Performance Tuning of Failure Detectors in Wireless Ad-Hoc Networks: Modelling and Experiments

Transcription

1 Performance Tuning of Failure Detectors in Wireless Ad-Hoc Networks: Modelling and Experiments Laboratoire ID-IMAG (UMR 5132), Projet Apache. MIRRA Project: Industrial collaboration INRIA - France Télécom R&D September 2, 2005 orine Marchand & Jean-Marc Vincent {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 1 / 24Pr

2 Outline 1 Wireless environment Wireless environment Distributed systems on a wireless network The Consensus Problem 2 Quality of Service of Failure Detectors Theoretical Concept Failure Detectors Implementation Quality of Service 3 Stochastic Models Statistical Description Independent assumption Model Contention on receiver 4 Experimentation Experimental environment Saturation Ideal Setting Experimentation Experimentation in perturbed environment 5 Conclusion and future works {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 2 / 24Pr

3 Outline 1 Wireless environment Wireless environment Distributed systems on a wireless network The Consensus Problem 2 Quality of Service of Failure Detectors Theoretical Concept Failure Detectors Implementation Quality of Service 3 Stochastic Models Statistical Description Independent assumption Model Contention on receiver 4 Experimentation Experimental environment Saturation Ideal Setting Experimentation Experimentation in perturbed environment 5 Conclusion and future works {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 3 / 24Pr

4 Wireless environment Environment Wireless Devices Laptop computers, Personal digital assistants (PDAs), Mobile phones,... Wireless Ad-Hoc Network WIFI, Bluetooth,... Principle Devices share services and collaborate to maintain the community. General distributed system problem Dynamical control of the architecture. State of the system observation and distributed decision process. orine Marchand & Jean-Marc Vincent EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 4 / 24Pr

5 Wireless environment Environment Wireless Devices Laptop computers, Personal digital assistants (PDAs), Mobile phones,... Wireless Ad-Hoc Network WIFI, Bluetooth,... Principle Devices share services and collaborate to maintain the community. General distributed system problem Dynamical control of the architecture. State of the system observation and distributed decision process. orine Marchand & Jean-Marc Vincent EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 4 / 24Pr

6 Operating systems on a wireless network Dynamic Architecture Heterogeneity of devices Behavior of wireless devices connections / disconnections Behavior of wireless network Unreliability of communications Variability of latencies Solution to maintain the consistency of the community Design and adapt distributed algorithms; to make some distributed decisions. (consensus, election, atomic broadcast, group membership,...) {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 5 / 24Pr

7 The Consensus Problem The impossibility of Fisher, Lynch & Paterson [Fischer-Lynch-Paterson 85] Some approaches to circumvent this impossibility result: Probabilistic algorithms [Canetti-Rabin 93] Self-Stabilizing algorithms [Tixeuil 00] An approach with partial synchrony [Dwork-Lynch-Stockmeyer 88]... The selected approach : Consensus + unreliable failure detectors [Chandra-Toueg 96] Interest: Dynamicity analysis of the environment is concentrated only inside failure detectors. Objective: Guarantee a quality of service for failure detectors. {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 6 / 24Pr

8 The Consensus Problem The impossibility of Fisher, Lynch & Paterson [Fischer-Lynch-Paterson 85] Some approaches to circumvent this impossibility result: Probabilistic algorithms [Canetti-Rabin 93] Self-Stabilizing algorithms [Tixeuil 00] An approach with partial synchrony [Dwork-Lynch-Stockmeyer 88]... The selected approach : Consensus + unreliable failure detectors [Chandra-Toueg 96] Interest: Dynamicity analysis of the environment is concentrated only inside failure detectors. Objective: Guarantee a quality of service for failure detectors. {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 6 / 24Pr

9 Outline 1 Wireless environment Wireless environment Distributed systems on a wireless network The Consensus Problem 2 Quality of Service of Failure Detectors Theoretical Concept Failure Detectors Implementation Quality of Service 3 Stochastic Models Statistical Description Independent assumption Model Contention on receiver 4 Experimentation Experimental environment Saturation Ideal Setting Experimentation Experimentation in perturbed environment 5 Conclusion and future works {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 7 / 24Pr

10 Unreliable Failure Detectors Principle: For each remote device, build an estimation of the global state. Local view of the global system. List of suspected devices Properties Accuracy: a correct process should not be suspected Completeness: an incorrect process should be suspected Quality of service Quality of information and reactivity false suspicion rate = function(reactivity) Implementation - false suspicion type 1: correct process suspected - false suspicion type 2: crashed process not suspected Risk analysis {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 8 / 24Pr

11 Unreliable Failure Detectors Principle: For each remote device, build an estimation of the global state. Local view of the global system. List of suspected devices Properties Accuracy: a correct process should not be suspected Completeness: an incorrect process should be suspected Quality of service Quality of information and reactivity Implementation false suspicion rate = function(reactivity) Timeout - false suspicion type 1: correct process suspected - false suspicion type 2: crashed process not suspected Risk analysis No Suspicion Heartbeat Receipt Suspicion {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 8 / 24Pr

12 Failure Detectors Implementation Need of information on remote devices Export local data Collect and analyse data coming from remote devices Info publication Consultation Informations Export Broadcast of information by anticipation (Heartbeat mecanism) Network Interface Export module Middleware Interface MW control ( policy & parameters ) Suspicion request Information List to query List of suspects Consensus interface MW control ( policy & parameters ) Network interface Import module Middleware Interface Suspicion request Informations Import Collect information Estimate the state of remote devices Information {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), 9 / 24Pr

13 Failure Detectors Parameters The running principle: Implemented mecanism: heartbeat. FD Export Process FD sending delay Sender Device FD Import Process FD receipt delay θ θ θ θ θ Receiver Device Parameters Heartbeat sending period. Estimate function of suspicions (timeout). [Bertier-Marin-Sens 03] Goal Estimate the timeout values according to the expected quality of service. EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

14 Failure Detectors Parameters The running principle: Implemented mecanism: heartbeat. FD Export Process FD sending delay Sender Device FD Import Process FD receipt delay θ θ θ θ θ Receiver Device Parameters Heartbeat sending period. Estimate function of suspicions (timeout). [Bertier-Marin-Sens 03] Goal Estimate the timeout values according to the expected quality of service. EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

15 Outline 1 Wireless environment Wireless environment Distributed systems on a wireless network The Consensus Problem 2 Quality of Service of Failure Detectors Theoretical Concept Failure Detectors Implementation Quality of Service 3 Stochastic Models Statistical Description Independent assumption Model Contention on receiver 4 Experimentation Experimental environment Saturation Ideal Setting Experimentation Experimentation in perturbed environment 5 Conclusion and future works {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

16 Statistical Description FD Export Process Network FD sending delay HB Interval (given) Network sending delay Sender Device Network FD Import Process Network receipt delay HB interval (measured) FD receipt delay Receiver Device Variability of HB arrivals λ 0 = emission beat rate X i = Heartbeat inter-arrivals. λ = 1 HB period (assumption : few losses λ = λ 0.(1 loss probability) orine Marchand & Jean-Marc Vincent {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

17 Statistical Description False Detection Probability θ = suspicion threshold (timeout) φ I (θ) = asymptotic false suspicion rate 1 φ I (θ) = λ lim n n nx (X i θ) + If the inter-arrivals {X i } of beats are independent and identically distributed, then : i=1 φ I (θ) = λe π [X θ] + where π is the distribution of X i. (renewal process) {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR 5132), / 24Pr

18 Independent assumption Model Variable Sending Delay Hypothesis : {X i } : renewal process (iid) Model suspicion rate no information on variance Exponential φ I (θ) = e λθ low variation coefficient Erlang(k,kλ) φ I (θ) = e kλθ P k (λθ) high variation coefficient Pareto(α) φ I (θ) = 1 (1+ α 2 θ )α 2 Suspicion probability related to reactivity Erlang model: Pareto model: {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

19 Independent assumption Model Variable Sending Delay Hypothesis : {X i } : renewal process (iid) Model suspicion rate no information on variance Exponential φ I (θ) = e λθ low variation coefficient Erlang(k,kλ) φ I (θ) = e kλθ P k (λθ) high variation coefficient Pareto(α) φ I (θ) = 1 (1+ α 2 θ )α 2 Suspicion probability related to reactivity Erlang model: False suspicion rate k=1 (exponential model) Pareto model: k= σ= 3 1e 04 k=3 k=5 k=4 1e θ σ=1.1 σ= θ {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

20 Contention on receiver (1) Variability of Heartbeat Arrivals Depends on the type of receiver (Laptop or PDA) + Correlation between inter-beats arrival periods Network FD Import Process variability of delivery Receiver Device HB contention on the receiver Heart beats Network Buffer Service Failure detector input process : {A n} n N service model : {S n} n N Delivery to upper layer GI/M/1 queue hypothesis : deterministic arrivals {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

21 Contention on receiver (2) Output process of a D/M/1 queue Suspicion rate computation Inter-output period : A = 1 λ ; β unique solution of β = L A(µ(1 β)) = e Aµ(1 β) f Z (x) = ( µ Rate of false suspicion : θ > A φ I (θ) = 2 β e µ(1 β)a ((1 β)e µ(1 β)x + e µx ) if x < A; µ 2 β e µx (e µ(1 β)a + (1 β)e µa ) if x A. 1 A(2 β)µ e µθ (e µ(1 β)a + (1 β)e µa ) θ A False suspicion probability, D/M/1 model {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

22 Contention on receiver (2) Output process of a D/M/1 queue Suspicion rate computation Inter-output period : A = 1 λ ; β unique solution of β = L A(µ(1 β)) = e Aµ(1 β) f Z (x) = ( µ Rate of false suspicion : θ > A φ I (θ) = 2 β e µ(1 β)a ((1 β)e µ(1 β)x + e µx ) if x < A; µ 2 β e µx (e µ(1 β)a + (1 β)e µa ) if x A. 1 A(2 β)µ e µθ (e µ(1 β)a + (1 β)e µa ) θ A False suspicion probability, D/M/1 model 10 False probability suspicion False probability suspicion µ=2 1 1e 05 µ= e e 15 1e 04 1e 20 1e 05 1e 25 1e 06 1e 30 A=1 A=1 1e 07 1e 35 θ θ {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

23 Outline 1 Wireless environment Wireless environment Distributed systems on a wireless network The Consensus Problem 2 Quality of Service of Failure Detectors Theoretical Concept Failure Detectors Implementation Quality of Service 3 Stochastic Models Statistical Description Independent assumption Model Contention on receiver 4 Experimentation Experimental environment Saturation Ideal Setting Experimentation Experimentation in perturbed environment 5 Conclusion and future works {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

24 Experimental environment Experimental Design: Devices (same OS, Java): Architecture 1 : 4 devices (2 Laptops + 2 PDAs) Architecture 2 : 6 devices (2 Laptops + 3 PDAs + 1 sensor Interconnection : b ad-hoc network Experimental duration: 15 min ( about 10,000 measurements) HB parameter Settings: Architecture 1 Architecture 2 Architecture 2 Highly loaded Ideal Setting Perturbed Environment HB emission period 100 ms 500 ms 500 ms Timeout none none none Reception process analysis Density of the delivery process timeout tuning {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

25 Experimental environment Experimental Design: Devices (same OS, Java): Architecture 1 : 4 devices (2 Laptops + 2 PDAs) Architecture 2 : 6 devices (2 Laptops + 3 PDAs + 1 sensor Interconnection : b ad-hoc network Experimental duration: 15 min ( about 10,000 measurements) HB parameter Settings: Architecture 1 Architecture 2 Architecture 2 Highly loaded Ideal Setting Perturbed Environment HB emission period 100 ms 500 ms 500 ms Timeout none none none Reception process analysis Density of the delivery process timeout tuning {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

26 Highly loaded system Distribution of the update times: 0.01 sender: PDA1 - receiver: PDA sender: PDA1 - receiver: laptop Probability Probability Times between 2 heartbeat receipts sender: laptop1 - receiver: PDA Times between 2 heartbeat receipts Probability Probability Times between 2 heartbeat receipts 0.08 sender: laptop1 - receiver: laptop Times between 2 heartbeat receipts If timeout value = 200 ms Quality of service highly depends on the type of the receiver orine Marchand & Jean-Marc Vincent {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

27 Highly loaded system Distribution of the update times: 0.01 sender: PDA1 - receiver: PDA sender: PDA1 - receiver: laptop Probability Probability Times between 2 heartbeat receipts sender: laptop1 - receiver: PDA Times between 2 heartbeat receipts Probability Probability Times between 2 heartbeat receipts 0.08 sender: laptop1 - receiver: laptop Times between 2 heartbeat receipts If timeout value = 200 ms Quality of service highly depends on the type of the receiver orine Marchand & Jean-Marc Vincent {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

28 Ideal Setting Experimentation Heartbeat Reception Analysis: 0.14 sender: laptop1 receiver: laptop2 sender: laptop1 receiver: pda2 sender: pda1 receiver: laptop sender: pda1 receiver: pda2 Probability Elapsed times between 2 receipts Timeout value: If the timeout value = 2 (HB period time) Then, the suspicion rate is around 10 3 if the receiver is a laptop 10 2 if the receiver is a PDA {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

29 Ideal Setting Experimentation Heartbeat Reception Analysis: 0.14 sender: laptop1 receiver: laptop2 sender: laptop1 receiver: pda2 sender: pda1 receiver: laptop sender: pda1 receiver: pda2 Probability Elapsed times between 2 receipts Timeout value: If the timeout value = 2 (HB period time) Then, the suspicion rate is around 10 3 if the receiver is a laptop 10 2 if the receiver is a PDA {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

30 Experimentation in perturbed environment Perturbation control : An external device is used to generate an external load (ping with 200kbytes/s) Heartbeat Reception Analysis: sender: Laptop1 receiver: Laptop2 sender: Laptop1 receiver: PDA2 sender: PDA1 receiver: Laptop2 sender: PDA1 receiver: PDA2 Probability Elapsed Times between 2 heartbeat receipts Results: Long non receiving period for PDA Some very small delays between HB receipts (after a long waiting time) Correlation between succesive waiting times of two HB (bursty receptions) {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

31 Experimentation in perturbed environment Perturbation control : An external device is used to generate an external load (ping with 200kbytes/s) Heartbeat Reception Analysis: sender: Laptop1 receiver: Laptop2 sender: Laptop1 receiver: PDA2 sender: PDA1 receiver: Laptop2 sender: PDA1 receiver: PDA2 Probability Elapsed Times between 2 heartbeat receipts Results: Long non receiving period for PDA Some very small delays between HB receipts (after a long waiting time) Correlation between succesive waiting times of two HB (bursty receptions) {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

32 Outline 1 Wireless environment Wireless environment Distributed systems on a wireless network The Consensus Problem 2 Quality of Service of Failure Detectors Theoretical Concept Failure Detectors Implementation Quality of Service 3 Stochastic Models Statistical Description Independent assumption Model Contention on receiver 4 Experimentation Experimental environment Saturation Ideal Setting Experimentation Experimentation in perturbed environment 5 Conclusion and future works {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

33 Conclusion Modelling Models for systems in stationary regime Contention on receivers Analytical formula Architecture Identification of FD QoS Experimental protocol Implementation in a distributed Middleware (Sidrah) {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr

34 Future works Modelling Non-stationary analysis : auto-regressive models Environment evolution model Application model : network sniffing Architecture Scaling the system : routing (dynamic graph model) Multilevel observation environment Dynamic control : contention prevention {Corine.Marchand,Jean-Marc.Vincent}@imag.fr EPEW - September 2005 ( Laboratoire September ID-IMAG 2, 2005 (UMR ), / 24Pr