In-Network Management. Rolf Stadler. Stockholm, Sweden



Similar documents
Green Master based on MapReduce Cluster

A Parallel Transmission Remote Backup System

Load Balancing Algorithm based Virtual Machine Dynamic Migration Scheme for Datacenter Application with Optical Networks

AN ALGORITHM ABOUT PARTNER SELECTION PROBLEM ON CLOUD SERVICE PROVIDER BASED ON GENETIC

Dynamic Provisioning Modeling for Virtualized Multi-tier Applications in Cloud Data Center

The impact of service-oriented architecture on the scheduling algorithm in cloud computing

IDENTIFICATION OF THE DYNAMICS OF THE GOOGLE S RANKING ALGORITHM. A. Khaki Sedigh, Mehdi Roudaki

Efficient Traceback of DoS Attacks using Small Worlds in MANET

Discrete-Event Simulation of Network Systems Using Distributed Object Computing

How To Make A Supply Chain System Work

Cyber Journals: Multidisciplinary Journals in Science and Technology, Journal of Selected Areas in Telecommunications (JSAT), January Edition, 2011

A particle Swarm Optimization-based Framework for Agile Software Effort Estimation

Statistical Pattern Recognition (CE-725) Department of Computer Engineering Sharif University of Technology

Automated Event Registration System in Corporation

AnySee: Peer-to-Peer Live Streaming

Optimal replacement and overhaul decisions with imperfect maintenance and warranty contracts

A DISTRIBUTED REPUTATION BROKER FRAMEWORK FOR WEB SERVICE APPLICATIONS

Load Balancing Control for Parallel Systems

Integrating Production Scheduling and Maintenance: Practical Implications

A particle swarm optimization to vehicle routing problem with fuzzy demands

Using Phase Swapping to Solve Load Phase Balancing by ADSCHNN in LV Distribution Network

6.7 Network analysis Introduction. References - Network analysis. Topological analysis

Simple Linear Regression

Study on prediction of network security situation based on fuzzy neutral network

Optimal Packetization Interval for VoIP Applications Over IEEE Networks

Maintenance Scheduling of Distribution System with Optimal Economy and Reliability

TESTING AND SECURITY IN DISTRIBUTED ECONOMETRIC APPLICATIONS REENGINEERING VIA SOFTWARE EVOLUTION

Low-Cost Side Channel Remote Traffic Analysis Attack in Packet Networks

A New Bayesian Network Method for Computing Bottom Event's Structural Importance Degree using Jointree

Proceedings of the 2010 Winter Simulation Conference B. Johansson, S. Jain, J. Montoya-Torres, J. Hugan, and E. Yücesan, eds.

Projection model for Computer Network Security Evaluation with interval-valued intuitionistic fuzzy information. Qingxiang Li

Web Service Composition Optimization Based on Improved Artificial Bee Colony Algorithm

Load and Resistance Factor Design (LRFD)

A PRACTICAL SOFTWARE TOOL FOR GENERATOR MAINTENANCE SCHEDULING AND DISPATCHING

Capacitated Production Planning and Inventory Control when Demand is Unpredictable for Most Items: The No B/C Strategy

Compressive Sensing over Strongly Connected Digraph and Its Application in Traffic Monitoring

Forecasting Trend and Stock Price with Adaptive Extended Kalman Filter Data Fusion

An Approach to Evaluating the Computer Network Security with Hesitant Fuzzy Information

Dynamic Two-phase Truncated Rayleigh Model for Release Date Prediction of Software

The Digital Signature Scheme MQQ-SIG

Proactive Detection of DDoS Attacks Utilizing k-nn Classifier in an Anti-DDos Framework

Security Analysis of RAPP: An RFID Authentication Protocol based on Permutation

ANOVA Notes Page 1. Analysis of Variance for a One-Way Classification of Data

VIDEO REPLICA PLACEMENT STRATEGY FOR STORAGE CLOUD-BASED CDN

Preprocess a planar map S. Given a query point p, report the face of S containing p. Goal: O(n)-size data structure that enables O(log n) query time.

A Hierarchical Fuzzy Linear Regression Model for Forecasting Agriculture Energy Demand: A Case Study of Iran

The Popularity Parameter in Unstructured P2P File Sharing Networks

Secure PHR Access Control Scheme in Cloud Computing

Software Aging Prediction based on Extreme Learning Machine

STOCHASTIC approximation algorithms have several

Credibility Premium Calculation in Motor Third-Party Liability Insurance

DHA: Distributed Decisions on the Switch Migration Toward a Scalable SDN Control Plane

How To Value An Annuity

Application of Grey Relational Analysis in Computer Communication

Bayesian Network Representation

Software Reliability Index Reasonable Allocation Based on UML

On the impact of heterogeneity and back-end scheduling in load balancing designs

Efficient, Safe and Sustainable Traffic at Sea

IP Network Topology Link Prediction Based on Improved Local Information Similarity Algorithm

The simple linear Regression Model

AP Statistics 2006 Free-Response Questions Form B

n. We know that the sum of squares of p independent standard normal variables has a chi square distribution with p degrees of freedom.

Fractal-Structured Karatsuba`s Algorithm for Binary Field Multiplication: FK

1. The Time Value of Money

Mobile Agents in Telecommunications Networks A Simulative Approach to Load Balancing

ADAPTATION OF SHAPIRO-WILK TEST TO THE CASE OF KNOWN MEAN

Construction of a system scanning the movement of human upper limbs

Report 52 Fixed Maturity EUR Industrial Bond Funds

STATISTICAL PROPERTIES OF LEAST SQUARES ESTIMATORS. x, where. = y - ˆ " 1

Applications of Support Vector Machine Based on Boolean Kernel to Spam Filtering

Banking (Early Repayment of Housing Loans) Order,

CIS603 - Artificial Intelligence. Logistic regression. (some material adopted from notes by M. Hauskrecht) CIS603 - AI. Supervised learning

Transcription:

To help protect your prvacy, owerot preveted ths exteral pcture from beg automatcally dowloaded. To dowload ad dsplay ths pcture, clck Optos the Message Bar, ad the clck Eable exteral cotet. I-Network Maagemet Rolf Stadler KTH Royal Isttute of Techology Stockholm, Swede 19th Iteratoal Coferece o Computer Commucato ad Networks ICCCN 2010 August 2 5, 2010, Zurch, Swtzerlad

Outle Nt Network kmaagemet I-Network Maagemet Case Study: Real-tme Motorg Wll t happe?

Maagemet Systems aalyze aalyze Maagemet System act observe act observe Network Maaged System Server Cluster

What s Network Maagemet? Network Maagemet refers to the actvtes, methods, procedures, ad tools that perta to the operato, admstrato, mateace ad provsog of etworked systems A. Clemm, 2006. Maagemet of Networks ad Networked Systems volves the followg fve tasks CAS. ault Maagemet Cofgurato Maagemet Accoutg Maagemet & User Admstrato erformace Maagemet Securty Maagemet defto from the telecom commuty, late 1980s. 4

Network Maagemet aradgms 2010 2000 1990 1980 2 Maagemet Autoomc Maagemet OO Maagemet Maagemet olces MIB Maagemet Iformato Base TMN Telecommucato Maagemet Network 5

Network Maagemet Cofereces Yearly coferece sprg: IEEE/II IM Iteratoal Symposum o Itegrated Network Maagemet IEEE/II NOMS Network Operatos ad Maagemet Symposum Sgle-track evet fall: IEEE DSOM Dstrbuted Systems Operato ad Maagemet IEEE CNSM Coferece o Network ad Servce Maagemet

Network Maagemet Jourals IEEE Trasactos o Network ad Servce Maagemet TNSM sce 2007 Joural of Network ad Servce Maagemet JNSM sce 1993, publshed by Sprger IEEE Commucatos Magaze Seres o Network ad Servce Maagemet twce a year

Nt Network kmaagemet I-Network Maagemet Case Study: Real-tme Motorg Wll t happe?

Today s Maagemet Systems for Tradtoal Network Techologes aalyze Maagemet System act observe Maagemet tellgece outsde maaged system. Clear separato betwee maagemet system ad maaged system, by desg. Maaged System

Today s Maagemet Systems for Tradtoal Network Techologes 2 aalyze Maagemet System Maaged System Motorg ad cofgurato, geerally CAS fuctos, performed o a per-devce bass. Successful for - small umber of odes <1000 - low rate of chage - log reacto cycles <1 sec

I-Network Maagemet: Key Idea polces drectos exceptos otfcatos aradgm Shft Reduce teractos betwee maagemet ad maaged systems lace maagemet fuctos sde the maaged systems Delegate tasks to a self-orgazg maagemet plae Eablg cocepts: embeddg, decetralzato, t self-orgazato

I-Network Maagemet: Egeerg Aspects polces exceptos otfcatos self-orgazg maagemet plae maagemet ode Maagemet odes wth processg capabltes sde devce, blade, applace eer teracto through eghborhood cocept overlayoverlay Maagemet fuctos execute as dstrbuted algorthms o overlay ygraph; ca be voked o each ode; are part of a self-orgazg maagemet plae 12

The Drvers for I-Network Maagemet Lack of maagemet frastructure eergy-costrat evromet ---sesor etworks, MANETs, vehcular etworks Avodg bottleecks large-scale systems ---access etworks, data ceters, maaged ed-devces Shorte reacto tme -dyamc evromets -msso-crtcal etworks State ca be estmated ad acted upo sde the etwork - ault maagemet - Routg, resource allocato 13

ault Resoluto Tmes Excessve OS messages force US Telco to brg dow parts of ATM etwork: 26 hrs Outage several Mllo US$ Impact Re esoluto Tm me Bad redudacy mplemetato forces traffc through a 64kbt udersea cable: 4 hrs Outage several Mllo Impact Source: Csco LS black hole ssue forces Arle to groud all plaes: 20 mutes Outage Lack of memory a swtch several Mllo US$ Impact causes Itermtted outages o tradg floor Impact: 1 Mllo per 1 mute Iadequate QoS o GgE lk bookstore mpacts 10 000 trasactos per secod: Mllos of US$ secods 1995 2000 2005 2010 2015

Sde Thought: A Revval of Network rogrammg? Itatves 1995-2005: Actve Networkg: actve packets wth state ad code, customzed packet processg o routers; pursued by Iteret commuty rogrammable Networks: focus o terfaces, e.g., for coecto maagemet, QoS; pursued by broadbad commuty, stadardzato IEEE 1520 Impact: specalzed techologes programmable layer 4/7 swtches, tellget frewalls, lmted dustral mpact o adopto by major maufacturers; operators ad provders valued operatoal safety over flexblty 15

Nt Network kmaagemets I-Network Maagemet Case Study: Real-tme Motorg Wll t happe?

Motorg Aggregates Aggregate tw 1 t,..., w t Aggregato fuctos,w,., w j,...,w j,., w,... w l t w k t w t Local varables w j t Sum w 1 1,,...,, w, Average, Max, Quatle Dstctve Elemets {w 1,..., w } Heavy htters { } Hstogram {w 1,..., w }

Decetralzed Motorg Aggregate tw 1 t,..., w t t Aggregato rotocol w l t w t w k t w j t

Challeges Estmato of etwork states, stuato awareess, threshold detecto. Uderstadg ad cotrollg trade-offs betwee accuracy, overhead, robustess, depedecy o the system sze, dyamcty, to buld tuable ad self-tug systems Uderstadg the sematcs of mgt operatos o a large system uder chage Uderstadg the mpact of estmato errors o the effect of maagemet decsos

A-GA: rotocol desg goals rovde a maagemet applcato wth a cotuous estmate of a aggregate sum of local values for a gve accuracy. Tuable trade-off: accuracy vs. overhead -lowest overhead for a gve accuracy objectve Dyamc y adaptato to chages -chages to local values, topology, falures Scalablty -overhead crease wth system sze s sublear A. Gozalez reto, R. Stadler: A-GA: A Adaptve rotocol for Cotuous Network Motorg wth Accuracy Objectves, IEEE Trasactos o Network ad Servce Maagemet TNSM, Vol. 4, No. 1, Jue 2007 D. Jurca, R. Stadler, H-GA: Estmatg Hstograms of Local Varables wth Accuracy Objectves for Dstrbuted Real-Tme Motorg, IEEE Trasactos o Network ad Servce Maagemet TNSM, Vol. 7, No. 2, Jue 2010. 20

I-Network Aggregato usg Spag Trees Maagemet Stato artal Aggregate Local varable Global Aggregate 25 Root 12 10 3 SumtSumw 1 t,..., w t hyscal Node Aggregatg Node Leaf Node 4 1 3 5 2 7 3 5 2 7 w 7 t 21

A-GA: rotocol desg prcples Creatg ad matag spag tree -Spag tree o maagemet overlay -BS tree based o self-stablzg stablzg protocol by Dolev, Israel, Mora 90 Icremetal -etwork aggregato g o spag tree - Aggregate computed bottom-up o odes of tree -Result avalable at root ode lterg updates -Reduce protocol overhead by flterg updates whle observg error objectve -Compute flters usg a dstrbuted heurstc S. Dolev, A. Israel, ad S. Mora, Self-stablzato of dyamc systems assumg oly read/wrte atomcty. ACM Symposum o rcples of Dstrbuted Computg ODC '90, Quebec Cty, Quebec, Caada, August,1990. 22

Local Adaptve lters Local varable or partal aggregate g Last update value lter wdth lter Exceeded: 1 Trggers a update to paret 2 lter s shfted tme Local flter o a ode Cotrols the maagemet overhead by flterg updates Drops updates wth small chage to partal aggregate erodcally adapts to the dyamcs of etwork evromet 23

roblem ormalzato d flter wdths to motor aggregate for a gve accuracy objectve, wth mmal overhead Overhead: max processg load ω over all maagemet processes Accuracy objectve: average error Mmze { } Max ω s.t. E[ E root ] ε percetle error Mmze Max { ω } Max ω st s.t. p E root >γ θ maxmum error Mmze { } Max ω s.t. E root κ 24

A Dstrbuted Heurstc The global problem s mapped oto a local problem for each ode { } Mmze ω π Max s.t. π E E out ε Attempts to mmze the maxmum processg load over all odes by mmzg the load wth each ode s eghborhood lter computato: decetralzed ad asychroous Each ode depedetly rus a cotrol cycle: every τ secods { request model varables from chldre compute ew flters ad accuracy objectves for chldre } compute model varables for local ode 25

A Stochastc Model for the Motorg rocess Model based o dscrete-tme Markov chas It relates for each ode - the error of ts partal aggregate - evoluto of the partal aggregate - the rate of updates seds - the wdth of the local l flter It permts to compute for each ode - the dstrbuto of estmato error - the protocol overhead Updates to paret G Node state λ S out E out Update rate Step szes Estmato Error lter wdth ω Update rate processg load Updates from chldre S E Step szes Estmato Error 26

Stochastc Model: leaf ode + + 0 otherwse X X j Estmatg step sze MLE Evoluto of local varable X Trasto Matrx. 0 otherwse < < + 0 0, j j j j X X j X t > + + s z s G z X d s s z Step Sze + 0 0 otherwse s z d G z X s S d d d z out Estmato Error Maagemet O erhead 0 otherwse. 0 1 S λ out G E 27 Maagemet Overhead 0 1 out S λ

Stochastc Model: aggregatg ode Step Sze: > + s k s G k S s Iput Output Δ Δ c c c c out s S s S γ > + s k d G k S s k s G k S s S d d d k s k out 0 Estmato Error: c out E E out G E E + c γ otherwse 0 c λ ω Maagemet Overhead: c γ 0 1 Δ S λ. c γ λ ω 0 1 Δ out S λ Trasto Matrx: 0, j j j j S t 28 Trasto Matrx: < < + 0 j j S S t

Model-based Motorg Error Objectve Estmato Error Estmato model varables Optmzato roblem Stochastc Model of Motorg rocess 0,035 0,03 0,025 0,02 0,015 0,01 0,005 0-100 -75-50 -25 0 25 50 75 100 Overhead 3 2,5 Updates/sec 2 1,5 1 0,5 0 ode 1 ode 2 ode 3 ode 4 ode 5 ode 6 ode 7 Step Szes lter Wdths Measured Estmated Aggregate Estmato x t Measuremets local varables Tree-based aggregato

Tradeoff: Accuracy vs Overhead 600 500 ε 0 ARC Upda ates/sec 400 300 200 100 0 ε 2 ε 5 T m 0.03 0.04 0.05 0.10 0.20 A-GA ε 10 ε 15 ε 20 0 5 10 15 20 25 30 Avg Error Overhead decreases mootocally Overhead depeds o the chages of the aggregate, ot o ts value. A-GA outperforms a rate-cotrol scheme ARC 30

Robustess mato Error Est 3500 2500 1500 500-500 140 145 150 155 160 165 170 Tme 175 Max xmum Load U Updates/sec 80 60 40 20 Node A fals Ed of Traset 0 140 145 150 155 160 165 170 Tme 175 Estmato error: several spkes durg sub-secod traset perod Overhead: sgle peak wth a log traset 31

A-GA rototype Lab testbed at KTH 16 motorg odes 16 Csco 2600 Seres routers Smartbts 6000 traffc geerator A-GA mplemeted Java Maagemet Stato Aggregato o Tree Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 hyscal Network 32

rototype: Maagemet Stato Iterface Select Aggregato ucto Select Accuracy Objectve Select Root Node Evoluto of the Aggregate True Value ad A-GA Estmato Overhead Dstrbuto ad Evoluto Show Aggregato Tree Real-tme Estmato of Error Dstrbuto ad Trade-off 33

rototype: Error Estmato by A-GA vs Actual Error 0,10 008 0,08 Measured Error Error Estmated by A-GA 0,06 0,04 0,02 Absolute Avg Error 0,00-30 -20-10 0 10 20 Error 30 Accurate estmato of the error dstrbuto Maxmum error >> average error oe order of magtude 34

Gossp vs. Tree-based Aggregato

Computg aggregates through gosspg ush Syopses [Kempe et al. 03] The protocol computes AVERAGE of the local varables x. After each roud a ew estmate of the aggregate s computed as s /w. Expoetal covergece o coected graphs rotocol Ivarats: s x,, r, r, w r r Roud 0 { 1. s x ; 2. w 1; 3. sed s, to self } w Roud r + 1 { * * 1. Let { sw, } be all pars set to durg roud r * 2. s s w ; l l l l l * w l 3. choose shares α 0 for all odes j, j such that α j, j 1 4. for all j sed α * s, α * w to each j }, j, j D. Kempe, A. Dobra, ad J. Gehrke, Gossp-based computato of aggregate formato, roc. 44th Aual IEEE Symposum oudatos Computer Scece OCS, Oct. 2003.

The G-GA protocol 5. for all j Neghbors { a. rs, j, rw, j rs, j, rw, j + rs m, rw m acks m, ackw m Roud 0 { morg : m j 1. s b. acks, j, ackw, j srs, j, srw, j + x ; 2. w 1 ; s m, w m morg : m j 3. L {} ; c. f detected_falurej { 4. for each ode j rs, 0,0. s, w s, w + rs,, rw,, j rw, j ; j j 5. for each ode j srs. rs, j, rw, j srs, j, srw, j 0,0, j, srw, j 0,0 ; 6. sed s,,0,0,0,0. L L \ j w to self; 7. for all j sed 0,0,0,0,0,0 to j } } Roud r+1 { } 1. Let M be all messages receved 6. for all j L { by durg roud r a. choose α, j 0 such that j α, j 1 2. s s m + xr, x 1, m M r ; w w m m M b. choose β, j 0 such that 3. for all j acks, j, ackw, j 0,0 j β, j 1 ad β, 0 4. L L org M c. srs, j, srw, j β, j α, s x, β, j α, w 1 d. sed α, js, α, jw, srs, j, srw, j, acks, j, ackw, j to j e. rs, rw rs + +α s, rw + α w } }, j, j, j, j, j, j

Accuracy vs. Overhead gossp- ad tree-based aggregato protocol GA ad G-GA 654 ode etwork GoCast overlay, coectvty 10 aggregato: AVERAGE UT trace 4 rouds/sec o falures. Wuhb, M. Dam, R. Stadler, A. Clemm Robust Motorg of Network-wde Aggregates through Gosspg, IEEE Trasactos o Network ad Servce Maagemet TNSM, Vol. 6, No. 2, Jue 2009.

Accuracy vs. alure Rate gossp- ad tree-based aggregato protocol GA ad G-GA 654 ode etwork GoCast overlay, coectvty 10 aggregato: AVERAGE UT trace 4 rouds/sec odes fal radomly, recover after 10 sec Tree-based aggregato outperforms gossp-based aggregato!

Nt Network kmaagemets I-Network Maagemet Case Study: Real-tme Motorg Wll t happe?

I-Network Maagemet Why t wll happe Compared to 5-10 years ago: New actors Google, Amazo, Mcrosoft, Apple New drvers data ceter etworkg, cloud computg, Advaces dstrbuted computg gossp protocols, algorthms for vrtual topologes, uderstadg protocols o dyamc topologes Eablers of etwork programmablty maufacturers Juper, Csco provde ope terfaces Opelow allows for programmable cotrol ad maagemet plaes 41