A protocol for mixed-criticality management in switched Ethernet networks

Transcription

1 A protocol for mxed-crtcalty management n swtched Ethernet networks Olver CROS, Laurent GEORGE Unversté Pars-Est, LIGM / ESIEE, France cros@ece.fr,lgeorge@eee.org Xaotng LI ECE Pars / LACSC, France xaotng.l@ece.fr Abstract In real-tme ndustral networks, provdng tmng guarantees for applcatons of dfferent crtcaltes often results n buldng separate physcal nfrastructures for each type of network at the prce of cost, weght and energy consumpton. Mxed-Crtcalty (MC) s a soluton frst ntroduced n embedded systems to execute applcatons of dfferent crtcalty on the same platform. In order to apply MC schedulng to off-the-shelves Swtched Ethernet networks, the key ssue s to manage the crtcalty change nformaton at the network level. The objectve of ths work s to propose a crtcalty change protocol for MC applcatons communcatng over Swtched Ethernet networks. The protocol reles on a global clock synchronzaton, as provded by the IEEE-1588v2 protocol, and a real-tme multcast based upon t, to preserve the consstency of the crtcalty level nformaton stored n all the Ethernet swtches. We characterze the worst-case delay of a crtcalty change n the network. Smulaton results show how the crtcalty change delay evolves as a functon of the network load. I. INTRODUCTION Nowadays, hghly-constraned ndustral systems found n defense, publc transports or home automaton have ncreasng needs n terms of relablty and performance. It s a common stuaton for such systems to ntegrate several ndependent network archtectures n order to transmt, n each network, messages of dfferent crtcalty (e.g. n bus: passenger nformaton, mechancal control nformaton, speed, etc.) such that each system can be certfed n solaton. Ths soluton s very expensve n terms of cost, weght and hence n terms of energy consumpton, as each network must have ts own nfrastructures, materals and wres. For example, the mechancal functons, trajectory control and the passenger nformaton are often treated n separated nfrastructures nsde a publc bus, wth dfferent dedcated materals. One soluton to ths problem s the mxed-crtcalty (MC) schedulng approach frst proposed n the context of unprocessor and multprocessor systems [1]. It executes several applcatons of dfferent crtcaltes on the same platform by adaptng certfcaton effort to the level of assurance needed at a gven crtcalty level. Snce a networked system s an nterconnect system for applcatons of dfferent crtcaltes, the objectve of ths work s to study how to manage crtcalty level nformaton n networked systems. Wth MC schedulng, each task s characterzed by the maxmum crtcalty level t s allowed to execute. A task can be non-crtcal, crtcal for the msson (msson-crtcal), crtcal for the safety of the vehcle (mechancal-crtcal), or for the safety of ts occupants (safety-crtcal). Each crtcalty level must provde guarantees on end-to-end transmsson delays, especally for hgh crtcal tasks. The more crtcal a task s, the more relable the guarantee should be. In the real-tme network context, we focus on how to ntegrate crtcalty management n networked systems. The frst pont s to bound the number of crtcalty levels we use. Baruah [2] showed that the complexty of MC schedulng problems s NP-hard n the strong sense. In order to lmt the complexty of our archtectures, we focus on a two-level crtcalty network, as presented n [3]. These two crtcalty levels are called low-crtcal (LO) and hgh-crtcal (HI) levels: only a set of predefned messages can be transmtted n HI crtcalty level, whereas all messages can be transmtted n LO crtcalty level. The man goal of the crtcalty management n networked systems conssts n provdng Qualty of Servce (QoS) guarantees (n terms of worst case end-to-end transmsson delays), specally for hgh crtcal messages. In ths context, we focus on a method to grant the consstency of the crtcalty level nformaton n a Swtched Ethernet network, to ensure bounds on end-to-end transmsson delays of messages as a functon of the crtcalty of nformaton sent. In unprocessor and multprocessor systems, the problem of characterzng the mpact of low-crtcalty tasks when a crtcalty change from LO to HI has been consdered by boundng the demand of carry-n jobs. Assurng determnstc communcatons n networked systems mples to be able to bound the end-to-end delay of all messages. In [4], we proposed a tool to evaluate the worst case end-to-end delay of any message sent on a Swtched Ethernet network relyng on a global clock-synchronzaton protocol. In ths paper, we also consder a clock synchronzed network, synchronzed wth the IEEE-1588v2 synchronzaton protocol and ts mplementaton n Precson Tme Protocol (PTP). On top of the clock synchronzaton, we buld a relable multcast to consstently swtch the crtcalty level on all the nodes of the network. In ths paper, frst we present n II the network model studed n ths work. In III, we llustrate a lnk utlzaton problem based on an example, and then we present the mportance of managng MC n network context. We propose a MC management protocol n IV. Fnally, we show by smulaton the performances of ths protocol n V, and conclude ths paper

2 as well as future perspectves of ths paper n VI. II. NETWORK MODEL A. Mxed-Crtcalty In ths paper, we consder a tree-based topology as those found n applcaton domans lke avoncs systems and recent and futur publc transport systems [5]. The network s composed of a set of nterconnected nodes, all organzed accordng to a tree-based structure wth one fnal collectng node denoted the snk node. An example of such topology s the one showed n fgure 1, wth S 4 as the snk node. ES 1 ES 2 ES 4 ES 5 ES 3 S 1 S 2 S 3 S 4 Fg. 1. Centralzed Network archtecture ES out Our goal s to propose a crtcalty level management protocol n a tree-based Swtched Ethernet network topology. The snk node s n charge of strorng the network crtcalty nformaton. All the nodes of the network have a local copy of the network crtcalty level nformaton. The protocol we propose should mantan the consstency of the crtcalty level nformaton n all the nodes of the network n the case of a crtcalty swtch. We then have two cases: If the network crtcalty level s LO, the frst node sendng a request to the snk-node, to change the crtcalty level to HI wll result n a multcast sent by the snk node the all the nodes of the network wth the new network crtcalty level. All the nodes recevng ths message should update ther local crtcalty level nformaton so as to keep t consstent after a crtcalty mode swtch. If the network crtcalty s HI, a swtch to LO crtcalty level can happen only f the snk node has receved from all the other nodes a request to swtch to LO mode. A smple multcast s not suffcent to guarantee the consstency of the network crtcalty level nformaton n all the nodes. We ntroduce a real-tme relable multcast protocol n secton IV-C as one soluton to ths consstency problem. We caracterze the maxmum tme needed to swtch the network crtcalty level from LO to HI, from the request of the frst node wllng to change the crtcalty level to HI, to the tme all nodes are allowed to change ther crtcalty level (after recevng the relable multcast from the snk node). B. Notatons and man hypothess In MC systems, representng dfferent levels of crtcalty nsde a system s mostly based on a choce between two dfferent hypotheses : ether a message has a dedcated worstcase transmsson tme (WCTT) for each crtcalty level, whch means that the flow of data sent n the entry ponts are longer n the case of HI modes. For example, as a plane s landng, t mght need more precse evaluaton of the alttude. It means that the alttude sensors wll send more complete, and so longer, data values. The second hypothese s not based on longer WCTT, but on a more frequent messages. Each message has now two dfferent perods, one for LO level and one for HI level. It corresponds to ncreasng the number of measures durng a crtcal phase : for example, durng landng, the measure of speed or alttude could have to be more frequent. We consder ths case n our analyss. A network s a set of nterconnected swtches communcatng through full-duplex lnks connectng end-systems. On each lnk, we can send one or several flows v, and each flow produces several messages. Each flow v s represented as a 3-tuple v = {P,C, T } where: P s the path of nodes followed by any message of v, startng from a source node to a destnaton node. We consder ths path as statcally defned by the desgner. C defnes the WCTT of any message of v sent n LO or HI network crtcalty level. Its perod s defned as T = {T LO,T HI }. Itsavector of dfferent perods of the flow, correspondng to LOcrtcal and HI-crtcal perod (we assume two network crtcalty levels n ths paper). In the case where a flow can only be sent n LO crtcalty level, that means that the flow wll not be sent by a swtch when the network crtcalty level s HI. Furthermore, we suppose that: Each message s ndependent from each other All message transmssons are non-preemptve All the swtchs use a Fxed Prorty (FP) schedulng wth FIFO schedulng n a specfc FP queue, denoted as FP/FIFO schedulng. III. PROBLEM STATEMENT A. An example We consder the case of a smple network composed of one swtch S (denoted S M s t support MC), schedulng flows wth FIFO schedulng and havng three entry ports ES1, ES2 and ES3 respectvely recevng flows v 1, v 2 and v 3, wth the followng parameters: Flow T LO (μs) T HI (μs) C (μs) u LO u HI v v v Imagne a scenaro where all the three flows are transmtted n the LO crtcalty level. The LO-utlzaton (u LO )ofthe network at the most loaded node S 4 s then u LO = u LO 1 + u LO 2 + u LO 3 = Then flow v 1 and flow v 2 ncrease ther workloads by reducng the perods of messages due to certan emergences. Then flow v 1 and flow v 2 are transmtted

3 n HI crtcalty level. Supposng that there s no crtcalty management, now the utlzaton at the node S 4 should be u HI 1 + u HI 2 + u LO 3 =1.13. It means that S 4 s overloaded n a mxed mode wth both crtcalty levels. We focus on the mpact of such an overloaded lnk on transmsson delays when crtcalty levem s not managed by the nodes. We suppose that, at t = 100 μs, the system becomes hgh-crtcal : v 1 and v 2 start emttng messages accordng to T HI and no more to T LO. Bascally, ths results n a strong ncrease n the transmsson delay for the frames of v 1 (see S n fgure 2). ES 1 ES 2 ES 3 S S MC Fg. 2. Transmsson delay wth crtcalty management We can observe that the transmsson delay of messages from v 1 ncreases drastcally wth tme. In fact, we can easly compute that the watng delay of each message from v 1 n the entry pont of S 4 ncreases of 50 ms at each new emsson. Thus, n classcal Ethernet context, swtches do not have nput buffers of nfnte sze, so a too hgh watng delay can result n droppng out a message, and then loss of data. If the network supports MC management (see S MC n fgure 2), we can observe that the transmsson delay of HIcrtcal messages s constant and that crtcalty management allowed us to fx the overload problem. We show next a protocol to mplement MC n network context. IV. A CRITICALITY-CHANGE PROTOCOL A. A two-phase protocol Transmttng and managng crtcalty level nsde network topology mples two dfferent condtons. Frst, we need to assure that all nodes n the network have the same crtcalty level. Secondly, n case of crtcalty level swtch, we have to be sure that all nodes change ther local crtcalty level nformaton preservng the consstency the network crtcalty level nformaton. Lke we showed n II, crtcalty level s managed by a central entty (the snk node). It means that, even f each node has ts own crtcalty level, t must be synchronzed at all tmes wth the one of the snk node. To assure ths condton, we propose our MC managng protocol. It conssts n assurng the consstent change of crtcalty level n all the nodes of a network topology. For ths, we use a relable multcast to ensure a total order (updates are scheduled n the same order n all nodes) for the update of crtcalty level nformaton n all nodes. Ths preserves the seralsablty of the crtcalty updates hence the consstency of the crtcalty level nformaton [6]. Ths MC managng protocol works n two phases for a LO to HI crtcalty change. The node n n charge of ntatng ths crtcalty level change sends a crtcalty change request to the snk node. We name t the swtch-crtcalty call (SCC) message, the delay needed to send ths message from node n to the snk s denoted Idelay n. Next, we need to send a crtcalty swtch message wth the new crtcalty level to all the nodes (except the snk node) of the network. Ths s the relable multcast phase ntated by the snk node that send a tmestamp of hs local clock n the multcast message (recall that we assume a global clock synchronzaton). The delay needed to send and receve ths message from the snk node to node n s denoted Mdelay n. The total crtcalty swtch delay S delay n the network N can be computed by : S delay =max n N (In delay + M n delay) (1) Upon recepton of a crtcalty swtch message, each node wll have to localy determne when to to the crtcalty swtch. We now how yo characterze the crtcalty swtch delay n the case of a LO to HI crtcalty swtch. B. Swtch-crtcalty call Suppose that the network crtcalty s LO. Callng for a crtcalty level change to HI conssts n sendng a specfc message from a node n n the network requrng ths crtcalty level change, and transmttng t to the central node responsble of crtcalty management. Ths SCC message s transmtted wth the hghest prorty (except PTP messages), dedcated for confguraton messages. Nevertheless, t can be delayed by other messages n the network: ether by PTP synchronzaton messages, or by other messages due to the non-preemptve effect (a message even wth the lowest prorty cannot be stopped once ts transmsson has started). The SCC message s consdered as a new message n the network. It means that t s defned by ts own path P c from n to the central node S h and ts WCTT C c. We consder that only one SCC message s sent by a swtch (the frst one receved). If another nodes ntate another SCC message wth the same crtcalty level, the swtch recevng t wll dscard the message. In other words, computng the delay needed to transmt the SCC conssts n evaluatng the delay needed for the message to be tranmtted from one node n to the snk node. To do ths, we use the trajectory approach, presented n [7]. The SCC

4 transmsson delay (noted as Idelay n ) can be computed by the general trajectory approach expresson. Gven that the call s done by a node n n the network, we apply the trajectory approach computaton method to the swtch-crtcalty call message c. Ths gves us the followng result : I n delay = ( 1+ j hpc ( 1+ j spc j spc hpc h P c\{n} W last c,j c,t t + S frst c,j max c ( max M frst c,j T LO j M frst c,j T LO j ) + A c,j + C j (2) + A c,j ) + C j (3) h P j (C j)) (4) +( P c 1) sl (5) h P c δ h c (6) C c (7) (2) s the delay nduced by messages wth hgher prortes than the one of message c. These messages can delay c f they arrve at one shared output port wth c durng the maxmzed nterval. For a hgher-prorty flow j, ts maxmzed arrval jtter s A c,j = Smax frstc,j c S frstc,j mn c, where (resp. Smn h c ) s the maxmum (resp. mnmum) S h max c delay of the message c from ts source node frst c tll the output port h. (3) s the delay nduced by messages wth the same prorty as message c. They are scheduled by FIFO polcy. Accordng to FIFO, messages arrvng at the output port where frst c,j (the frst output port where they meet message c) durng the maxmzed temporal nterval [M frstc,j,t+ Smax frstc,j c ] can delay message c. (4) ndcates the transmsson delay of a message sequence ncludng message c. Ths delay s maxmzed by consderng the transmsson tme of the largest frame n the sequence at each output port along P c. (5) s the electroncal latency nduced by the transmsson through wres (wth sl, the electroncal latency nduced by the transmsson between two nodes) (6) represents the delay nduced by the non-preemptve effect (see [8]) of flows wth a lower prorty (7) fnally, we substract C c because the global transmsson delay Wc,t lastc corresponds to the delay between the emsson of c and ts startng nstant of emsson n the fnal snk node As SCC message has a hgh prorty (dedcated for confguraton messages), the only hgher prorty messages that wll delay us are the PTP synchronzaton messages. We consder that PTP messages and swtch-crtcalty call messages are the only one to be sent n ths level of prorty and hgher, so there s no other message wth the same prorty as SCC. It means that we just need to compute the number of PTP messages generated n each node encountered along the path of the swtch-crtcalty call. Moreover, we suppose that PTP messages have ther specfc worst case transmsson tme, noted as C PTP. But, for the sake of readablty, we consder that C c and C PTP have the same value (the hghest one of the two). Even f t s a pessmstc assumpton, we can consder t true as PTP and SCC messages are both confguraton messages, defned wth a small and close number of bytes n Ethernet protocol. Consderng these hypotheses, we obtan (wth F PTP, the PTP synchronzaton frequency) : I n delay = F PTP h P c δ h c j hpc ( S frst c,j max c M frst ) c,j c + A c,j +( Pn 1) (sl +2 Cc) (8) C. Relable multcast In order to update the crtcalty level nformaton from the central node to all the nodes n the network, we must be sure that the crtcalty swtch order s receved by all the nodes and s executed preservng the consstency of the crtcalty level nformaton. Thus, n order to preserve the consstency of the current crtcalty level n all the nodes, we need to guarantee a total order n the crtcalty updates: two consecutve crtcalty swtches have to be executed n the same order n all the nodes. A relable real-tme multcast s a method to send the same nformaton to all nodes n a network provdng total order for the update of the crtcalty level n all nodes. In [6], the authors show how to buld a real-tme relable multcast provded that worst case messages end-to-end delays can be bounded from above. In ths paper, we adapt ther soluton to the context of mxed crtcalty management wth the trajectory approach to compute worst case end-to-end message delays. To mplement ths, we need a determnstc computaton of the transmsson delay of the nformaton. Snce we can assure a bounded transmsson delay for nformaton n each physcal lnk of the network, we wll be able to provde guarantees on transmsson delays n the whole network. That bulds the determnsm of the delay. Suppose a network N, composed of a set of nodes S = {S 1,S 2,...S n,es 1,ES 2,...ES m }.WenoteMdelay n the delay needed by node n to receve the relable real-tme multcast nformaton from the central node. Now, we can compute the transmsson delay of the crtcalty-swtch order (whch s a message) from the central node to all the nodes n the network. Ths multcast delay s noted as M delay. Even f we mplemented PTP, we need to consder clock accuracy ɛ for each node, as t mpacts the relable multcast protocol [6]. The multcast delay of the whole network can then be deduced by takng the maxmum value of accuracy on any node. We have ɛ =max (ɛ n). We obtan [6]: n N M delay =max (M delay)+ɛ n (9) n N Mdelay n, for a node n, s then computed by the addton of dfferent elements: the swtchng latency sl nduced by

5 electroncal transmsson between two nodes, and the WCTT of the crtcalty-swtch message, noted as C o. For clarty purposes, we consder that the WCTT of the swtch-crtcalty message s the same n each node: n N,C o = C n o = C c. For a node n the network wth a central node (snk node) S h, the delay needed to receve the order drectly depends on the dstance between and S h. We note ths dstance d h. Furthermore, we make the hypothess that the swtchng latency sl s the same for each physcal lnk (lke n IV-C), and so that sl =0ms : the electroncal delay generated by the dstance between each node s null. We then obtan the followng expresson of M delay : Mdelay n = d n (C c + sl)+ɛ n M delay = max (d n) (C c + sl)+ɛ (10) n N As we are computng the multcast delay, we are computng the worst case delay needed for the farthest node from the snk node to receve the swtch crtcalty order. At the end of ths multcast delay M delay, we are sure that all the nodes n the network receved the crtcalty change nformaton. The crtcalty swtch occurs on any node at ts local tme t m +M delay, where t m (respectvely M delay ) s the tmestamp (the swtchng delay) sent by the snk node n the crtcalty swtch multcast message. Hence when all nodes have receved the crtcalty swtch request, hence preservng the consstency of the crtcalty level nformaton n all nodes. All the nodes swtch almost at the same tme, wth a tme dfference bounded by ɛ the clock synchronzaton accuracy. D. Crtcalty-swtch message Gven the expresson of Idelay n (8) and M delay(??), we obtan the global expresson of the crtcalty-swtch delay S delay n the network. Gven the hypotheses that the swtchng latency s constant, and that PTP, SCC and the relable multcast message have the same WCTT (noted as C c ), we obtan: S delay = F PTP h P c δ h c j hpc ( S frst c,j max c M frst ) c,j c + A c,j + (2 max(dn) 1) (Cc + sl)+cc(max(dn) 1) n N n N + ɛ (11) When we compute the total delay needed to operate a crtcalty swtch n the network, we then compute the delay represented by PTP synchronzaton messages, the nonpreemptve effect nduced by all messages n the network, the WCTT of crtcalty swtch messages and fnally the delay nduced by electroncal latency and clock jtter. V. SIMULATION In order to provde estmatons of the transmsson delay of crtcalty nformaton nsde a network, we provde smulaton results usng ARTEMIS [4]. To provde these results, we based our approach on the topology descrbed on fgure 1, and on randomly-generated tasksets to smulate traffc load n the topology. To generate these tasksets, we used the UUnfast generaton algorthm presented n [9]. But as ths method was desgned for processor-context smulaton, we frst adapted t to network context. A. UUnfast The UUnfast method s a random tasksets generaton method, frst presented n [9] and used to generated tasksets n mono and multcore contexts. It conssts n three steps : Frst, we generate random perods n a confgurable tme nterval [ɛ; T ], where ɛ s the clock accuracy(see IV) and T s the global smulaton tme. These perods can be of any sze. On a second pont, we generate a random value n [0, 1] accordng to a unform law, called the utlsaton of a frame. It represents the ndvdual load represented by the frame. Then, based on the generated perod and utlsaton, we compute the WCET of the task. When t comes to adapt UUnfast method to a network context, t results n two dfferent problems to solve. Frst, when we generate a random flowset, we have to specfy a targetted load for the whole system (to compute the utlsaton). But n the network, the load s dfferent n each node. To solve ths, we decded to focus on the load on the snk node of the topology : as t s the central node, we assume ths s the one wth the hghest load, or at least wth the most mportant one to focus on. Secondly, we had to modfy the UUnfast method n order to generate LO and HI crtcal messages. So we ntroduced a crtcal rate n the method whch defnes, randomly and accordng to ths rate, the average number of crtcal messages nsde the whole generated flowset. As we based our approach on crtcal perods change, each frame was defned ether wth just one LO-crtcal perod, or wth one LO and one HI crtcal perod. B. Impact of the load We dd generate dfferent flowsets, each one representng a dfferent schedulng scenaro. We computed the utlzaton of each flow wth a unform dstrbuton based on the network load, and fnally deduce the WCTT of the flow. As we are workng wth ethernet(ieee 802.3), the sze of each frame n constraned n sze between 64 and 1518 bytes. Wth a 100 Mb/s bandwth, all the WCTT n our network are bound between 4.9μs and 115,8μs. We made a scenaro wth a set of 50 dfferent flows (correspondng to a classcal context of use). In our evaluaton, we fxed a global smulaton tme of t = 500μs, whch s enough to observe and bound the dfferent delays we want to focus on. We made the load represented by the flows ncreasng from 0.4 to 1. The computed load was the one n the central node (S 4 ) whch receves all the flows. Wth

6 these generated flows, we wanted to evaluate the mpact of the network load on the crtcalty swtch transmsson delay. Assgnng the hghest prorty for mxed crtcalty management messages (dedcated for confguraton messages) allows the crtcalty-swtch messages to not be delayed more than once by messages wth a lower prorty (non-preemptve effect). We verfed ths hypothese by evaluatng the MC delay swtch as a functon of the network load. Thus, we need to add to ths delay the one due to PTP synchronzaton messages, consdered wth a hgher prorty (see IV). We can observe n fgure 3 that, bascally, the delay of the crtcalty-swtch as a functon of the network load s lnear. the network. As ths swtch delay s bound, the transmsson of HI-crtcal messages can be assured n our topology n a bound and known tme. Fg. 3. Swtch-crtcalty delay/load We now analyse the mpact of the non-preemptve delay for flows sent wth a smaller prorty than those of MC messages. C. Non-preemptve effect We pcked the same parameters as for the prevous scenaro, but we also decded to put a lmt on the hghest WCTT n the network than for MC messages. We smulated 4 dfferent cases wth dfferent lmts n the hghest WCTT. The results obtaned (see fgure 4) shows that the crtcaltyswtch delay s strongly nfluenced by the hghest WCTT n the network. To evaluate ths nfluence, we lmted the hghest WCTT to small szes: 20 μs (262B), 30 μs (393B) and 40 μs (524B). We can observe that, at the hghest loads, the transmsson tme s nearly constant. The pont s: n the UUnfast task generaton method we presented n V-A, the task model bases the WCTT computaton on the load. It means that the hghest WCTT ncreases wth the load. That explans why, n the frst scenaro wthout any partcular lmt, we obtaned an ncreasng transmsson tme wth the ncreasng load. On the contrary, when lmtng the hghest WCTT n the network, we obtan an constant swtchng delay (mpacted from 1% to 6% n our examples at the hghest network load, due to error margns we tolerated n the load computaton). As explaned n IV, attrbutng to the swtch-crtcalty and relable multcast messages the hghest prorty allows us to make the crtcalty messages ndependant from the network load (the mpact s very lmted). No matter the traffc n the network, we can then compute the crtcalty swtch delay n Fg. 4. Swtch-crtcalty delay/load wth lmted hghest WCTT VI. CONCLUSION A. Concluson In ths paper, we present a crtcalty-change protocol n a clock synchronzed Swtched Ethernet network, n the case of two crtcalty levels. Ths crtcalty change protocol s based on a relable real-tme multcast used update the crtcalty nformaton n all the nodes of the network whle preservng ts consstency. The real-tme multcast bulds a total order for the updates of the crtalty nformaton. It reles on the based IEEE1588v2 global clock synchronzaton. We characterze the worst case delay for a crtcalty change wth the trajectory approach. Through smulaton, we generate random scenaros to test the tme needed for a crtcalty change. We show that the crtcalty swtch delay s harldy mpacted by the network load. As a further work, we wll show how to characterze the end to end response tme of HI messages n the case of a LO to HI crtcalty swtch. REFERENCES [1] S. Baruah, A. Burns, and R. Davs, Response-tme analyss for mxedcrtcalty systems, n RTSS [2] S. Baruah, Mxed crtcalty schedulablty analyss s hghly ntractable, [3] A. Burns and R. Davs, Mxed crtcalty systems: A revew. Department of Computer Scence, Unversty of York, 2013, vol. Tech. Rep. [4] O. Cros, F. Fauberteau, L. George, and X. L, Smulatng real-tme and embedded networks schedulng scenaros wth artems, n WATERS, [5] H.-T. Lm, L. Völker, and D. Herrscher, Challenges n a future p/ethernet-based n-car network for real-tme applcatons, n Proceedngs of the 48th Desgn Automaton Conference, ser. DAC 11. New York, NY, USA: ACM, 2011, pp [Onlne]. Avalable: [6] L. George and P. Mnet, A ffo worst case analyss for a hard real-tme dstrbuted problem wth consstency constrants, n Proceedngs of the 17th Internatonal Conference on Dstrbuted Computng Systems, [7] S. Martn, P. Mnet, and L. George, End-to-end response tme wth fxed prorty schedulng: trajectory approach versus holstc approach, n Internatonal Journal of Communcaton Systems, vol. 18, no. 1. John Wley & Sons, Ltd., 2005, pp [8] X. L, O. Cros, and L. George, The trajectory approach for afdx ffo networks revsted and corrected, n RTCSA 14, [9] E. Bn and G. C. Buttazzo, Measurng the performance of schedulablty tests, Journal of Real-Tme Systems, pp , 2005.