I Introduction and motivation
Designing and analysing a distributed real-time system is a very challenging task. The main source of complexity arises from the large number of parameters to consider: task priority, computation times and deadlines, synchronisation, precedence and communication constraints, etc. Finding the “optimal” values for the parameters is not easy, and often the robustness of the solution strongly depends on the exact values: a small change in one parameter may completely change the behaviour of the system and even compromise the correctness. For these reasons, designers are looking for analysis methodologies that enable incremental design and exploration of the space of parameters.
Task computation times are particularly important parameters. In modern processor architectures, it is very difficult to precisely compute worst-case computation times of tasks, and estimations derived by previous executions are often used in the analysis. However, estimations may turn out to be optimistic, hence an error in the estimation of a worst-case execution time may compromise the schedulability of the system.
The goal of this research is to characterise the space of the parameters of a real-time system for which the system is schedulable, i.e. all tasks meet their deadlines. Parametric analyses for real-time systems have been proposed in the past, especially on single processors [1, 2, 3, 4].
In this paper, we investigate the problem of doing parametric analysis of real-time distributed systems scheduled by fixed priority. We consider an application modelled by a set of pipelines of tasks (also called transactions in ), where each pipeline is a sequence of tasks that can be periodic or sporadic, and all tasks in a pipeline must complete before an end-to-end deadline. We consider that all nodes in the distributed system are connected by one or more CAN bus .
Ii Related Work
There has been a lot of research work on parametric schedulability analysis, especially on single processor systems. Bini and Buttazzo  proposed an analysis of fixed priority single processor systems based on Lehoczky test . Later, Bini, Di Natale and Buttazzo  proposed a more complex analysis, which considers also the task periods as parameters. Such results are summarised and extended in Bini’s PhD thesis .
Parameter sensitivity can be also be carried out by repeatedly applying classical schedulability tests, like the holistic analysis [13, 5]. One example of this approach is used in the MAST tool [9, 10], in which it is possible to compute the slack (i.e. the percentage of variation) with respect to one parameter for single processor and for distributed systems by applying binary search in that parameter space .
A similar approach is followed by the SymTA/S tool , which is based on the event-stream model . Another interesting approach is the Modular Performance Analysis (MPA)  which is based on Real-Time Calculus . In both cases, the analysis is compositional, therefore less complex than the holistic analysis; nevertheless, these approaches are not fully parametric, in the sense that it is necessary to repeat the analysis for every combination of parameters values in order to obtain the schedulability region.
Model checking on Parametric Timed Automata (PTA) can be used for parametric schedulability analysis, as proposed by Cimatti, Palopoli and Ramadian . In particular, thanks to generality of the PTA modelling language, it is possible to model a larger class of constraints, and perform parametric analysis on many different variables, for example task offsets. Their approach has been recently extended to distributed real-time systems .
Also based on PTA is the approach proposed by André et al. . Their work is based on the Inverse Method  and it is very general because it permits to perform analysis on any system parameter. However, this generality can be paid in terms of complexity.
In this paper, we first propose an extensions of the methods in  for distributed real-time systems. We also propose a model of a distributed real-time systems in PTA, and compare the two approaches against classical holistic analysis.
Iii System model
We consider distributed real-time systems consisting of several computational nodes, each one hosting one single processor, connected by one or more shared networks. We consider preemptive fixed priority scheduling for processors, as this is the most popular scheduling algorithm used in industry today, and non-preemptive fixed priority scheduling for networks. In particular, the CAN bus protocol is a very popular network protocol that can be analysed using non-preemptive fixed priority scheduling analysis . We will consider extensions to our methodology to other scheduling algorithms and protocols in future works.
For the sake of simplicity and uniformity of notation, in this paper we use the same terminology to denote processors and communication networks, and tasks and messages. Therefore, without loss of generality, from now on we will use the term task to denote both tasks and messages, and the term processor to denote both processors and networks
A distributed real-time system consists of a set of task pipelines to be executed on a set of processors . In order to simplify the notation, in the following we sometime drop the pipeline index when there is no possibility of misinterpretation.
A pipeline is a chain of tasks , and each task is allocated on one possibly different processor. A pipeline is assigned two fixed parameters: is the pipeline period, and is the end-to-end deadline. This means that the first task of the pipeline is activated every units of time, and every activation is an instance (or job) of the task. We denote the -th instance of task as . Every successive task in the pipeline is activated when the corresponding instance of the previous task has completed; finally, the last task must complete before units of time from the activation of the first task. Therefore, tasks must be executed in a sequence: job cannot start executing before job has completed.
A task can be a piece of code to be executed on a CPU, or a message to be sent on a network. More precisely, a real-time periodic task is modelled by the following fixed parameters:
is the task period. All tasks in the same pipeline have period equal to the pipeline period ;
is the task priority; the higher , the larger the priority;
is the task fixed deadline; all jobs of must complete within from their activation.
Also, a task has the following free parameters:
is the worst-case computation time (or worst-case transmission time, in case it models a message). In this paper we want to characterise the schedulability of the system in the space of the computation times, so is a free parameter.
is a variable denoting an upper bound on the task worst-case completion time. We will call this variable actual task deadline or simply task deadline. Of course, we require that . Remember that fixed priority does not use the task deadline for scheduling, but just for schedulability analysis. As we will see later, we will use this variable for imposing precedence constraints on pipelines. We say that a task has constrained deadline when , and unconstrained deadline when .
is the task start time jitter (see below).
As anticipated, a task consists of an infinite number of jobs . Each job is activated at time , can start executing (or can be sent on the network) no earlier than time , with , executes (or is transmitted over the network) for units of time, and completes (or is received) at . For the task to be schedulable, it must be . A sporadic task has the same parameters as a periodic task, but parameter denotes the minimum inter-arrival time between two consecutive instances. We also define the -th level hyperperiod as .
In this paper, we use the following convention. All tasks belonging to a pipeline are activated at the same time . However, only the first task can start executing immediately: . The following tasks can only start executing when the previous task has completed: . The task jitter is the worst case start time of a task: .
A scheduling algorithm is fully preemptive if the execution of a lower priority job can be suspended at any instant by the arrival of a higher priority job, which is then executed in its place. A scheduling algorithm is non-preemptive if a lower priority job can complete its execution regardless of the arrival of higher priority jobs. In this paper, we consider preemptive fixed priority scheduling for CPUs, and non-preemptive fixed priority scheduling for networks.
Iv Analytic method
In this section we describe a novel method for parametric analysis of distributed system. The method is based on the sensitivity analysis by Bini et al. [2, 1], and extends it to include jitter and deadline parameters.
Iv-a Single processor preemptive fixed priority scheduling
There are many ways to test the schedulability of a set of real-time periodic tasks scheduled by fixed priority on a single processor. In the following, we will use the test proposed by Seto et al.  because it is amenable to parametric analysis of computation times, jitters and deadlines.
With respect to the original formulation, we now consider tasks with constrained deadlines (i.e. can be less than or equal to ).
Consider a system of sporadic tasks with constrained deadlines and zero jitter, executed on a single processor by a fixed priority scheduler. Assume all tasks are ordered in decreasing order of priorities, with being the highest priority task.
Task is schedulable if and only if:
where is the set of all possible vectors of
is the set of all possible vectors ofpositive integers.
Notice that, with respect to the original formulation, we have separated the case of from the rest of the inequalities.
The theorem allows us to only consider sets of linear inequalities, because the non-linearity has been encoded in the variables . The resulting system is a set of inequalities in disjunctive and conjunctive form. Geometrically, this corresponds to a non-convex polyhedron in the space of the variables .
How many vectors do we have to consider? If the deadline is known, the answer is to simply consider all vectors corresponding to the minimal set of scheduling points by Bini and Buttazzo . If is unknown, we have to consider many more vectors: more specifically, we must select all multiples of the period of any task with priority higher than , until the maximum possible value of the deadline. All vectors until time can be computed as:
If a task is part of a pipeline with end-to-end deadline equal to , then (keep in mind that, by now, the deadline is supposed to not exceed the task period). Therefore, we have to check all .
The number of vectors (and correspondingly, the number of inequalities) depends on the relationship between the task periods. In real applications, we expect the periods to have “nice” relationships: for example, in many cases engineers choose periods that are multiples of each others. Therefore, we expect the set of inequalities to have manageable size for realistic problems.
We have one such non-convex region for every task . Since we have to check the schedulability of all tasks on a CPU, we must intersect all such regions to obtain the final region of schedulable parameters.
Iv-B Unconstrained deadlines and jitters
We now extend Seto’s test to unconstrained deadlines and variable jitters. When considering a task with deadline greater than period, the worst-case response time may be found in any instance, not necessarily in the first one (as with the classical model of constrained deadline tasks). Therefore, we have to check the workload not only of the first job, but also of the following jobs of . Let use define , i.e. the number of jobs of contained in the -level hyperperiod. Then, task is schedulable if and only if the following system of inequalities is verified:
The correctness of the test is proved by the following Lemma.
Consider a system . Let be a task set obtained from by substituting with having computation time , deadline and the same priority .
If for every , task completes before its deadline, then the first jobs of will also complete before their deadlines.
By induction. Base of induction: the response time of job corresponds to the response time of the first job of (trivially true). Therefore, if is schedulable, also the first job of is schedulable.
Now, the induction step. Suppose the Lemma is valid for , we are now going to prove that is also valid for . By assumption, the first job of is schedulable for . As a consequence of the validity of the Lemma, also the first instances of are schedulable. Let be the finishing time of the first job of . We have two cases: either , or .
In the first case, job is only subject to the interference of higher priority tasks. Therefore, its worst case response time correspond to the situation in which all higher priority tasks arrive at the same time (critical instant), and it is therefore equal to the response time of the first job , hence also schedulable. We can conclude that the Lemma is true without further induction steps.
In the second case, the job has to wait for the previous job to finish before it can start executing. In particular, there is no idle time in interval . Therefore, the response time of job coincides with the response time of task , and if the second one is schedulable, also job is schedulable.
To take into account the task jitter, we can appropriately adjust the last term that accounts for the task deadline, and the set .
Task is schedulable if:
We report here a sketch of the complete proof. For every higher priority interfering task , the worst case situation is when the first instance arrives at , whereas the following instances arrive as soon as it is possible. Therefore, the scheduling points must be modified from to . For what concerns task , the critical instant corresponds to the situation in which the first instance can only start at , hence the available interval is . ∎
Notice that the introduction of unconstrained deadline adds a great amount of complexity to the problem. In particular, the number of non-convex regions to intersect is now , which is dominated by . So, the proposed problem representation does not scale with increasing hyperperiods; however, as we will show in Section VI, the problem is tractable when periods are harmonic or quasi-harmonic, as it often happens in real applications.
Iv-C Non preemptive scheduling
In this paper we model the network as a non-preemptive fixed priority scheduled resource. In non-preemptive fixed priority scheduling, the worst-case response time for a task can be found in its longest -level active period . A -level active period is an interval such that the amount of processing that needs to be performed due to jobs with priority higher than or equal to (including itself) is larger than 0 , and equal to 0 at instants and . The longest can be found by computing the lowest fixed point of the following recursive function :
In order to find the worst-case response time of task , all jobs that appear in the longest need to be checked, with .
To obtain the worst-case response time, we compute first its worst-case start time. When there is no jitter, George et al.  give the following formula to compute the worst-case start time of a job :
Note that is the computation time of the preceding jobs. Since a lower priority task’s execution cannot be preempted, this could “push” one job of a higher priority task to interfere with its future jobs.
Observe that the iterating computation of in Equation (6) is non decreasing and (when the system utilisation is no larger than 1) , so the length of will not exceed .
In this paper, the worst-case execution time of the tasks are considered free parameters. However, can still be upper bounded by . Now, we can derive a similar feasibility test for non preemptive scheduling as in Theorem 2.
A non preemptive task is schedulable if :
where is defined as in Theorem 2, and is the blocking time that task suffers from lower priority tasks:
Term is an additional free variable used to model the blocking time that a task suffers from lower priority tasks. It is possible to avoid the introduction of this additional variable by substituting it in the inequalities with a simple Fourier-Motzkin elimination.
Like in the preemptive case, for every non preemptive task, this theorem builds a set of inequalities. The system schedulability region is the intersection of all the sets. The complexity of this procedure is the same as for the preemptive case.
Iv-D Distributed systems
Until now, we have considered the parametric analysis of independent tasks on single processor systems, with computation times, deadlines and jitter as free parameters. In particular, the equations in Theorem 2 and Theorem 3 give us a way to express the constraints on the system in a fully parametric way: all solutions to the system of Inequalities (4) and (8) are all the combinations of computations times, deadlines and jitters that make the single processor system schedulable.
It is important to make one key observation. If we fix the computation times and the jitters of all tasks, and we leave the deadlines as the only free variables, the worst-case response time of each task can be found by minimising the deadline variables. As an example, consider the following task set (the same as in ) to be scheduled by preemptive fixed priority scheduling on a single processor:
We consider as a parameter and set up the system of inequalities according to Equation (4). After reduction of the non-useful constraints, we obtain
Notice that is actually the worst-case response time of .
The second key observation is that a precedence constraint between two consecutive tasks and in the same pipeline can be expressed as . This basically means that the worst-case response time of task should never exceed the jitter (i.e. worst-case start time) of task . Therefore, we have a way to relate tasks allocated on different processors that belong to the same pipeline.
Finally, the last task in every pipeline, let us call it must complete before the end-to-end deadline: .
We are now ready use inequalities in (4) as building blocks for the parametric analysis of distributed systems. The procedure to build the final system of inequalities is as follows:
For each processor, we build the system of inequalities (4), and for every network the system of inequalities in (8). All these systems are independent of each other, because they are constraints on different tasks, so they use different variables. The combined system contains variables, where is the total number of tasks.
For every pipeline, we add the following precedence constraints:
For the first task in the pipeline, let us denote it as , we set its jitter to 0: .
For every pair of consecutive tasks, let us denote them as and , we impose the precedence constraint: ;
For the last task in the pipeline, let us denote it as , we impose that it must complete before its end-to-end deadline .
Such constraints must intersect the combined system to produce the final system of constraints.
To give readers an idea how the parameter space of a distributed system would look like, here is a very simple example, built with the goal of showing the general methodology without taking too much space. We consider a system with two processors (and no network), two tasks and , and one pipeline consisting of two tasks, and .
To make sure that for each task we have one single inequality (see Equation (4)) we set up the deadlines short enough so that one schedulability point for each task needs to be considered, thus avoiding complex disjoints.
Based on the analysis in this section, we derive a set of constraints, where , and are the free variables for the tasks.
In the first two lines, we show the “trivial” inequalities: all values must be non-negative, and every deadline must not exceed the corresponding maximum deadline specified in the table. The inequalities at line 3 and 4 and the inequalities at line 5, 6 and 7 are (reduced) constraints (according to Theorem 2) on the schedulability of tasks on processor 1 and 2, respectively. Finally, the inequalities in the last line are the ones imposed by the precedence constraints between and .
A real system will produce a much more complex set of constraints. For each task we will need to prepare a set of disjoint inequalities, that must be intersect with each other: this may greatly increment the number of inequalities to be considered. Also, often we need to model the network. Therefore, we prepared a software tool to automatically build and analyse the set of inequalities for a distributed system.
The analytic method proposed in this section has been implemented in RTSCAN , a C/C++ library publicly available as open source code that collects different types of schedulability tests. The code for the parametric schedulability analysis uses the PPL (Parma Polyhedra Library) , a library specifically designed and optimised to represent and operate on polyhedra. The library efficiently operates on rational numbers with arbitrary precision: therefore, in this work we make the assumption that all variables (computations times, deadlines and jitter) are defined in the domain of integers. This does not represent a great problem, since in practice every value is multiple of a real-time clock expressed as number of ticks.
An evaluation of this tool, and of the complexity of the analysis presented here, will be presented in Section VI.
V The Inverse Method approach
V-a Parametric Timed Automata
Timed Automata are finite-state automata augmented with clocks, i.e., real-valued variables increasing uniformly, that are compared within guards and invariants with timing delays . Parametric timed automata (PTAs)  extend timed automata with parameters, i.e., unknown constants, that can be used in guards and invariants.
Formally, given a set of clocks and a set of parameters, a constraint over and is a conjunction of linear inequalities on and . Given a parameter valuation (or point) , we write when the constraint where all parameters within have been replaced by their value as in is satisfied by a non-empty set of clock valuations.
A PTA is with a finite set of actions, a finite set of locations, the initial location, a set of clocks, a set of parameters, a constraint over , the invariant assigning to every a constraint over and , and a step relation consisting of elements , where , , is the set of clocks to be reset, and the guard is a constraint over and .
The semantics of a PTA is defined in terms of states, i.e., couples where and is a constraint over and . Given a point , we say that a state is -compatible if . Runs are alternating sequences of states and actions, and traces are time-abstract runs, i.e., alternating sequences of locations and actions. The trace set of corresponds to the traces associated with all the runs of . Given and , we denote by the (non-parametric) timed automaton where each occurrence of a parameter has been replaced by its constant value as in . One defines as the set of states reachable from a set of states in exactly steps under , and .
Detailed definitions on parametric timed automata can be found in, e.g., .
The Inverse Method exploits the model of Timed Automata and the knowledge of a reference point of timing values for which the good behaviour of the system is known. The method synthesises automatically a dense zone of points around the reference point, for which the discrete behaviour of the system, that is the set of all the admissible sequences of interleaving events, is guaranteed to be the same. Although the principle of the inverse method shares similarities with sensitivity analysis, its algorithm proceeds by iterative state space exploration. Furthermore, its result comes under the form of a fully parametric constraint, in contrast to sensitivity analysis. By repeatedly applying the method, we are able to decompose the parameter space into a covering set of “tiles”, which ensure a uniform behaviour of the system: it is sufficient to test only one point of the tile in order to know whether or not the system behaves correctly on the whole tile.
V-B System model with PTAs
In this section, we show how we modelled a schedulability problem as defined in III, similarly to what has been done in . In the current implementation, we only model pipelines with end-to-end deadlines no larger than their periods. Moreover, all pipelines are strictly periodic, and have 0 offset. This means that the results of the parametric analysis produced by this model are only valid for periodic synchronous pipelines.
We illustrate our model with the help of an example of two pipelines with , , , , being a preemptive processor and being non-preemptive. We have that and .
In Figure 1, we show the model of a pipeline. A pipeline is a sequence of tasks that are to be executed in order: when a task completes its instance, it instantly activates the next one in the pipeline. Once every task in the pipeline has completed, the pipeline waits for the next period to start.
In Figure 2, we present how we model a preemptive processor. The processor can be idle, waiting for a task activations. As soon as a request has been received, it moves to one of the states where the corresponding higher priority task is running. If it receives another activation request, it moves to the state corresponding to the highest priority task running. Moreover, while a task executes, the scheduler automaton checks if the corresponding pipeline misses its deadline. In the case of a deadline miss, the processor moves to a special failure state and stops any further computation.
In Figure 3, we present the model a non-preemptive processor. Similarly to the previous case, the processor can be idle, waiting for an activation request. As soon as a request as been received it moves to a corresponding state, setting a token corresponding to the activated task to . If another request is sent at the same time, it sets a corresponding token to , and moves to the state where the highest priority task will be running. Once a task is completed, the processor set the corresponding token to , and according to the token set to , moves to the state where the highest priority task will be running. Similarly to the previous case, while a task executes, the automaton checks for deadline misses, and in that case it stops any further computation by moving to a special failure state.
Since we model periodic pipelines, and the model explores all possible traces, we expect that schedulability region will be larger that the one obtained with other techniques which only consider sporadic pipelines (like the analysis proposed in Section IV). An assessment of this difference is provided in the next section.
We evaluated the effectiveness and the running time of three different tools for parametric schedulability analysis: the RTSCAN tool, which implements the analytic method described in Section IV; the IMITATOR tool , described in Section V; and the MAST tool .
We highlight that the three tools are implemented in different languages, and use different ways to optimise the analysis; RTSCAN is implemented in C/C++ and uses on the PPL library; IMITATOR is implemented in OCaml, but it also used the PPL libraries for building regions; finally, MAST is implemented in Ada.
For MAST, we have selected the “Offset Based analysis”, proposed in . For IMITATOR, we consider all pipelines as strictly periodic, and only deadlines less than periods. We evaluated the tools on two different test cases, in increasing order of complexity. We will first present the results, in terms of schedulability regions, for two different test cases. In order to simplify the visualisation of the results, for each test case, we will present the 2D region of two parameters only: however, all three methods are general and can be applied to any number of parameters.
In Section VI-C, after discussing some important implementation details, we will and present the execution times of the three tools.
Vi-a Test case 1
The first test case has been adapted from  (we reduced the computation times of some tasks to position the system in a interesting schedulability region). It consists of 2 processors, connected by a CAN bus, three simple periodic tasks and one pipeline. The parameters are listed in Table I. Processor 1 and 3 model two different computational nodes that are scheduled by preemptive fixed priority, and Processor 2 models a CAN bus with non-preemptive fixed priority policy. The only pipeline models a remote procedure call from CPU 1 to CPU 3. All tasks have deadlines equal to periods, and also the pipeline has end-to-end deadline equal to its period. Only two messages are sent on the network, and if the pipeline is schedulable, they cannot interfere with each other. We wish to perform parametric schedulability analysis with respect to and .
The resulting regions of schedulability from the tools RTSCAN and MAST are reported in Figure 4, whereas the region produced by IMITATOR is reported in Figure 5 in green (the non schedulable region is also painted in red). The small white triangles are regions which do not contain any integer point, and have not been explored by the IMITATOR tool.
In this particular test, RTSCAN dominates MAST. After some debugging, we discovered that the analysis algorithm currently implemented in MAST does not consider the fact that the two messages and
cannot interfere with each other, and instead considers a non-null blocking time on the network. This is probably a small bug in the MAST implementation that we hope will be solved in a future version.
Also, as expected, the region computed by IMITATOR dominates the other two tools includes the other two regions. The reason is in the different model of computation: IMITATOR considers fully periodic and synchronous pipelines, therefore it produces all possible traces that can be generated in this case. Both RTSCAN and MAST, instead, compute upper bounds on the interference that a task can suffer from higher priority task and from blocking time of lower priority tasks. Therefore, they can only provide a sufficient analysis.
Vi-B Test case 2
The second test case is taken from . It consists of two pipelines on 3 processors (with id 1, 3 and 4) and one network (with id 2). We actually consider two versions of this test case: in the first version (a) pipeline is periodic with period and end-to-end deadline equal to the period. In the second version (b), the period of the first pipeline is reduced to (as in the original specification in ). The full set of parameters is reported in Table II, where all values are expressed in microseconds. We perform parametric analysis on and .
For version (a) we run all tools and we report the regions of schedulability in Figure 6 for RTSCAN and MAST. In this case, MAST dominated RTSCAN. The reason is due to the offset based analysis methodology used in MAST, which reduces the interference on one task from other tasks belonging to the same pipeline. RTSCAN does not implement such an optimisation (it will be the topic of future extensions) and hence it is more pessimistic.
The results from IMITATOR are shown in Figure 7. Again, the region produced by IMITATOR dominates the one produced by the other two tools.
For version (b) we run only RTSCAN and MAST, because in the current version of IMITATOR we can only model constrained deadline systems. The results for version (b) are reported in Figure 8. In this case, MAST dominates RTSCAN. Again, this is probably due to the fact that MAST implements the offset-based analysis.
Vi-C Execution times
Before looking at the execution times of the three tools in the three different test cases, it is worth to discuss some detail about their implementation. First of all, all the three tools are single threaded, therefore we did not use any kind of parallelisation technique in order to have a fair comparison. The RTSCAN tool uses the technique described in Section IV, and as a result it produces a disjunction of convex regions, each one corresponds to a set of inequalities in AND. Typically, the number of convex regions produced by the tool is relatively small. Also, there is no need to “explore” the space of feasible points, as these regions are naturally obtained from the problem constraints.
IMITATOR also produces a disjunction of convex regions. However, these regions are typically smaller and disjoints. Moreover, to produce a region, IMITATOR needs to start from a candidate point on which to perform a sensitivity analysis. More specifically, it works as follows: 1) it starts from a random point (typically the centre of the interval) and computes a region around it. Then, it searches for the next candidate point outside of the already found regions. The key factor here is how this search is performed. Currently, IMITATOR search for a candidate point in the neighbourhood of the current region. This is a very general strategy that works for any kind of PTA. However, the particular structure of schedulability problems would probably require an ad-hoc exploration algorithm.
MAST can perform schedulability analysis given a full set of parameter values, and it returns either a positive or a negative response. In addition, MAST can perform sensitivity analysis on one parameter (called slack computation in the tool), using binary search on a possible interval of values. This latter strategy can be used to implement parametric analysis: we select a interval of values for each free parameter that we wish to analyse. Then, we perform a cycle on all values of one parameter (with a predefined step) and we ask MAST to compute the interval of feasible values for the other parameter.
This may not be the smartest way to proceed: it is possible, for example, to implement binary search on the full 2D space of free parameters to accelerate the execution time of the tool. We defer the implementation of such an algorithm as a future extension.
All experiments have been performed onto a PC with 8Gb of RAM, an Intel Core I7 quad-core processor, working at 800 Mhz per processor.
We are now ready to present and discuss the execution times of the tools in the three test cases, which are reported in Table III, together with the length of the hyperperiod for the test cases. As you can see, for small problems RTSCAN performs very well. In test case 2b, the execution time of RTSCAN is much larger than the one obtained from test case 2a. This is due to the fact that in test case 2b, one pipeline has end-to-end deadline greater than the period, and therefore RTSCAN needs to compute many more inequalities (for all points in the hyperperiod).
As for MAST, in test cases 2a and 2b (where the time units are expressed in microseconds, and therefore are quite large), the search has been run with a step of 100 for a good compromise between precision and execution time. However, we believe that a smarter algorithm for exploring the space of parameters can really improve the overall execution time.
Finally, IMITATOR greatly suffers from a similar problem. We observed that the tool spends a few seconds for computing the schedulability region around each point. However, the regions are quite small, and there are many of them: for example, in test case 2a IMITATOR analysed 257 regions. Also, it spends a large amount of time in searching for neighbourhood points. Therefore, we believe that a huge improvement in the computation time of IMITATOR can be achieved by coming up with a smarter way of exploring the schedulability space.
|2b||300,000||1 m 47s||33m19s||–|
Vii Conclusions and Future Works
In this paper we presented two different approaches to perform parametric analysis of distributed real-time systems: one based on analytic methods of classic schedulability analysis; the other one based on model checking of PTA. We compared the two approached with classical holistic analysis.
The results are promising, and we plan to extend this work along different directions. Regarding the analytic method, we want to enhance the analysis including static and dynamic offsets, following the approach of . Also, in the future we will to extend the model to consider mutually exclusive semaphores and multiprocessor scheduling.
Regarding IMITATOR, we plan to improve the algorithm to explore the space of parameters: one promising idea is to use the analytic method to find an initial approximation of the feasible space, and then extend the border of the space using PTAs. We also plan to collaborate with the team at Universidad de Cantabria that develops the MAST tool on novel algorithms for exploring an N-dimensions parameter space.
We would like to express our gratitude to Michael González Harbour and Juan M. Rivas, from the Universidad de Cantabria, for their support to installing and using the MAST tool.
The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement No. 246556.
-  E. Bini, “The design domain of real-time systems,” Ph.D. dissertation, Scuola Superiore Sant’Anna, 2004.
-  E. Bini, M. Di Natale, and G. C. Buttazzo, “Sensitivity analysis for fixed-priority real-time systems,” Real-Time Systems, Apr. 2007, dOI: 10.1007/s11241-006-9010-1.
-  A. Cimatti, L. Palopoli, and Y. Ramadian, “Symbolic computation of schedulability regions using parametric timed automata,” 30 2008-Dec. 3 2008, pp. 80–89.
-  É. André, L. Fribourg, U. Kühne, and R. Soulat, “IMITATOR 2.5: A tool for analyzing robustness in scheduling problems,” in Proceedings of the 18th International Symposium on Formal Methods (FM’12), ser. Lecture Notes in Computer Science, D. Giannakopoulou and D. Méry, Eds., vol. 7436. Paris, France: Springer, Aug. 2012, pp. 33–36. [Online]. Available: http://www.lsv.ens-cachan.fr/Publis/PAPERS/PDF/AFKS-fm12.pdf
-  K. Tindell and J. Clark, “Holistic schedulability analysis for distributed hard real-time systems,” Microprocess. Microprogram., vol. 40, no. 2-3, pp. 117–134, Apr. 1994. [Online]. Available: http://dx.doi.org/10.1016/0165-6074(94)90080-9
-  R. I. Davis, A. Burns, R. J. Bril, and J. J. Lukkien, “Controller area network (can) schedulability analysis: Refuted, revisited and revised,” Real-Time Systems, vol. 35, pp. 239–272, 2007. [Online]. Available: http://dx.doi.org/10.1007/s11241-007-9012-7
-  É. André and R. Soulat, The Inverse Method. ISTE Ltd and John Wiley & Sons Inc., 2013, 176 pages.
-  L. Fribourg, D. Lesens, P. Moro, and R. Soulat, “Robustness analysis for scheduling problems using the inverse method,” in Proceedings of the 19th International Symposium on Temporal Representation and Reasoning (TIME’12), M. Reynolds, P. Terenziani, and B. Moszkowski, Eds. Leicester, UK: IEEE Computer Society Press, Sept. 2012, pp. 73–80. [Online]. Available: http://www.lsv.ens-cachan.fr/Publis/PAPERS/PDF/FLMS-time12.pdf
-  M. González Harbour, J. Gutiérrez, J. Palencia, and J. Drake, “Mast: Modeling and analysis suite for real-time applications,” in Proceedings of the Euromicro Conference on Real-Time Systems, Delft, The Netherlands, June 2001.
-  U. d. C. Grupo de Computadores y Tiempo Real, “Mast: Modeling and analysis suite for real-time applications,” http://mast.unican.es/.
-  E. Bini and G. C. Buttazzo, “Schedulability analysis of periodic fixed priority systems,” IEEE Transactions on Computers, vol. 53, no. 11, pp. 1462–1473, Nov. 2004.
-  J. Lehoczky, L. Sha, and Y. Ding, “The Rate Monotonic Scheduling Algorithm: Exact Characterization and Average Case Behavior,” in Proc. Real Time Systems Symposium, 1989, pp. 166–171.
-  J. C. Palencia and M. Gonzalez Harbour, “Schedulability analysis for tasks with static and dynamic offsets,” in Proc. 19th IEEE Real-Time Systems Symp., 1998, pp. 26–37.
-  R. Henia, A. Hamann, M. Jersak, R. Racu, K. Richter, and R. Ernst, “System level performance analysis - the symta/s approach,” Computers and Digital Techniques, IEE Proceedings -, vol. 152, no. 2, pp. 148 – 166, mar 2005.
-  K. Richter and R. Ernst, “Event model interfaces for heterogeneous system analysis,” in Proceedings of the conference on Design, automation and test in Europe, ser. DATE ’02. Washington, DC, USA: IEEE Computer Society, 2002, pp. 506–. [Online]. Available: http://dl.acm.org/citation.cfm?id=882452.874327
-  E. Wandeler, L. Thiele, M. Verhoef, and P. Lieverse, “System architecture evaluation using modular performance analysis: a case study,” Int. J. Softw. Tools Technol. Transf., vol. 8, no. 6, pp. 649–667, Oct. 2006. [Online]. Available: http://dx.doi.org/10.1007/s10009-006-0019-5
-  L. Thiele, S. Chakraborty, and M. Naedele, “Real-time calculus for scheduling hard real-time systems,” in Circuits and Systems, 2000. Proceedings. ISCAS 2000 Geneva. The 2000 IEEE International Symposium on, vol. 4. IEEE, 2000, pp. 101–104.
-  T. Le, L. Palopoli, R. Passerone, and Y. Ramadian, “Timed-automata based schedulability analysis for distributed firm real-time systems: a case study,” accepted to Interntional Journal on Software Tools for Technology Transfer, –.
-  D. Seto, D. P. Lehoczky, and L. Sha, “Task period selection and schedulability in real-time systems,” in RTSS, Dec 1998.
-  E. Bini and G. C. Buttazzo, “Schedulability analysis of periodic fixed priority systems,” IEEE TC, vol. 53, pp. 1462–1473, 2004.
-  R. J. Bril, J. J. Lukkien, and W. F. J. Verhaegh, “Worst-case response time analysis of real-time tasks under fixed-priority scheduling with deferred preemption revisited,” in Proceedings of the 19th Euromicro Conference on Real-Time Systems, ser. ECRTS ’07. Washington, DC, USA: IEEE Computer Society, 2007, pp. 269–279. [Online]. Available: http://dx.doi.org/10.1109/ECRTS.2007.38
-  L. George, D. D. Voluceau, and B. L. C. C. (france, “Preemptive and non-preemptive real-time uni-processor scheduling,” 1996.
-  G. Lipari and Y. Sun, “Real-time schedulability analyzer (rtscan),” Web page: https://github.com/glipari/rtscan, January 2013.
-  R. Bagnara, P. M. Hill, and E. Zaffanella, “The Parma Polyhedra Library: Toward a complete set of numerical abstractions for the analysis and verification of hardware and software systems,” Dipartimento di Matematica, Università di Parma, Italy, Quaderno 457, 2006, available at http://www.cs.unipr.it/Publications/. Also published as arXiv:cs.MS/0612085, available from http://arxiv.org/.
-  R. Alur and D. L. Dill, “A theory of timed automata,” Theoretical Computer Science, vol. 126, no. 2, pp. 183–235, Apr. 1994.
R. Alur, T. A. Henzinger, and M. Y. Vardi, “Parametric real-time reasoning,”
Proceedings of the twenty-fifth annual ACM symposium on Theory of computing, ser. STOC’93. New York, NY, USA: ACM, 1993, pp. 592–601.