Enhancing Real-Time CAN Communications by the Prioritization of Urgent Messages at the Outgoing Queue

Transcription

1 Enhancing Real-Tie CAN Counications by the Prioritization of Urgent Messages at the Outgoing Queue ANTÓNIO J. PIRES (1), JOÃO P. SOUSA (), FRANCISCO VASQUES (3) 1,,3 Faculdade de Engenharia da Universidade do Porto INESC Porto Rua Dr. Roberto Frias, Porto PORTUGAL {julio.pires, jpsousa, Abstract: - To ensure the correct behaviour of a Networked Control Syste, the counication network ust provide a reliable and tiely counication service. The two coponents with the highest ipact on the counication delays are the Mediu Access Control (MAC) protocol and the local counication stack, therefore, the usage of an adequate counication stack is of utost iportance to guarantee the tiing correctness of a feedback control application. In this paper, we propose the use of state-of-the-art scheduling algoriths to anage the outgoing queue of a local counication stack. We deonstrate that it is possible to iprove the responsiveness of applications supported by the CAN counication protocol, by using just a light scheduling iddleware to adequately schedule the outgoing queue. We also show that ipleenting such iddleware even on top of COTS counication hardware, opens the possibility to enhance the counication process by iniizing the nuber of deadline isses for highly loaded network scenarios. Keywords: - CAN Counication, Fieldbus Systes, Real-Tie Counication, Networked Control Systes, Priority Inversion, Message Scheduling. 1 Introduction Fieldbus networks becae increasingly popular in coputer-controlled systes. Coputer-controlled systes wherein the control loops are closed through a real-tie fieldbus network are called Networked Control Systes (NCS). Figure 1 represents an exaple of a Networked Control Syste. Fieldbus Network Figure 1: Networked Control Syste (NCS) A fieldbus network interconnecting sensors, actuators and controllers in a feedback control syste, ust be able to: a) support periodic essage streas, in order to convey the control-related periodic data between the controller and the set of related sensors / actuators; b) guarantee upper-bounded response ties for the essage transfers, in order to cope with the controlrelated delays; c) guarantee a predictable tiing behavior in the presence of a variable network load due to traffic non related to the control application (such as alars, surveillance video/audio streas, etc.); d) guarantee a predictable tiing behaviour in the presence of a faulty counication behaviour. In addition, a well-known proble when using a fieldbus network is the presence of induced jitter, that is, the variability of the tie interval between consecutive transfers. For instance, in spite of periodically requesting the transfer of a specific sensor value, the actual transfer will not be iediately executed, as essages need to be scheduled for transission in a shared resource (the counication ediu). As a consequence, in soe of the control cycles the sensor essage will be transferred earlier in the cycle period, and in soe other cycles it will be transferred later. The real-tie service provided by the fieldbus network will just guarantee that the sensor essage will always be transferred before its deadline. Controller Area Network (CAN) [1] was originally designed for use within road vehicles, to solve cabling probles arising fro the growing use of icroprocessor-based coponents in vehicles. Due to its interesting characteristics, CAN is also being considered for the autoated anufacturing and distributed process control environents [], to support sall-scale Networked Control Systes.

2 Several studies on how to guarantee the tiing requireents of essages in CAN networks are available (e.g. [3]), thus providing pre-run-tie schedulability conditions for the analysis of the tiing requireents of NCS traffic. More recently, the CAN counication protocol has attracted an increased attention fro the research counity, naely in what concerns the guarantee of real-tie constraints [4], the ipleentation of innovative scheduling approaches, either server-based [5] or cyclic table-based [6], and the provision of reliable counication services [7]. However, when assessing the ost coon COTS counication boards, one of the perceived drawbacks is that the outgoing counication queues are FIFO queues. Therefore, the essage transfer in CAN networks using such counication boards is prone to priority inversions. That is, lower priority essages will be regularly transferred before higher priority essages. As a consequence, essage deadlines will be frequently issed, whereas they could be respected if the adequate scheduling strategies were ipleented. This paper assesses a set of state-of-the-art scheduling algoriths that have been proposed for scheduling essages in CAN networks. Specifically, it assesses how such algoriths can be ipleented to re-order the outgoing counication queue and, as a consequence, to guarantee the real-tie behavior of CAN counications, iniizing or even avoiding the priority inversion proble. The ain goal of this paper is to deonstrate that, using COTS hardware and a light scheduling software layer ipleenting adequate real-tie essage scheduling strategies, it is possible to iprove the responsiveness of real-tie applications supported by the CAN counication protocol. The reainder of this paper is organized as follows. Section describes soe of the ost relevant available research works on response tie analysis of CAN networks. In Section 3 we briefly describe the experiental setup used to ipleent and to assess the proposed iddleware. Then, in Section 4, we present the ain strategies to schedule the local outgoing queues, which are proposed in this paper. These strategies will be assessed in Section 5. Finally, soe conclusions are drawn in Section 6. CAN Real-Tie Counications In [3], the authors addressed in detail the analysis of real-tie counications in CAN, assuing fixed priorities for essage streas. In such case, the worst-case response tie of a queued essage, easured fro the release of the queuing task to the tie the essage is fully transitted, is: R = J + I + C (1) J is the queuing jitter of essage strea S, inherited fro the worst-case response tie R sender() (where sender() denotes the task which queues the essage ). The ter I represents the worst-case queuing delay - longest tie between placing the essage in the priority-ordered outgoing queue - and the start of the essage transission. The deadline onotonic (DM) priority assignent [8] can be directly ipleented in a CAN network, by setting the identifier field of each essage strea to a unique priority, according to the DM rule. Therefore: I = B + j hp( ) I + J j + τ bit Cj () Tj where B is the worst-case blocking factor, which is equal to the longest tie taken to transit a lower priority essage, and is given by: B = ax{ 0, C k } (3) k lp( ) The set lp() is the set of essage streas with lower-priority than essage strea S. τ bit is the tie taken to transit a bit on the bus and hp() is the set of essage streas in the syste with higher-priority than the essage strea S. C is the longest tie taken to transit a essage fro strea S. CAN has a 47 bit overhead per essage, and a stuff width of 5 bits. Only 34 of the 47 bits of overhead are subject to stuffing, so C can be defined as: db C = s τbit (4) where db is the nuber of data bytes in the essage. Alternatives for the fixed priority assignent are the dynaic priority schees, such as the nonpreeptive Earliest Deadline First (EDF). In [], the authors analyze how the EDF scheduling algorith [9] could be used to schedule CAN essages. In this work, the authors propose the use of a ixed traffic scheduler (MTS), which attepts to provide a high utilization of the counication ediu while using the standard 11-bit forat for the identifier field. The goal of the MTS scheduler is to ake the identifier fields of different essage streas to reflect the deadlines of essages. However, considering that each essage ust have a unique identifier field (which is a requireent of CAN), they suggested the division of the identifier field into three

3 sub-fields, as shown in Figure 3, where a) is for the essages that are to be scheduled according to the EDF. b) is for essages to be scheduled according to the DM, and c) is for low-priority essages. For the higher-priority essage, the deadline field is derived fro the deadline of the essage. To deal with the case where two essages have the sae deadline, the one with the highest uniqueness code will win (note that Zuberi and Shin assue a Wired- OR bus, thus being '1' the doinant bit). The uniqueness code also serves to identify the essage for reception purposes. 1 Identifier 11 bits deadline 5 bits a) uniqueness 5 bits Identifier 11 bits 0 1 Identifier 11 bits DM Priority 9 bits b) coponents for CAN counications, both for hardware and software. 3.1 Hw Fraework When selecting a hardware fraework, iportant topics are the icrocontroller faily, the eory size (both progra and data eory), the internal peripherals of the icrocontroller and the available prograing tools (and its cost). After the analysis of all these characteristics, we have set up a fraework with the following ites: 1) One IXXAT Autoation ipc-i30 board in a desktop PC, for the reception/transission of control data in the CAN. ) Several Phytec PhyCORE AduC81 odules for the reote data acquisition or data output nodes. (Figure 5). 0 0 Low Priority 9 bits c) Figure 3: Identifier field of a MTS essage This type of encoding raises two probles: the first is that the reaining slack tie of a essage changes with every clock tick. This requires identifiers of all essages to be continually updated, and also that each local clock ust be synchronized. The second is that in a typical syste, essage streas ay have largely different deadlines, which raises a proble with the length of the identifier field (only 5 bits to encode the deadline). To solve the second proble, the authors divided the tie into regions and encoded deadlines according to which region they fall in. Deadlines are then expressed relatively to a periodically increasing reference called the start of epoch (SOE). Figure 4 illustrates this concept for =. SOE Deadline sub-field region l th end of epoch Figure 4: Deadline field, encoded in only two bits 3 Experiental Setup In order to define an adequate experiental setup, we have ade a careful analysis of the available COTS Figure 5: PHYTEC phycore AduC81 The ain purpose of this fraework was to ipleent and assess a set of state-of-the-art scheduling algoriths that have been proposed for scheduling essages in CAN networks. Such scheduling strategies were ipleented in the PhyCORE odules, and a desktop PC, through the IXXAT Autoation ipc-i30 board, was used to onitor the tiing characteristics of the network traffic. 3. Network Traffic Load A highly loaded network scenario was ipleented, in order to assess the proposed scheduling strategies. Specifically, we are interested in the coparison of the tiing results (nuber of deadline isses) when different essage scheduling approaches are used to handle a highly loaded network scenario. Therefore, the first step is to derive such load scenario. We consider a CAN network transferring 8 essage streas, with priorities set according to the Deadline Monotonic rule [8], that is: the essage strea with the saller deadline (d i ) will be assigned with the highest priority identifier. All the transferred essages have the sae essage length: C i =5,5s and a deadline value equal to its periodicity (d i = T i ).

4 The periodicities of the essage streas were set according to the following rules: a) The essage strea with saller periodicity generates a essage every 30s; b) The periodicity (Ti) of the other essage streas are related by a integer coefficient a, which enables setting up different network loads for related sets of essage streas. Therefore, the overall network load (U) can be evaluated using the following equation: C C C C U = T1 at1 4aT1 6aT (6) 1 Considering that T 1 = 30s and C=5,5s, the a paraeter can be expressed as follows: 11 a = (7) 360U 13 For the experiental setup, the eight above defined essage strea sets were used, with a network load ranging fro 70% up to 95%. Figure 6 illustrates two of those essage strea sets, taken fro an experience with only two reote nodes (odd and even streas). Message T C Strea i (s) i, d i (s) U i (%) F1 5, ,3 F 5, ,3 F3 5,5 60 9, F4 5,5 60 9, F5 5,5 11 4,5 F6 5,5 11 4,5 F7 5, ,0 F8 5, ,0 total: 70,% C i (s) T i, d i (s) U i (%) F1 5, ,3 F 5, ,3 F3 5, ,7 F4 5, ,7 F5 5,5 69 8,0 F6 5,5 69 8,0 F7 5, ,3 F8 5, ,3 total: 94,6% Figure 6: Exaple of essage strea sets Finally, the essage scheduling was observed during a 60s snapshot and the nuber of deadline isses for each scheduling strategy were counted and copared. The total nuber of generated essages during the 60s snapshot, for each essage strea and for each network load, can be observed in Figure 7. Message F F F F F F F F Figure 7: Total nuber of generated essages 4 Scheduling of the Outgoing Counication Queue Our ain goal was to assess a set of state-of-the-art scheduling algoriths, used to schedule essages in CAN networks. Specifically, we were interested in the assessent of the outgoing counication queue and, as a consequence, to guarantee the real-tie behaviour of the CAN counication protocol. Three scheduling strategies were devised to anage the software-ipleented outgoing counication queue. The ipleented setup is briefly described in Figure 8. Message Generation SJA1000 CAN Counication Controller Proposed iddleware Queue Scheduling Mechaniss SJA1000 outgoing buffer CAN Network Figure 8: Inside a reote node 4.1 Message generation Messages are generated at each node by the periodic interrupt of one of the icrocontroller s internal tiers and according to several predefined periodicities. A axiu of four essage streas are generated at each node and stored on the outgoing essage queue according to one of three possible scheduling disciplines. 4. Local FIFO Scheduling This scheduling strategy is ipleented just for coparison purposes. This is the algorith

5 traditionally used in CAN counication boards for scheduling the outgoing essage queue. In order to enable a direct coparison between the results obtained when using different scheduling strategies, we ipleented the FIFO Scheduling algorith in the proposed iddleware using a circular buffer with onotonic input and output pointers. 4.3 Local Priority Scheduling This scheduling strategy anipulates the positions where generated essages are placed. Such anipulation is supported by a tied pointer increented every 5s. Whenever a essage is generated and oved to the circular buffer, it is placed in the (pointer + T i ) position, where T i is the periodicity of essage strea M i expressed in ultiples of 5 illiseconds. This eans that the essage with the earliest deadline in the circular buffer is the one that will be scheduled for transission. If the SJA1000 is unable to iediately transfer the scheduled essage and a newly arrived essage to the circular buffer has an earlier deadline, then those two essages will be swapped. This eans that the previously scheduled essage is oved back to the circular buffer. 4.4 Local Priority Scheduling with Teporal Epochs The third scheduling strategy ipleents the concept of Teporal Epochs proposed by Zuberi and Shin [], that was briefly explained in Section. The proposed schee ipleents the Mixed Traffic Scheduler (MTS) to schedule essages siilarly to the EDF scheduling algorith [9]. Basically, it encodes the deadline field (DF) according to the following set of rules (as detailed in Figure 3a): x tieout x tieout, if < DF = (8) x tieout 55, if together with the ipleentation of the Local Priority Scheduling proposed in the previous subsection. 5 Experiental Evaluation The assessent of the proposed scheduling strategies was done for a highly loaded network, according to the Network Traffic Load scenario defined in subsection Deadline isses The obtained results during a 60s snapshot are illustrated in Figures 9, 10 and 11. Deadline Misses Deadline Misses FIFO queues in each node Figure 9: Results for FIFO Scheduling Local Priority Scheduling Figure 10: Results for Local Priority Scheduling Deadline Misses Local Priority Scheduling and Teporal Epochs Figure 11: Results for Local Priority Scheduling with Teporal Epochs Clearly, the nuber of deadline isses is totally unacceptable for the case of the FIFO Scheduling approach as there are deadline isses for the highest priority streas M1 and M even with network loads as sall as 75%.

6 With the Local Priority Scheduling strategy, the results are significantly different. While for interediate network load scenarios, there were no experienced deadline isses, for higher network load scenarios there was a significant increase in the nuber of deadline isses, but now just for the lowest priority streas M6 and M8. Finally, when ipleenting the Local Priority Scheduling with Teporal Epochs, the obtained results are clearly better than the results for any of the previous scheduling strategies. The nuber of deadline isses both for interediate and for highly loaded network scenarios is significantly saller than in the previous cases. This behaviour forecasts an iproved capability to support real-tie applications. 5. Scheduling code overhead The three ipleented strategies were scrutinized for code overhead and the results of this analysis are illustrated in Figure 1. Msg generation FIFO scheduling EDF scheduling MTS scheduling Other % of CPU tie Figure 1: Code profiler results As it can be observed, the results of the code profiler tool reveal that the iddleware ipleentation represents an overhead saller than 1%, even for the case of the used sall perforance icrocontroller. When coparing the proposed scheduling approaches, it also reveals a very sall code overhead for the second and third strategies over the FIFO strategy. 6 Conclusions In this paper, we propose the use of state-of-the-art scheduling algoriths to anage the outgoing queue of CAN counication stacks, in order to iprove the responsiveness of supported applications. We showed that it is possible to reduce the occurrence of priority inversions in the counication ediu, and therefore it becoes possible to decrease the nuber of deadline isses even for highly loaded network scenarios. We provided early ipleentation results of deadline isses versus network load for two scheduling algoriths and showed their better perforance and sall coputing overhead when copared with the FIFO scheduling ethod, aking the very attractive for ipleentation on low end processors. References: [1] ISO (1993). Road Vehicle - Interchange of Digital Inforation - Controller Area Network (CAN) for High-Speed Counication. ISO. [] Zuberi, K. and Shin, K. (1997). Scheduling essages on Controller Area Network for Real Tie CIM Applications. In IEEE Transactions on Robotics and Autoation, Vol. 13, No., pp [3] Tindell, K., Burns, A. and Wellings, A. (1995). Calculating Controller Area Network (CAN) Message Response Tie. In Control Engineering Practice, Vol. 3, No. 8, pp [4] Cavalieri, S., Meeting Real-Tie Constraints in CAN, IEEE Trans. on Industrial Inforatics, 005, vol. 1, no., pp [5] Nolte, T.; Nolin, M.; Hansson, H.A., Real- Tie Server-Based Counication With CAN, IEEE Trans. on Industrial Inforatics, 005, vol. 1, no. 3, pp [6] L. Aleida, P. Pedreiras, J. Fonseca, The FTT- CAN protocol: why and how, IEEE Transactions on Industrial Electronics, Volue 49, Issue 6, Dec. 00 Page(s): [7] L. Pinho, F. Vasques, Reliable Real-Tie Counication in CAN Networks", IEEE Transactions on Coputers, Volue 5, Issue 1, Dec. 003 Page(s): [8] N. Audsley A. Burns M. Richardson A. Wellings, Hard Real-Tie Scheduling: The Deadline Monotonic Approach. In Proceedings of the 8th IEEE Workshop on Real-Tie Operating Systes and Software, Atlanta, [9] C. Liu, J. Layland, "Scheduling Algoriths for Multiprograing in a Hard-Real-Tie Environent. In Journal of the ACM, Vol 0, No. 1, January 1973, pp [10] K. Etschberger; Controller Area Network: Basics, Protocols, Chips and Applications; IXXAT Autoation, 001.