Controller Area Network (CAN) Schedulability Analysis for Messages with Arbitrary Deadlines in FIFO and Work-Conserving Queues

Transcription

1 Cotroller Area Networ CAN) Schedulability Aalysis for Messages with Arbitrary Deadlies i FIFO ad Wor-Coservig Queues Robert Davis 1 ad Nicolas Navet 2 1 Real-Tie Systes Research Group, Uiversity of Yor, UK 2 INRIA / RTaW, Nacy, Frace

2 Outlie Cotroller Area Networ CAN) Bacgroud Motivatio ad wor-coservig queues Schedulig odel ad aalysis for priority queues Aalysis for Wor-coservig queues Priority assiget Case study Ipact of wor-coservig queues Epirical ivestigatio Suary ad coclusios Recoedatios 2

3 CAN Bacgroud Cotroller Area Networ CAN) Siple, robust ad efficiet serial couicatios bus for ivehicle etwors Developed origially by BOSCH i 1983, stadardised i 1993 ISO 11898) Average faily car ow has approx Electroic Cotrol Uits ECUs) coected via CAN CAN adatory for cars ad light trucs sold i USA sice 2008 O Board Diagostics) Today alost every ew car sold i Europe uses CAN Sales of icroprocessors with CAN capability approx 750 illio i

4 Schedulig Priority queues) CAN Schedulig Messages copete for access to the bus based o essage ID priority) With each ode ipleetig a priority queue, etwor ca be odelled as if there was a sigle global queue Oce a essage starts trasissio it caot be pre-epted Resebles sigle processor fixed priority o-pre-eptive schedulig Schedulability Aalysis for CAN assuig priority queues) First derived by Tidell i 1994 [14,15] fro earlier wor o fixed priority pre-eptive schedulig Calculates worst-case respose ties of all CAN essages Used to chec if all essages eet their deadlies i the worst-case Sigificat flaws i the origial aalysis corrected by Davis et al. [6] i

5 Motivatio: Wor-coservig ad FIFO queues Aalysis i [6] oly holds if every ode ca always eter its highest priority ready essage ito bus arbitratio This ay ot always be the case: Device drivers ay ipleet FIFO rather tha priority queues Sipler to ipleet, less code / lower CPU load It ay ot be possible to abort a lower priority essage i a trasit buffer A issue if there are fewer trasit buffers tha trasitted essages The CAN cotroller ay eter essages ito bus arbitratio based o trasit buffer uber rather tha essage ID priority) May result i high priority essages beig delayed by lower priority oes placed i trasit buffers with lower ubers Precise queuig policy used ay be difficult to quatify / aalyse 5

6 Wor-coservig queues Recogise that the syste itegrator ay lac coplete iforatio e.g. whe CAN odes supplied by 3 rd parties) Wor-coservig queues Assue oly that the cos stac o a ode esures that if there are ready essages queued by a API call but ot yet trasitted) the oe of the will be etered ito arbitratio wheever arbitratio start o the bus Wor-coservig: re-orderig ot peritted WQN) Arbitrary wor-coservig queue but evertheless esures that istaces of the sae essage are ot re-ordered Exaple: FIFO Wor-coservig: re-orderig peritted WQR) Arbitrary wor-coservig queue ay re-order istaces of the sae essage Udesirable i practice but is the ost geeral case Exaple: LIFO 6

7 Schedulability Aalysis: Model T Trasissio starts J C Iitiatig evet Message queued ready to trasit D R Trasissio copletes Each CAN essage has a: Uique priority idetifier) Maxiu trasissio tie C Miiu iter-arrival tie or period T Deadlie D Maxiu queuig itter J Additioal otatio for wor-coservig queues Group M) set of essages trasitted by the ode that trasits essage L lowest priority of ay essage i group M) f bufferig tie logest tie that a istace of essage ca tae fro beig queued to beig able to eter ito priority based arbitratio f = 0 for priority queued essages) 7

8 Ipact of bufferig delay High priority essages delayed fro eterig priority based arbitratio due to a wor-coservig rather tha priority based queuig policy ca ipact the schedulability of essages set by other odes Fro the perspective of other odes o the etwor, such a essage ca be odelled as havig additioal itter equal to its axiu bufferig tie f Allows aalysis to be derived fro the case where all odes use priority queues 8

9 9 Schedulability Aalysis for Priority Queues Schedulability test derived fro [6]) for essage set via a priority queue o a heterogeeous etwor Worst-case respose tie for a istace of essage occurs withi a priority-level busy period Assuig istaces of all higher priority essages released at the start of the busy period with axiu itter, with subsequet istaces of these essages released as soo as possible Blocig at the start of the busy period due to logest lower priority essage: Busy period: Nuber of istaces of essage ready i the busy period: ) ax ) lp B C = M hp M hep C T f J v C T J v B v = ) ) ) ) 1 + = + T J v Q 1

10 Schedulability Aalysis for Priority Queues Iterval fro start of busy period to whe istace q of essage begis trasissio w J f τ bit w q) = B + qc + C hp ) T Iteratio starts with w q) = B + qc eds whe w q) = w q) + 1 or whe J + w q) qt + C > D Respose tie of istace q of essage : R q) = J + w q) qt + C Worst-case respose tie of essage : R = ax R q)) q= 0.. Q 1 Note the ipact of a essage set by aother ode ipleetig a worcoservig queue. The queuig policy ca delay essage fro eterig priority based arbitratio by up to f ad hece the ipact o essage ca be odelled as essage havig additioal itter equal to f 10

11 Aalysis for wor-coservig queues Ai: To obtai a upper boud o the worst-case respose tie for essage i group M) set by a ode ipleetig a worcoservig queuig policy i a heterogeeous etwor Strategy: Mae pessiistic) worst-case assuptios: Assue that all essages i M) are trasitted at the lowest priority L of ay essage i the group This caot result i shorter respose ties for ay of the istaces of essages fro M) The ode sedig essages i M) ca the be odelled as ipleetig a priority queue, albeit with ust oe priority Assue that istaces of essage lose ties i this priority queue to istaces of other essages i the group Model essages set by other odes as beig priority queued with additioal itter of f use aalysis derived fro that for priority queues 11

12 12 Schedulability Aalysis for Wor- Coservig Queues Schedulability test for essage set via a wor-coservig queue o a heterogeeous etwor Pessiistic assuptio that all essages i M) are trasitted at priority L Blocig at the start of the busy period due to logest lower priority essage: Busy period: Nuber of istaces of essage ready i the busy period: + = + T J v Q 1 M L hp M L C T f J v C T J v B v = ) ) ) 1 ) ax ) L lp B C =

13 Schedulability Aalysis for Wor- Coservig Queues Geeral fraewor: Copute legth of the iterval fro start of busy period to whe istace q of essage begis trasissio + 1 * ) = + +, w + J + f + τ bit w q B L qc I q w ) + C hp L M T ) ) Iteratio starts with w q ) = BL + qc eds whew q) = w q) + 1 or whe J + w q) qt + C > D Respose tie of istace q of essage : R q) = J + w q) qt + C Worst-case respose tie of essage : R = ax q= 0.. Q 1 R q)) * Need to istatiate the iterferece ter I q, w ) for other essages M) set by the sae ode for the wor-coservig queuig policy 13

14 14 Schedulability Aalysis for Wor- Coservig Queues For wor-coservig queues WQN) that do ot perit re-orderig of istaces of the sae essage: With re-orderig of istaces WQR): later istaces of the sae essage M bit WQN C T J w w q I + + = ) ), τ = M bit WQR C T J w w q I ) ), τ bit C q T J w ) 0, ax τ

15 Schedulability test for Wor- Coservig Queues Bufferig delay: Upper boud give by f = R J C Proble: If priorities of essage groups are iterleaved, the bufferig delay of oe essage ca deped o the respose tie of aother essage ad viceversa Resolved by otig that bufferig delays are ootoically o-decreasig w.r.t. respose ties ad viceversa 15

16 Adacet priority orderig Adacet priority orderig: Messages withi a group M) have adacet priorities o iterleavig with other essages Optial partial orderig: If a priority orderig Q exists that is schedulable accordig to the previous schedulability test, the a schedulable adacet priority orderig also exists Regardless of the priority orderig of priority queued essages, all essages sharig a worcoservig queue should have adacet priorities but ot ecessarily cosecutive values) PQ-1 WQ-1 PQ-2 PQ-3 WQ-2 PQ-4 WQ-3 PQ-5 PQ-1 PQ-2 PQ-3 PQ-4 WQ-1 WQ-2 WQ-3 PQ-5 16

17 Adacet priorities With Adacet Priorities: No eed to accout for bufferig tie so f = 0 for all essages set via wor-coservig queues This is because if a essage is of higher priority tha essage, the crucially, so are all of the other essages that share the queue with, hece all cotribute to the queuig delay of essage, ad the order i which they are actually set o the bus is irrelevat Settig f = 0 for all essages: siplifies the aalysis o repeats of the while loop ust calculate the essage respose ties) Reoves a sigificat aout of pessiis 17

18 Optial priority assiget OPA-FP/WQ algorith: Based o Audlsey s greedy Optial Priority Assiget OPA) algorith Optial for etwors with a ix of priority queues ad wor-coservig queues w.r.t. the schedulability test preseted 18

19 Case Study: Autootive Geerated usig NETCARBENCH 16 ECUs, 80 essages, oially 500 Kbit/s, load 36.5% Message periods 20, 50, 100, 200, or 1000 s 5-8 data bytes i each essage Deadlies equal to periods, queuig itter 5s, 18 essages set via a gateway 9 of which coe fro aother etwor. The deadlie of these essages is 2x period, ad itter = period Experiets Cofig. 1: All priority queues, essage priorities i trasissio deadlie ootoic order TDMPO) i.e. D-J order Cofig. 2: Gateway assued to use a wor-coservig queue, essage priorities agai TDMPO Cofig. 3: As Cofig 2, but usig OPA-FP/WQ algorith to set essage priorities 19

20 Cofig 1: All priority queues Mi bus speed 191 Kbit/s Max bus Util. 95.8% Tie s) Messages TDMPO) Deadlie - Jitter PQ-odes Gateway 20

21 Cofig 2: Gateway WQN Tie s) Mi bus speed 624 Kbit/s +230%) Max bus Util. 29.3% Messages TDMPO) Deadlie - Jitter PQ-odes Gateway WQN-ode 21

22 Cofig 3: Gateway WQN Tie s) Mi bus speed 393 Kbit/s +105%) Max bus Util. 46.6% 10 Deadlie - Jitter PQ-odes Gateway WQN-ode Messages TDMPO-FP/WQ) 22

23 Case Study: Suary Cofig. Node type Priority order Mi. bus speed Max. bus Utilisatio 1 All PQ TDMPO 191 Kbit/s 95.8% 2 Gateway WQN TDMPO 624 Kbit/s +230%) 29.3% 3 Gateway WQN OPA- FP/WQ 393 Kbit/s +105%) 46.6% 23

24 Epirical evaluatio Exaied 10,000 radoly geerated sets of essages: 80 essages i each set, 8 data bytes per essage 8 odes o the etwor Rado allocatio of essages to odes Log-uifor distributio of essage periods 10s 1000s First ode assued to be a gateway: essage deadlies = 2x period, itter = period Other odes: essage deadlies = period, itter uifor distributio 2.5 5s) 11-bit idetifiers Cofiguratios Cofig. 1: All PQ odes - TDMPO Cofig. 2: Two WQN odes TDMPO-WQ/FIFO Cofig. 3: Four WQN odes TDMPO-WQ/FIFO Cofig. 4: All WQN odes TDMPO-WQ/FIFO Cofig. 5: All PQ odes rado priorities 24

25 Epirical results 1600 #1 PQ No WQN odes) #5 PQ - Rado Priorities #4 WQN All WQN odes) #2 WQN ad PQ Quarter WQN odes) #3 WQN ad PQ Half WQN odes) #4 WQN All WQN odes) #5 PQ - Rado Priorities Frequecy 800 #1 PQ No WQN odes) 600 #3 WQN ad PQ Half WQN odes) #2 WQN ad PQ Quarter WQN odes) Breadow Utilisatio 25

26 Evaluatio Epirical evaluatio of 10,000 essage sets 8 odes, 80 essages, 8 data bytes per essage periods s log uifor distributio) itter 2.5-5s uifor distributio) 1 gateway: deadlies = 2x period, itter = period Cofig. Node type Priority order Average Max. bus utilisatio 1 All PQ TDMPO 85.5% 89.5%) 2 2 WQN, 6 PQ TDMPO-FP/WQ 49.9% 62.7%) 3 4 WQN, 4PQ TDMPO-FP/WQ 38.0% 44.9%) 4 All WQN TDMPO-FP/WQ 25.5% 28.4%) 5 All PQ Rado 16.4% 18.4%) Figures i bracets are without larger deadlies ad itter for essages set by the gateway 26

27 Suary ad Coclusios Itroduced sufficiet schedulability test for CAN etwors with a ix of odes usig wor-coservig e.g. FIFO) ad priority queues Exteded previous aalysis for FIFO queues ad costraied deadlies to wor-coservig queues ad arbitrary deadlies Aalysis reduces to that for FIFO queues whe all essages have costraied deadlies see paper) FIFO aalysis fro ECRTS 11 holds for arbitrary worcoservig queues whe all essages have costraied deadlies. For wor-coservig queues ad the aalysis preseted, Adacet priority orderig is optial withi each essage group Modified OPA algorith provides a optial priority orderig w.r.t. our aalysis) for a set of essages set via worcoservig priority queues 27

28 Suary ad Coclusios Exaied perforace of wor-coservig queues / aalysis via case study ad epirical evaluatio Sigificat reductio i perforace icreased bus speed is required ad a large decrease i ax. bus utilisatio e.g. 80% dow to 30%) Maily caused by uavoidable priority iversio Why are FIFO queues used Mae the device driver ore efficiet less processor load) Easier to ipleet But local gai coes at a cost uderiig priority based arbitratio o CAN sigificat perforace pealty 28

29 Recoedatios To obtai the best possible perforace Use a appropriate priority orderig e.g. based o trasissio deadlies) Use priority queues Avoid usig wor-coservig / FIFO queues FIFO wheever possible Arbitrary wor-coservig / FIFO queues ca cause sigificat perforace degradatio Whe there are ay essages i a queue, with a rage of trasissio deadlies that iterleave with those of other essages o the etwor result is sigificat priority versio 29

30 Questios? 30