Digital On-Demand Computing Organism Dod Org SPP OC Kolloquium DFG SPP 1183 “Organic Computing” München, Oktober 7/8, 2010 KIT – The cooperation of Forschungszentrum Karlsruhe GmbH and Universität Karlsruhe (TH)
Talk Overview Project Motivation and Overview Current Work: Organic Hardware Organic Middleware Organic Low Power Management Organic Middleware Demonstrator Platform Overview Scenarios Conclusion SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 2 7.10.2010
DodOrg Motivation DodOrg Scenario: Classic Scenario: Only those scenarios can be handled System reaction based on indications (higher level of abstraction) that were considered in advance, where the cause can be detected, e.g. CRC/bit error rate, network bottleneck, environmental change or change on where the corresponding reaction application level had been explicitly programmed. P roper reaction possible even if Lack of adaptation leads to insufficient Scenario was not considered in advance. reactions (e.g. shutdown …) Cause was not detected, Reaction was not explicitly programmed. F lexible response to changed environmental situation Scenario detection: recognize that something is different Adapt to changed requirements either by known path or gradual process of Demonstrator platform rearrangement (optimization, healing) Plasticity: Stabilization but not “ petrification ’’ SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 3 7.10.2010
Phase III: Project Objectives Stability Robustness The ability of the system to Extending the stable system provide the required service property towards more serious while reacting upon external system changes . and internal events. + Attack resistance + Oscillation avoidance + Fault resistance + Normal operating conditions + Increased tolerance - Faulty components - Increased overhead - Malicious attacks SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 4 7.10.2010
DodOrg: Refined Layer Model Brain Level Self-Adaptation Application Application Testbed Nervous Self-Optimization Monitoring System (all groups) Self-Healing Biological Considerations (Brändle) Application API Organic Monitoring System (Karl) Proactive Intelligent Data Analysis Middleware Monitoring, Organic Middleware Stable Hormone Organ Level Feedback Interaction (Brinkschulte) Heart Hormone Level Computation Temperature, Thermal-aware Distributed Low Power Local Traffic Energy Management (Henkel) distribution Dynamic Power Management Cell Level Real-time considerations Myo- cardial Hardware OPC Extension Organic Processing Cell Monitoring Cells (Becker) Stabiltiy Aspects SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 6 7.10.2010
Organic Processing Cells (OPCs): Chip To Chip Communication (Prof. Becker) Goals Seamless and transparent expansion of the on chip communication services (artNoC) Dynamic OPC resource pool physical growth of DodOrg organism Challenges System 1 System 1 System 2 System 2 Physical Connection # I/O Pins IO IO µProc µProc IO IO Cell Cell Cell Cell Cell Cell Bandwidth Latency IO IO Hot Plug /Unplug FPGA FPGA IO IO Cell Cell Cell Cell Cell Cell Control Flow Overhead Transparent Transparent Dynamic Address Space Extension Extension IO IO µProc µProc IO IO Cell Cell Cell Cell Cell Cell Broadcast Broadcast Real-time Real-time Adaptive routing Adaptive routing SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 7 7.10.2010
Organic Processing Cells (OPCs): Chip To Chip Communication (Prof. Becker) Off-Chip-Interface (OCI) R R Parallel Serial R R Asynchronous Transmission Off-Chip Interface artNoC Handshake Signals Chip Chip 38 2 38 2 artNoC Flit-Data 3 Phase Operation 38 2 38 2 R R R R R R R R Link Negotiation 38 2 38 2 Backchannel Detection R R R R R R R R # Virtual Channels (VCs) # Real-Time Channels Data Transmission 38 2 38 2 Link Cutback Terminate open artNoC VCs R R R R SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 8 7.10.2010
Organic Processing Cells (OPCs): Chip To Chip Communication (Prof. Becker) Data Transmission: Adaptive Control Flow Protocol Router_a Router_b Goal: Increase Serial Link Efficiency port_in_b port_out_a OCI_a OCI_b Differential Control Flow Transmission data(7:0) data(7:0) 1 VC used: P(C ch ) < 0,1 vc_id(1:0) vc_id(1:0) 1 > 1 VC used: P(C ch ) > 0,5 valid valid Amount of Control Flow mainly depends on VC usage full(3:0) full(3:0) rtc_ready(2:0) rtc_ready(2:0) 2 Protocol Modes bc_grant bc_grant Differential Mode port_in_a port_out_b Auto Sequence Mode 90 data(7:0) data(7:0) 80 vc_id(1:0) vc_id(1:0) 1 70 link efficiency % valid valid 60 full(3:0) full(3:0) 50 rtc_ready(2:0) rtc_ready(2:0) 40 Diff Mode bc_grant bc_grant 30 Auto Mode 20 10 0 1 VC 2 VC's 50/50 3 VC's 2 VC's 10/90 50/20/30 virtual channel usage SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 9 7.10.2010
Organic Monitoring: State Classification and Evaluation (Prof. Karl) Objectives Application Application Testbed Self-Adaptation Monitoring (all groups) Classification of the system state Self-Optimization Biological Considerations (Brändle) Proactive Intelligent Data Analysis Application API Organic Monitoring System (Karl) with regard to environmental Self-Healing Middleware Monitoring, Organic Middleware Feedback Stable Hormone conditions (Brinkschulte) Interaction Hormone Level Computation Identification of bottlenecks Temperature, Distributed Low Power Application Hardware Stable Energy Local Traffic Management (Henkel) Determine the outcome of an Distribution Dynamic Power Management Real-time considerations optimization cycle Requirements Hardware Organic Processing Cells Monitoring OPC Interaction, (Becker) Status Providing these information to Status Metrics Organic Middleware and Thermal Plasticity Aspects Management Monitoring Raw & Cooked Flexible, rule-based Approach Configuration Feedback Data Based on the data gathered by our low-level monitoring-infrastructure Evaluation rules are defined at runtime in a dedicated learning phase Middleware Thermal Management Rules can be updated at runtime to adapt to new environmental conditions SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 10 7.10.2010
Organic Monitoring: State Classification and Evaluation (Prof. Karl) Rule Layout rule = (t, v, p1, p2, p3) One Rule for each event type Rule Creation Determine the normalized occurence ratio for each event type at a predefined time Creation of a histogram Determination of the points p1, p2, p3 Evaluation Using a simple transfer function (TF) TF converts the ratio into an evaluation score System State is classified using a weighted arithmetic mean of all evaluation scores SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 11 7.10.2010
Organic Monitoring: State Classification and Evaluation (Prof. Karl) Rule Layout rule = (t, v, p1, p2, p3) One Rule for each event type Rule Creation Determine the normalized occurence ratio for each event type at a predefined time Creation of a histogram Determination of the points p1, p2, p3 Evaluation Using a simple transfer function (TF) TF converts the ratio into an evaluation score System State is classified using a weighted arithmetic mean of all evaluation scores SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010v 12 7.10.2010
Organic Monitoring: Outlook (Prof. Karl) Phase/ Trend Detection and Prediction Prediction of future system states Identification of potentially harmful system states in advance Avoiding Bad System States through Proactivity Initiating proper system changes to avoid bad or harmful system states (e.g. high temperature or performance bottleneck) Introducing a feedback-loop for on-line evaluation of the system changes Proactive self-healing and self-optimization Avoid this state Current Predicted Past System States System State System States SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 13 7.10.2010
Organic Low Power Management: Managing Energy-Distribution (Prof. Henkel) Energy Distribution: goals Power source Organic Middleware Low energy consumption Energy Influencing Assigned Avoidance of local thermal hot-spots Input Hormone Expression Tasks Energy Distribution: main concept Peers (in neighbored OPCs) Cost Function Local traffic Temperature Each OPC has a Local Energy Budget Energy level Efficiency Organic Monitoring Determines the local available Trade & RT criteria Future energy level energy Negotiate Energy Budget Actual energy level Power / RT Manager Manager Global Power Source Actual power state Assigns energy budgets to OPCs Policy Fill Consume (pulse-based) P 1 Energy Budget Manager P 2 Fade Power Local Energy P 3 Agent controlling Local Energy States Budget P 4 Budget Receives temperature data each Voltage / frequency Scheduled pulse Legend: setting Tasks Negotiates & Trades energy data / information OPC actions budget with neighboring OPCs Influences Power Manager policies SPP 1183 OC Kolloquim – München, 7.-8. Oktober 2010 14 7.10.2010
Recommend
More recommend
