Analysis and Op,miza,on of Mixed- Cri,cality Applica,ons on Par,,oned Distributed Architectures

Transcription

1 Analysis and Op,miza,on of Mixed- Cri,cality Applica,ons on Par,,oned Distributed Architectures Domițian Tămaș- Selicean, Sorin Ovidiu Marinescu and Paul Pop Technical University of Denmark

2 Outline MoBvaBon SeparaBon of mixed- cribcality applicabons At processing element level At communicabon level Problem formulabon and example OpBmizaBon strategy Experimental results Conclusions 2

3 Mo,va,on Safety is the property of a system that will not endanger human life or the environment A safety- related system needs to be cer,fied A Safety Integrity Level (SIL) is assigned to each safety related funcbon, depending on the required level of risk reducbon There are 4 SILs: SIL4 (most cribcal) SIL1 (least cribcal) SIL0 (non- cribcal) not covered by standards SILs dictate the development process and cerbficabon procedures 3

4 Mo,va,on Real Bme applicabons implemented using distributed systems Mixed- cribcality applicabons share the same architecture SIL3 Federated Architecture SIL4 SIL1 Integrated Architecture SIL3 SIL2 SIL4 SIL4 SIL1 ApplicaBon A 1 PE ApplicaBon A 2 ApplicaBon A 3 SoluBon: parbboned architecture 4

5 Separa,on at PE- level ParBBon = virtual dedicated machine ParBBoned architecture SpaBal parbboning protects one applicabon s memory and access to resources from another applicabon Temporal parbboning parbbons the CPU Bme among applicabons 5

6 Separa,on at PE- level PE 1 PE 2 PE 1 PE 2 PE 3 PE 3 ParBBon ParBBon slice Major Frame Temporal parbboning StaBc parbbon table Repeated with a period MF ParBBon switch overhead Each parbbon can have its own scheduling policy A parbbon has a certain SIL 6

7 Separa,on at Network- level ES 1 ES 3 NS 1 NS 2 ES 2 End System Network Switch ES 4 Full- Duplex Ethernet- based data network for safety- cribcal applicabons Compliant with ARINC 664p7 Aircraa Data Network 7

8 Separa,on at Network- level ES 1 τ 1 ES 3 vl 2 τ 2 τ 5 NS 1 NS 2 ES 2 τ4 vl 1 virtual link τ 3 ES 4 Highly cribcal applicabon A 1 : τ 1, τ 2 and τ 3 τ 1 sends message m 1 to τ 2 and τ 3 Non- cribcal applicabon A 2 : τ 4 and τ 5 τ 4 sends message m 2 to τ 5 8

9 Separa,on at Network- level ES 1 τ 1 dataflow link l 1 l 2 dp 1 l 3 τ 2 τ 5 ES 3 NS 1 NS 2 l 4 ES 2 τ4 vl 1 dp 2 ES 4 dataflow path τ 3 Highly cribcal applicabon A 1 : τ 1, τ 2 and τ 3 τ 1 sends message m 1 to τ 2 and τ 3 Non- cribcal applicabon A 2 : τ 4 and τ 5 τ 4 sends message m 2 to τ 5 9

10 TTEthernet Traffic classes Time Triggered (TT) based on stabc schedule tables Rate Constrained (RC) determinisbc unsynchronized communicabon ARINC 664p7 traffic Best Effort (BE) no Bming guarantees provided 10

11 Applica,on Model SCS apps transmit TT messages FPS apps transmit RC messages 11

12 Problem formula,on Given A set of applicabons The cribcality level (or SIL) of each task A set of N processing elements (PEs) and topology of the network The set of TT and RC frames The set of virtual links The size of the Major Frame and of the ApplicaBon Cycle Determine The mapping of tasks to PEs The sequence and length of parbbon slices on each processor The assignment of tasks to parbbons The schedule for all the tasks and TT frames in the system Such that All applicabons meet their deadline The response Bmes of the FPS tasks and RC frames is minimized 12

13 Mapping and parbboning opbmizabon Mo,va,onal Example 1 13

14 Mo,va,onal Example 1 14

15 Mo,va,onal Example 1 15

16 OpBmizaBon of TT message schedules Mo,va,onal Example 2 vl 1 ES 1 vl 2 NS 1 ES 3 ES 2 vl 3 period (us) deadline (us) C i (us) M f1 F RC vl 1 f2 F TT vl 2 f3 F TT vl 3 16

17 IniBal TT schedule Mo,va,onal Example 2 ES 1 vl 1 period (us) deadline (us) C i (us) M vl 2 NS 1 ES 3 f1 F RC vl 1 f2 F TT vl 2 ES 2 vl 3 f3 F TT vl 3 17

18 OpBmized TT schedule Mo,va,onal Example 2 ES 1 vl 1 period (us) deadline (us) C i (us) M vl 2 NS 1 ES 3 f1 F RC vl 1 f2 F TT vl 2 ES 2 vl 3 f3 F TT vl 3 18

19 Op,miza,on Strategy Tabu Search meta- heurisbc Task mapping and parbbon slice opbmizabon (TO) Considering TT frame schedules fixed TT frame schedules opbmizabon (TM) Considering the task mapping and parbbon slices fixed Tabu Search Minimizes the cost funcbon Explores the solubon space using design transformabons 19

20 Op,miza,on Strategy Degree of schedulability Captures the difference between the worst- case response Bme and the deadline Cost FuncBon Cost( )= c1 = Ai max(0,r i D i ) ifc 1 > 0 c 2 = Ai (R i D i ) ifc 1 = 0 20

21 Op,miza,on Strategy: Design Transforma,ons ParBBon slice moves resize parbbon slice swap two parbbon slices join two parbbon slices split parbbon slice into two Task moves re- assign task to another parbbon 21

22 Op,miza,on Strategy: Design Transforma,ons 22

23 Op,miza,on Strategy: Design Transforma,ons 23

24 Op,miza,on Strategy: Design Transforma,ons 24

25 Op,miza,on Strategy: Design Transforma,ons 25

26 Op,miza,on Strategy: Design Transforma,ons 26

27 Op,miza,on Strategy: Design Transforma,on Task re- assignment move To another parbbon of the same applicabon To a parbbon of another applicabon To a newly created parbbon Empty parbbons are deleted 27

28 Op,miza,on Strategy: Design Transforma,ons TT frame moves advance frame transmission Bme advance frame predecessors transmission Bme postpone frame transmission Bme postpone frame successors transmission Bme RC frame moves reserve space for RC frame resize reserved space for RC frame remove reserved space for RC frame 28

29 Frame Representa,on for Moves ES 1 ES 3 NS 1 NS 2 ES 2 vl 1 ES 4 [NS f 1, NS 2 ] 1,1 [ES f 1, NS 1 ] 1,1 [NS f 1, NS 2 ] 1,1 [NS f 1, NS 2 ] 1,1 29

30 Design transforma,ons: Postpone move 30

31 Design transforma,ons: Advance move 31

32 Design transforma,ons: Reserve space for RC 32

33 Design transforma,ons: Resize RC reserved space 33

34 RC Frame End- to- End Analysis On a dataflow link, a RC frame can be delayed by: scheduled TT frames queued RC frames technical latency policy specific: Bmely block pre- empbon 34

35 ES 1 vl 2 vl 1 NS 2 vl 3 NS 1 ES 4 vl 4 NS 3 NS 2 NS 1 f 3,j NS 3 NS 1 NS 1 ES 4 RC Frame End- to- End Analysis ES 1 NS 1 f 1,i Let us assume frame f 5 has a deadline of 775 µs. f 4,1 f 2,1 f 4,1 f 3,j f 1,i The worst-case end-to-end delay R C [NS fi of an RC frame f i 1, ES 4 sent on a virtual link vl ] i = M ( f i ) is the sum of the worst- Q queueing TL TT NS delays 1 Q [ν j,ν k ] f i [NS 1, ES 4 ] on each Q RC [NS 1, network ES 4 ] node (ES or NS) ν j (which is the source of ar dataflow f1 link [ν j,ν k ] vl i ) and the tr mission duration Q [ν C [ν j,ν k ] j,ν f i k ] for f i = Q TT each f i,[ν j,ν k ] + dataflow link QRC f i,[ν j,ν k ] + [ν j,ν k ] vl frame transits: QTL ν j f 1 In ca frame instance misses its deadline due to a blank interval bi [ i on dataflow link [ν j,ν k ], the algorithm can apply either a rem blank or a resize blank move to the blank interval. In Fig. 9 apply a resize blank move on the blank interval bi 1 on dataflow [NS f 2,1 1,ES 2 ]. Thus, we resize bi 1 from 175 µs to 100 µs, whic lows us to advance the scheduled send time for frames f 4 an and consequently, allows frame f 5 to be delivered before its d line. 7.2 RC frame end-to-end delay analysis R fi = (Q [ν j,ν k ] f i +C [ν j,ν k ] f 1 ) ν j,ν k V [ν j,ν k ] vl i 35

36 RC Frame End- to- End Analysis Approaches for analysis of ARINC 644p7 network traffic: Network Calculus, (Boyer, 2008) Finite State Machine, (Saha, 2007) Timed Automata, (Adnan, 2010) Trajectory Approach, (Bauer, 2009) We use the method proposed in (Steiner, 2011) it takes into account also the TT traffic it is pessimisbc: does not ignore frames that already delayed a RC frame on a previous link assumes uniformly distributed intervals of equal length reserved for RC traffic 36

37 Experimental Results: TO Benchmarks 5 synthebc 2 real life test cases from E3S TO compared to: Straighnorward SoluBon for Tasks (SST) Simple parbboning scheme, each applicabon A i is allocated a total Bme proporbonal to the ublizabon of tasks of A i on the processor they are mapped to 37

38 Experimental Results: TO Set Tasks PEs SST Sched. TO Sched. avg. % Tasks Tasks increase in d All All All All All All All All

39 Experimental Results: TO Set Tasks PEs SST Sched. TO Sched. avg. % Tasks Tasks increase in d All All All All All All All All

40 Experimental Results: TM Benchmarks 7 synthebc 1 real life test case based on the SAE AutomoBve benchmark TM compared to: Straighnorward SoluBon for Messages (SSM) Builds TT schedules with the goal to opbmize the end- to- end response Bme of the TT frames without considering the RC traffic 40

41 Experimental Results: TM Set Test case ES NS Messages Frame D cost instances [%] automotive

42 Experimental Results: TM Set Test case ES NS Messages Frame D cost instances [%] automotive

43 Experimental Results: TM Set Test case ES NS Messages Frame D cost instances [%] automotive

44 Conclusions ApplicaBons of different cribcality levels can be integrated onto the same architecture only if there is enough separabon: SeparaBon at PE- level achieved with IMA. SeparaBon at network- level using TTEthernet. We proposed a Tabu Search based opbmizabon of task mapping and allocabon to parbbons, and of Bme parbbons. Only by opbmizing the implementabon of the applicabons, taking into account the parbcularibes of IMA and TTEthernet, are we able to support the designer in obtaining schedulable implementabons. 44

45 45