy levelActual energy level

Energy level

Actual power state

► Organic MiddlewareCost Function

Used for computaion of local eager values

► Organic MonitoringExternal Cost Function parameters

Used to select apt power management policy

► OPCConfigure Power State


Organic Low Power Management: Managing Energy-Distribution(Prof. Henkel)

Cost Function

Organic MiddlewareInfluencing

Hormone Expression

Power / RTManager

Organic M

onitoring

Consume

Fade

Trade &Negotiate

Policy

Energy BudgetManager

Local EnergyBudget P4

P3

P2

P1

Fill

EnergyInput

EfficiencyRT criteria

Temperature

Local traffic

Power sourceP

eers

(in

neig

hbor

ed O

PC

s)

OPC

Voltage / frequencysetting

PowerStates

AssignedTasks

ScheduledTasks

data / informationactions

Legend:

Future energy levelActual energy level

Energy level

Actual power state

► Energy Distribution: goalsLow energy consumptionAvoidance of local thermal hot-spotsConvergent system behavior (plasticity)

► Energy Distribution: main conceptEach OPC has a Local Energy Budget

Determines the local available energyGlobal Power Source

Assigns energy budgets to OPCs (pulse-based)

Energy Budget ManagerAgent controlling Local Energy BudgetNegotiates & Trades energy budget with neighboring OPCsInfluences Power Manager policies


Organic Low Power Management: Agent Negotiation(Prof. Henkel)

OPC requiresenergy to rundemanding task

OPC negotiates energybudget with neighborsusing cost function basedon supply and demand

OPC

Local EnergyBudget

Energy Flow

Energy Cost

0/3

0/3 0/3

0/3 0/30/3

0/3 0/3

0/3

Energy units used/free

0/3

0/1 0/3

7/7 0/30/3

0/1 0/3

0/3

Energy units used/free

0/2

0/2 0/3

7/0 0/30/3

0/2 0/3

0/2


Organic Low Power Management: Power Trading – Getting buyand sell Values(Prof. Henkel)

Sell value:

Buy value:

Sell value must also consider running tasks:

Sell value

Buy value

Wan

ts to

buy

Wan

ts to

sel

l

Free units (used units constant)

1 2 3 4 5

Threshold atstable-state

Weights and γ are dependant on OPC type, total amount of energy units and number OPCs

pi is priority of task i

sell

-buy

sell

-buy

Agent of OPC n sells to neighbor i if:(selln – buyn) – (selli – buyi) > threshold


Organic Low Power Management: Power Trading – Agent Communication(Prof. Henkel)

0

10

20

30

40

50

60

70

80

90

100

1 5 9 13 17 21 25

time (pulses)

Tota

l age

nt c

omm

unic

atio

n

1 taskmapped

2 tasksmapped

3 tasksmapped

zero communication indicates stable state is reached


Organic Processing Cells Work plan: Close Control Loop Effects, Metrics, Cost Functions (Prof. Becker)

Foundation laid byDNA-configuration controlAdaptive routingAutomated test systemHardware prototype

ChallengesDynamics of cell interactionInterference with Middleware/Low-Power Management

Research GoalsMetricsOptimizationPlasticity

Optimization & Evaluation• HW Structure• Cell Interaction• Guarantees, Stability

Comparison to classical approaches

Metrics• Power• Availability• Performance• …

Abstract OC-HW Model

Library ofOC-CellsLibrary ofOC-CellsLibrary ofOC-Cells

Input: Organic Hardware Architecture (Phase I)

Determine level ofsystem abstraction


Organic Processing Cells: Close Control Loop Effects: Example Link Fault Detection (Prof. Becker)

Analysis of neighbor broadcast Broadcast uses flooding Assumption on fault free link: receive neighbor BC over adjacent link

BC-T

artNoC 1 artNoC 2

artNoC 3 artNoC 4Link Fault None

Routing Status Fault free

Link Fault None

Routing Status Fault free

C4

C4

C4

C4

South

C2

Detour S

C2

OPC-Status Interface

OPC-Status Interface

Broadcast-TokenBroadcast messagefrom cell 4

BC-T

Broken link

include statusBC-Round 1

C4

BC-Round 2

C2


Accu

mulat

ed ad

dition

al ha

rdwa

re co

st pe

r OPC

[TSM

C 13

0]

Poten

tial fl

exibi

lity/ fa

ult to

leran

ce

Organic Processing Cells: OPC-Hardware Cost –Synergetic Effects (Prof. Becker)

Basic OPC;µC-DP, xy-routing artNoC

Broadcast and Multicast

Error recovery;deadlock, livelock, dead end

Full adaptive routing

Fault aware routing

Observer; Link failure detection,deadlock detection

Status interface;OPC node and link status,OPC heartbeat

Clock and Power Management

Observer ;temperaturemonitoring

One cycle clock switch

Adaptive routing; West First

0,024

mm2

0,007

mm2

0,016

mm2

0,01 m

m2

0,005

mm2

0,02 m

m2

0,04 m

m2

FPGA

Impl.

FPGA

Impl.

FPGA

Impl.


Organic Processing Cells : Xilinx-Virtex-II Pro Hardware Prototype (Prof. Becker)

Power Measurement Setup

►2D-Partial and dynamic reconfiguration of FPGA-datapath

►OPC-based distributed clk-management

►Core OPC functionality

►Derive cost and performance figures for OPC HW-model


Organic Processing Cells : OPC Power Model (Prof. Becker)

Power Down

Change Power Mode

Unspecified/Empty

ReconfiguringPower Up

Specified/Programmed

Idle

Processing

UnconnectedConnected

OPC- Model

Cp = 0 mW

Cp = 76 mW

Cp 100 MHz = 100 mW

Cp100 MHz = 120 mW

Cp = 10 mW

Cp = 86 mW(Power Up/Down)

Change Power Mode(Clock Scaling)

Cp = 82 mW

Cp 100 MHz = 92 mW

Cp 100 MHz = 100 mW

40µs40µs

880 µs/Col


Conclusion and Plans for Phase III

►Outcome of Phase IIConcepts individually tested and applicability provenMonitoring: hormone-inspired associative event coding and use of associative countersMiddleware: reaching stable hormone and mapping situations while still being able to react to changes (plasticity)Low-Power-Processing: local agent-based energy budget distributionProcessing Cells: abstract OC-hardware model and evaluation of cell interaction metrics and cost-functions


Outlook: Stability Phase III

►Phase IIISelf-optimization scenariosConflict avoidance throughproactivityCell-level low-power issues and interactionOff-chip communicationRobustness during Development/Processing PhasePrototype implementation: application phase detection and fault-tolerance


Organic Monitoring: Outlook(Prof. Karl)

►Trend DetectionPrediction of future system statesIdentification of potentially harmful system states in advance

►Avoiding Potential Conflicts through ProactivityBased upon Trend Detection and Event CorrelationInitiating required system chances to avoid bad or harmful system states (e.g. high temperature)Proactive self-healing and self-optimization

Past System States CurrentSystem State

PredictedSystem States

Avoid this state


Organic Middleware: Outlook of the Third Phase(Prof. Brinkschulte)

► Robustness and StabilityExamination of dynamic aspects of task (re-)allocation and operation during normal conditions (also in the presence of internal or external disturbances)System changes due to Self-Adaptation and Self-Optimization must lead to new stable conditionsRobustness against mal-behaving internal/external components (comparable to illness in a biological system)

Being able to react to „ill“ OPCsCounter-measures against malicious attacks

Immune mechanisms for advanced Self-Healing and Self-Protecting Aspects to increase Robustness (together with the Organic Processing Cell Capabilities and the Organic Monitoring)Examination and Evaluation of different Self-optimization ScenariosQuality Analysis of the Artificial Hormone System


Organic Low Power Management: Outlook(Prof. Henkel)

Thermal test simulation

Thermal issues are a major concernfor the robustness of the DodOrg System and are directly influenced bythe power distribution

Thermal models need to be developed andincorporated in the power negotiation processImprovement of negotiation process using econ-omic learning

Thermal-aware power negotiation21.09.200928 SPP 1183 OC Kolloquim – Augsburg, 21.-22. September 2009

Organic Processing Cells: Outlook Phase III Stability and Robustness (Prof. Becker)

OPC-Lifecycle

Development Phase:

Reconfiguration

Processing Phase:

Calculation

Global View: Phases are concurrent within DodOrg Cell-Array

Ongoing Change

Adaptation Mechanisms

Resulting Configurations


Organic Processing Cells: Outlook Phase III Goals (Prof. Becker)

►Chip To Chip CommunicationSeamless and transparent expansion of the on chip communication servicesDynamic OPC resource pool physical growth of the whole DodOrg organism New challenges for all subprojects

►Robustness on hardware-level during development phaseScope

Loading new ConfigurationEstablish-Inter-Cell-DatapathPower Up/Down Cells

Goal : Reach Fail Safe State after development phase ►Robustness on hardware-level during processing phase

ScopeOPC-Datapath (packet sender)OPC to OPC communication Path ( artNoC-Network)

Goal: Cell immune System with Cell-Input/Output-Guidance Mechanism


Questions?

Thank you for your attention!




Organic Processing Cells(Becker)

Organic Low Power Management (Henkel)

Orga

nic M

onito

ring S

ystem

(Kar

l)

List of Publications:


• D. Kramer, R. Buchty, and W. Karl, “A Scalable and Decentral Approach to Sustained System Monitoring“, ACACES,2009

• R. Buchty and W. Karl, “Design Aspects for Self-Organizing Heterogeneous Multi-Core Architectures“, IT - Information Technology Journal 5/08, 2008

• R. Buchty, D. Kramer, and W. Karl, “An Organic Computing Approach to Sustained Real-time Monitoring“, BICC08, 2008

• R. Buchty, O. Mattes, and W. Karl, “Self-aware Memory: Managing Distributed Memory in an Autonomous Multi-master Environment,“ ARCS, 2008

• R. Buchty and W. Karl, A Monitoring) “Infrastructure for the Digital on-demand Computing Organism (DodOrg)“, IWSOS, 2006

• Hans-Peter Löb, Rainer Buchty, Wolfgang Karl, “A Network Agent for Diagnosis and Analysis of Real-time Ethernet Networks“, CASES, 2006

• U. Brinkschulte and A. von Renteln, “Analyzing the Behavior of an Artificial Hormone System for Task Allocation”, ICATC, 2009

• U. Brinkschulte , A. von Renteln, and M. Weiss, “Examining Task Distribution by an artificial hormone system based middleware”, ISORC, 2008

• U. Brinkschulte, M. Pacher and A. von Renteln, “An Artificial Hormone System for Self-Organizing Real-Time Task Allocation”, in Organic Computing, 2007

• U. Brinkschulte, A. von Renteln, and M. Pacher, “Reliability of an Artificial Hormone System with Self-X Properties”, PDCS, 2007

• T. Ebi, M. A. Al Faruque, and J. Henkel, “TAPE: Thermal-aware Agent-based Power Economy for Multi/Many-Core Architectures”, ICCAD 2009 (accepted)

• M. Shafique, L. Bauer, and J. Henkel, “REMiS: Run-time Energy Minimization Scheme in a Reconfigurable Processor with Dynamic Power-Gated Instruction Set” , ICCAD 2009 (accepted)

• M. A. Al Faruque, R. Krist, J. Henkel: ”ADAM: Run-time Agent-based Distributed Application Mapping for on-chip Communication", DAC 2008

• C. Schuck, B. Haetzer, and J. Becker, “An Interface for a Decentralized 2d-Reconfiguration on Xilinx Virtex-FPGAs for Organic Computing“, ReCoSoC, 2008

• C. Schuck, M. Kuehnle, M. Huebner, and J. Becker, “A framework for dynamic 2D placement on FPGAs“ , IPDPS, 2008

• C. Schuck, S. Lamparth, J. and Becker, ”artNoC - A Novel Multi-Functional Router Architecture for Organic Computing”, FPL, 2007

Documents

Digital On-Demand Computing Organism - DodOrgprojects.aifb.kit.edu/effalg/oc/inhalte/FilesColloquiumAugsburg/Karl_DodOrg.pdfScalability Application API Application Monitoring Hardware