Report

A Forward end-to-end delays Analysis for packet switched networks Georges Kemayo, Frédéric Ridouard, Henri Bauer, Pascal Richard LIAS, Université de Poitiers, ISAE/ENSMA, France RTNS’2014 October 08-10, 2014, Versailles, France LIAS - ISAE/ENSMA - Université de Poitiers Outline 1 Context The AFDX network 2 State of the art End-to-End delay variability Network Calculus and Trajectory Approach 3 Contribution Forward end-to-end delay Analysis (FA) 4 Conclusion and future work 2 The AFDX network: Generalities (1/2) Context State of the art Contribution Conclusion Analysis of AFDX networks used in avionics systems AFDX = Avionics Full Duplex Switched Ethernet Fully static Switched Ethernet network « End/Systems » interconnected by « switches » and « physical links » A physical link between 2 components is full duplex no loss of frames due to collisions ES1 ES2 ES3 ES4 v1 v2 v3,v4 v5,v6 S1 S2 S4 v1,v2 v3 v3,v4 v5,v6 S3 v1,v3 v2 v3,v4,v5 S5 v6 ES5 ES6 ES7 ES8 8 End/Systems ESi 5 Switches Si 6 Virtual links vi 3 The AFDX network: Generalities (2/2) Context State of the art Contribution Conclusion AFDX = Avionics Full Duplex Switched Ethernet End/System: Message 1 Message n Switch: … … Control and Routing 4 The AFDX network: Virtual link Context State of the art Contribution Conclusion Virtual link = static, unidirectionnal and monotransmitter logical channel generated by only one source End/System towards one or many End/Systems (multicast) avionics data flows correspond to virtual links Any virtual link respects a traffic contract: Fmin ≤ data frame lenght ≤ Fmax BAG = minimum inter-generation time of frames on its source End/System: ≥ BAG < BAG ES1 (v1) guaranteed bandwidth for any data flow: Fmax / BAG ES1 ES2 ES3 ES4 v1 v2 v3,v4 v5,v6 S1 S2 S4 v1,v2 v3 v3,v4 v5,v6 S3 v1,v3 v2 v3,v4,v5 S5 v6 ES5 ES6 ES7 ES8 8 End/Systems ESi 5 Switches Si 6 Virtual links vi 5 The AFDX network: Notion of ETE delay Context State of the art Contribution Conclusion The ETE delay of a data frame in the AFDX: S1 ES1 S2 ES2 ETE delay Variable waiting durations in buffers (difficult to evaluate) Objective: Guarantee the worst ETE delay of any frame of any flow vi crossing the AFDX (Mandatory for certification) Necessity to use a method to compute the worst ETE delay 6 LIAS - ISAE/ENSMA - Université de Poitiers Outline 1 Context The AFDX network 2 State of the art End-to-End delay variability Network Calculus and Trajectory Approach 3 Contribution Forward end-to-end delay Analysis (FA) 4 Conclusion and future work 7 End-to-End delay Variability Context State of the art Contribution Conclusion The variability of the waiting duration in each crossed buffer implies: The ETE delay is between a lower bound and an exact worst case Can miss some rare scenarios leading to the exact worst ETE delay ETE delay distribution obtained by simulation Combinatorial explosion when computing the exact worst ETE delay Simulation Model Checking Network Calculus, Trajectory Approach miss of rare Scenarios Lower bound of the ETE delay Worst ETE delay Exact worst case Upper bound of (observed) ETE delay the ETE delay time 8 Network Calculus and Trajectory Approach Context State of the art Contribution Conclusion What is the necessity to design a new method? FA (Forward end-to-end delay Analysis) correct the disadvantages of the existing methods Network Calculus Trajectory Approach FA Theory (min,+) algebra Real-time scheduling Real-time scheduling Policy FIFO, … FIFO, … FIFO Serialization (coming soon Pessimism (Bauer et al. IEEE TII’10) (Li et al. RTNS’11) ) (to be studied) Global charge > 1 Absence of optimism (Kemayo et al. ETFA’13) Global charge: sum of charges of all the flows encountered on any crossed node Serialization: 1 2 S Frames 1 and 2 not are serialized they can delay each other 2 1 S Frames 1 and 2 are serialized, frame 1 cannot delay frame 2 9 LIAS - ISAE/ENSMA - Université de Poitiers Outline 1 Context The AFDX network 2 State of the art End-to-End delay variability Network Calculus and Trajectory Approach 3 Contribution Forward end-to-end delay Analysis (FA) 4 Conclusion and future work 10 Analysis of the worst ETE delay of flows with FA (1/10) Context State of the art Contribution Conclusion Modelization of AFDX by the FA method (1/2) 1 output port (Switch or End/System) 1 virtual link vi a network node flow, characterized by: ≥ Ti vi i i Ci Maximum transmission time: Ci = Fmax / R (R = rate of the physical link) Minimum inter-generation time between two consecutive frames: Ti = BAG 11 Analysis of the worst ETE delay of flows with FA (2/10) Context State of the art Contribution Conclusion Modelization of AFDX by the FA method (2/2) v1 ES1 AFDX : S1 v2 ES2 v3,v4 ES3 S3 v3 S2 v5,v6 ES4 v1,v2 1 2 v3,v4 S5 v5,v6 S4 ES1 ES2 Modelization v ,v 3 4 by FA : ES3 v5,v6 ES4 1 v3,v4,v5 v6 2 ES5 ES6 ES7 ES8 S32 v2 S1 v1 v2 v2 2 v2 v1 v1 v1,v3 v2 1 v3 v3,v4 v5,v6 S21 S22 S4 v3 v3,v4 v5 v6 S31 S51 S52 v1,v3 v3,v4,v5 v6 12 Analysis of the worst ETE delay of flows with FA (3/10) Context State of the art Contribution Conclusion The FA principle Worst ETE delay Ri of a flow vi: i - Maximum delay incurred by fi to arrive on lasti: S maxlast i - Maximum backlog encountered by fi on lasti ( FIFO policy) … firsti Generation time of fi i S maxlast i i Bklglast i Ci Worst arrived time of fi fi lasti Ri i i Ri S max last Bklg last Ci i i 13 Analysis of the worst ETE delay of flows with FA (4/10) Context State of the art Contribution Conclusion lasti Computation of the maximal delay S maxi i How to compute S maxlast ? i firsti - Iterative computation, knowing that Smaxi Generation time of fi Bklgih … firsti 0 Ci fi h S max ih L h 1 S maxih1 S max ih 1 S max ih Bklg ih Ci L L = propagation delay of a frame on the link between h and h+1 14 Analysis of the worst ETE delay of flows with FA (5/10) Context State of the art Contribution Conclusion Maximum interference of frames of a same flow on a node h (1/3) Question: how to compute Bklgih on a node h ? Theorem: considering a temporal interval [a,b] on h, the scenario leading, for any flow vj, to its greatest amount of work is obtained when: … first j Tj Tj Tj S max hj S max hj S max hj fj S min hj h a Sminhj 1 Sminhj C j L b with first j S min j 0 15 Analysis of the worst ETE delay of flows with FA (6/10) Context State of the art Contribution Conclusion Maximum interference of frames of a same flow on a node h (2/3) Example: determination of the worst case backlog of a single flow vj on h: [a, b] = [20, 110] 80 … first j h Tj 40 120 Tj 40 hh max 90 SSSmax 70 maxj jhj120 60 Sminhj 30 Tj Tj 80 0 90 120 6070 a 20 40 120 6070 90 Tj 120 90 70 60 fj 30 b 110 When Smaxhj 60, no other frame of vj can catch up the frame fj on h When Smaxhj 70, one frame of vj catches up the frame fj on h When S max hj 90, always one frame of vj still catches up the frame fj on h When Smax hj 120, two frames of vj catch up the frame fj on h 16 Analysis of the worst ETE delay of flows with FA (7/10) Context State of the art Contribution Conclusion Maximum interference of frames of a same flow on a node h (3/3) Example: determination of the worst case backlog of a single flow vj on h: [a, b] = [20, 110] 80 first j Tj 40 … 120 Tj 40 maxhjhhjj70 90 120 SSmax Sminhj 30 Tj Tj 80 0 120 70 90 120 7090 h 40 Tj 120 90 70 fj 30 a 20 General case: … first j b 110 k frames Tj Tj … S max hj S max hj Tj fj Smin hj k0 h a b 17 Analysis of the worst ETE delay of flows with FA (8/10) Context State of art Contribution Conclusion Maximum interference of frames of all the flows crossing a node h: Bklgih … first j Tj Tj Tj … S max hj S max hj fj S min hj t=b-a h a b Usage of the request bound function theory for computing the maximal transmission duration of vj frames arrived in [a,b]: RBFi h (t ) Total transmission duration of frames of all the flows crossing h and arriving in [a,b]: W h (t ) RBF v j h i h (t ) Bklgih is computed based on W h (t ) 18 Analysis of the worst ETE delay of flows with FA (9/10) Context State of the art Contribution Conclusion Comparative study of the FA method on an AFDX example (1/2) v2 v1 v1 ES1 S1 v1 v2 v2 ES2 v3,v4 ES3 v5,v6 ES4 v2 S32 v3 v3,v4 v5,v6 S21 v3 v3,v4 S22 v5 S4 Ci v6 100 v2 100 10 S51 v1,v3 v3,v4,v5 S52 v6 Ti v1 v3 S31 50 v4 30 v5 30 v6 50 L = 16 19 Analysis of the worst ETE delay of flows with FA (10/10) Context State of the art Contribution Conclusion Worst case ETE delays Comparative study of the FA method on an AFDX example (2/2) FA NC NCNS TA v1 82 85 85 82 v2 72 73 73 72 v3 (S31) 82 87 87 82 v3 (S51) 112 100.8 126.7 82 NCNS = Network Calculus Not v4 112 100.8 126.7 82 taking into account the Serialization v5 112 100.8 126.7 N/A TA = Trajectory Approach taking v6 82 77.6 92 72 FA = Our approach whitout serialization NC = Network Calculus taking into account the serialization into account the serialization Remarks: 1) TA is not optimistic in this example 2) TA cannot compute the delay of flow v5, its global charge is greater than 1 3) FA is better than NCNS, but can compete with NC only for flows v1, v2, v3 4) FA is more pessimistic than TA, but obtains same results for flows v1, v2, v3 5) No method is better than the others 20 LIAS - ISAE/ENSMA - Université de Poitiers Outline 1 Context The AFDX network 2 State of the art End-to-End delay variability Network Calculus and Trajectory Approach 3 Contribution Forward end-to-end delay Analysis (FA) 4 Conclusion and future work 21 Conclusion and Perspectives Context State of the art Contribution Conclusion Conclusion: Analysis of the ETE delay of flows on AFDX networks Comparative study of the ETE delay computation methods for the AFDX Proposition of a new method, Forward end-to-end delay Analysis, for the AFDX FIFO policy Future work on FA: Take into account the serialization Large cases comparison with others existing methods Extension to other policies: Fixed priorities, … 22 LIAS - ISAE/ENSMA - Université de Poitiers Thank you for your attention! ? 23