Abstract
System modelling with a unified modelling language (UML) is an active research area for developing real-time system development. UML is widely used modelling language in software engineering community, to specify the requirement, and analyse the target system successfully. UML can be used to provide multiple views of the system under design with the help of a variety of structural and behavioural diagrams at an early stage. UML-RT (unified modelling language-real time) is a language used to build an unambiguous executable specification of a real-time system based on UML concepts. This paper presents a unified modeling approach for a newly proposed rate monotonic scheduling algorithm-shortest job first (RMA-SJF) for partitioned, semipartitioned and global scheduling strategies in multiprocessor architecture using UML-RT for different system loads. As a technical contribution, effective processor utilization of individual processors and success ratio are analyzed for various scheduling principles and compared with EDF and D_EDF to validate our proposal.
1. Introduction
The integration of object modelling and design methods and real-time scheduling theory is the key to successful use of object technology for real-time software. Surprisingly, many past approaches to integrate the two either restrict the object models or do not allow sophisticated schedulability analysis techniques [1]. Visual notations and model abstractions aid object-oriented designer to understand the problem space at an early stage of development cycle. Nowadays, embedded real-time systems are implemented as hardware and software configurations, where the software components have become key for a successful system [2]. The dominance of software in real-time embedded systems design caused the interest in methodologies with widely accepted notations in the software community, such as the unified modeling language (UML). UML-RT (unified modelling language-real time) is a language used to build an unambiguous executable specification of a real-time system based on UML concepts [3]. The integration of schedulability analysis with the industry standard unified modelling language-real time (UML-RT) allows real-time developers to detect and prevent costly design mistakes at an early stage of development.
Technological trends on high performance computational systems are moving towards execution platforms made up of multiple programmable and dedicated processing elements implemented on a single chip known as multiprocessor system-on-chip (MPSoC). In recent years, model based system level design has gained considerable attention in multiprocessor system-on-chip, since it simplifies the application behaviour and reveals the top-level structure of the behaviour, abstracting out the low-level details [4]. A critical issue for MPSoC design is to evaluate the expected performance, early in the design process before hardware implementation [5] and hence real-time scheduling and schedulability analysis for multiprocessor systems have become an important research area [6].
Algorithms for multiprocessor scheduling can be classified as partitioned scheduling and global scheduling. Tasks which are assigned statically to processors and task migration between processors are not allowed in partitioned scheduling, whereas, in later approach, ready tasks are enqueued in a global queue and tasks are assigned dynamically to available processors. All fixed-priority algorithms based on partitioned scheduling [7–10] and global scheduling [11, 12] have lower utilization bounds. Based on fixed priority, there is a little advantage of global scheduling over partitioned scheduling and rate monotonic algorithm is extensively researched and implemented successfully in conjunction with UML profile for schedulability analysis [13].
When scheduling single-processor systems, nonpreemptive has been considered inferior because of its poor responsiveness. However, in multiprocessor systems, high-priority tasks will still have a chance to execute on available processors to meet deadlines. Moreover, nonpreemptive scheduling algorithms are easier to implement and have lower runtime overhead [14]. In [14], under many parameter settings, experimental simulation surprisingly shows that, in multiprocessor environment, global nonpreemptive fixed-priority scheduling (NP-FP) outperforms global preemptive fixed-priority scheduling (P-NP).
In this paper, a MPSoC system is modelled using UML-RT and schedulability of nonpreemptive rate monotonic algorithm-shortest job first (RMA-SJF) for periodic task sets on three homogeneous multiprocessors is analysed for various load conditions. As a technical contribution, we present our simulation experiments comparing the success ratio and effective processor utilization for global, semipartitioned, and partitioned techniques to validate our work.
This paper improves upon previous publications in the following aspects.(1)Our result shows better performance than conventional EDF algorithm and D_EDF [15].(2)In [15] periodic tasks are preemptive whereas our contribution considers nonpreemptive tasks, which were considered inferior to preemptiveness, and still achieves good schedulability for overloaded conditions.(3)Runtime overhead will be much more reduced in our proposal because EDF is dynamic whereas rate monotonic algorithm is a static algorithm.
The paper is constructed as follows. Overview of UML-RT is presented in Section 2. We present MPSoC model using UML-RT in Section 3. Rate monotonic algorithm (nonpreemptive) for global, partitioned and semi-partitioned scheduling are analyzed in Section 4. Simulation experiments and performance evaluation are done in Section 5. Section 6 concludes the paper.
2. Object-Oriented Models
Object-oriented programming upgrades procedural programming in the areas of adaptability, understandability, and code reusability. Real-time object-oriented model must be expressed unambiguously and explicitly to represent synchronization and concurrency among processes. Currently available object-oriented real-time models are weak in specifying temporal and behavioural requirements and also lack schedulability analysis [16]. RTSO-RAC model [17] does not explore timeliness and schedulability aspects of the model. OPM/T [18] does not describe the effects of priority assignments and concurrency of the model. In TMO [16] (time triggered and message triggered objects), the model success depends on designers’ knowledge of underlying hardware platform. CHAOS [19] does not specify timeliness aspects of real-time system and seems reasonable only for soft-real-time system. Some of the popular tools including Rhapsody, ObjectTime developer, and IBM Rational Rose Real Time provide a framework for analysing the problem space in real-time domain.
Object-oriented models like ADARTS (ADA based design approach for real time), CODARTS (concurrent design approach for real time), HRT-HOOD (hard-time hierarchical object-oriented design) [20], UML-RT [21], and ROOM (real-time object-oriented modelling) [22] use object-oriented notations to capture temporal properties of a real-time system [23]. Limitations of ADARTS and CODARTS are that they are designed mainly for ADA and use limited number of views. With the introduction of UML, ObjectTime has cooperated with Rational software to develop UML-RT, which uses UML’s in-built extensibility mechanisms to integrate ROOM concepts within UML. UML-RT and the code generation technology of ObjectTime Developer have been integrated into Rational Rose in the new product Rational Rose Real Time [1]. UML-RT includes all the modelling capabilities of ROOM. The recently standardized UML profile for modelling and analysis of real-time and embedded systems (UML MARTE profile) has been provided by OMG group. But, still when it comes to scheduling, UML-RT profile introduces a set of common scheduling annotations which are fairly sufficient to perform schedulability analysis [24].
2.1. UML and UML-RT
Rational Rose Real time is a software development environment tailored to the demands of real-time software [9]. Developers use Rational Rose Real Time to create models of the software system based on the UML constructs, generate the implementation code, compile, and then run and debug the application. Rational Rose Real Time can be used through all phases of the software development lifecycle, from initial requirements analysis through design, implementation, test, and final deployment.
The tool, named UML-RT for real time, developed by the Rational Corporation, uses UML to express the original ROOM (real-time object-oriented modelling) concepts and their extensions. It includes constructs for modelling both the structure and behaviour of event-driven real-time systems.
Rational Rose Real Time includes features for(i)creating UML-RT models using the elements and diagrams defined in the UML-RT;(ii)generating complete code implementations (applications) for those models;(iii)executing, testing, and debugging models at the modelling language level using visual observation tools.
In UML-RT the three principle constructs for modelling structure are capsules, ports, and connectors. UML-RT is a profile that extends UML with stereotyped active objects, called capsules, to represent system components. The internal behavior of a capsule is defined using statecharts; its interaction with other capsules takes place by means of protocols that define the sequence of signals exchanged through stereotyped objects called ports. The UML-RT profile defines a model with precise execution semantics; hence it is suitable to capture system behavior and support simulation or synthesis tools (e.g., Rose RT).
2.1.1. Capsules
The fundamental modelling constructs of UML-RT are capsules.(i)They are possibly distributed architectural active objects that interact with other capsules exclusively through one or more ports.(ii)The behaviour of a capsule is modelled in the state transition diagram that can process (send and receive) messages via their ports, while its (hierarchical) structure is modelled in the capsule structure diagram.
2.1.2. Ports
Messages are sent to and received from capsule instances through objects known as ports. Ports connected to a state machine of a capsule (end port) can handle messages sent to them.
2.1.3. Connectors
The key communication relationships between capsule roles are captured by connectors. They interconnect capsule roles that have similar public interfaces through ports.
2.1.4. Protocols
(i)They define a set of messages exchanged between a set of capsules.(ii)Messages are defined from the perspective of both the receiver and the sender.
3. Multiprocessor Model Using UML-RT
The behaviour of a real-time model is composed of periodic independent tasks with three homogeneous processors. A real-time system must be analysed to ensure that all the tasks are schedulable. A multiprocessor system is modelled with active objects called capsules in UML-RT and shown in Figure 1. Two capsules are created, one for generating tasks named “gentask” and the other named “scheduler” to schedule the generated tasks. Using the property of inheritance, three capsule instances are created for the three homogeneous processors where the tasks are executed and named as “processor1,” “processor2,” “processor3.” The ports used in each capsule are shown in Table 1.
3.1. “gentask” Capsule
This is the task activation model where task sets are generated. The tasks to be scheduled are triggered in the “gentask” capsule. The output port “gout” of the “gentask” capsule is connected to the input port “sin” of the “scheduler” capsule. The input port “gin” of the “gentask” capsule is connected to the output port “sout” of the “scheduler” capsule.
3.2. “scheduler” Capsule
In “scheduler” capsule, priorities are assigned to each task and tasks are scheduled as per the algorithm.
3.3. “processor1” Capsule
The output port “p1out” of the “scheduler” capsule is connected to the input port “pin11” of the “processor1” capsule. Likewise, the output port “pout11” of the “processor1” capsule is connected to the input port “p1in” of the “scheduler” capsule.
3.4. “processor2” Capsule
The output port “p2out” of the “scheduler” capsule is connected to the input port “pin22” of the “processor2” capsule. Likewise, the output port “pout22” of the “processor2” capsule is connected to the input port “p2in” of the “scheduler” capsule.
3.5. “processor3” Capsule
The output port “p3out” of the “scheduler” capsule is connected to the input port “pin33” of the “processor3” capsule. Likewise, the output port “pout33” of the “processor3” capsule is connected to the input port “p3in” of the “scheduler” capsule.
4. System Description
For a periodic task system, let task set consist of a set of “” tasks . Each task is defined by two parameters: computation time and interrelease time or period . So, a task set can be denoted by , where “” is the th task considered. For a task , the task utilization is , and let the system load . A set of all higher priority tasks of are denoted by ; similarly, a set of all lower priority tasks of are denoted by .
The scheduling strategy of a task to be schedulable in an “M” identical nonpreemptive multiprocessor architecture is analyzed. Depending on the number of processors “M” in the system, the total tasks in the system are split into higher priority tasks to , intermediate tasks to , and lower priority tasks to . Therefore, there must be tasks in the system. The work area is shown in Figure 2. For a task , is the th release time and it is an integer multiple of . The condition for a task to be schedulable is that the processor must start to execute at least at or the total work load contributed by other tasks in work area must be less than or equal to , so that the processor is available to execute on or before . Therefore, the contribution of total work load of the other tasks in the work area is analyzed under pessimistic conditions.
The work load in can be categorized into three parts as reported by Guan et al. [6], shown in Figure 3.
The total work contributed in the work area can be classified into three works: initial, intermediate, and final work.
Initial Work. Any task will contribute the initial work, if its execution started before and finishes in the work area.
Final Work. Any task will contribute the final work, if its execution started in work area and finishes later to .
Intermediate Work. The contribution of any task started on or later to and finishes on or before .
Theorems and lemmas are analyzed considering that the entire task contributes maximum in the work area, and task , released at , is always scheduled before . The given computation times are arranged in nondecreasing order. From the known computation time, let
The maximum computation times are assigned as shown from (1): A task will be schedulable if the following is satisfied:
Equation (2) gives the total work load contributed by the other tasks in the work area. Equation (3) is the schedulability condition for a task to be schedulable on an “” identical nonpreemptive multiprocessor architecture. Equation (3) can be written as follows as stated by Guan et al. [6]: where , and are the initial, intermediate, and final work contribution by other tasks, respectively.
The final work contributions are bounded by the second term in the RHS of (4). Therefore, (4) can be written as follows: Recall that Equation (6) is substituted in (5) to obtain (7) and (8): Equation (9) must be satisfied for the task to be schedulable.
Lemma 1. An optimum condition for the highest priority task to be schedulable by a work conserving nonpreemptive algorithm is given by
Proof. To prove (10), consider the time duration given below:(i).
Tasks in execution before in “M” processors will be because is the highest priority task considered. For “M” processor system, “M” initial works contributed by “M” lower priority tasks (with maximum computation time) under pessimistic conditions are considered. Figure 4 shows that there are “M” tasks , which are in execution during .
Therefore, initial work contributed by is obtained from (1) and summation of “M” initial work is considered as total initial work, for task , and it is shown in:
To Prove by contradiction, if highest priority task satisfies (10) but is not schedulable, then task misses its deadline. For to miss its deadline, “M” processors must be continuously busy in the work area , which means that . This contradicts the assumption that satisfies (10). Therefore, (10) is the optimum condition for task to be schedulable.
Lemma 2. Maximum waiting time of highest priority task for a “M” processor system is shown in
Proof. To prove consider, consider(i).
Tasks in execution before in “M” processors will be because is the highest priority task considered. In worst case condition, “M” tasks having maximum computation time will contribute initial work to as proved in Lemma 1. Consider(ii).
Task is the highest priority task, released at , and is waiting to be executed in any of the “M” processors. The task that finishes first will be the task having minimum execution time during . Therefore, during when any one processor is free, starts to execute. It will be the minimum time of all initial works as given in (12).
Lemma 3. The maximum waiting time of periodic task for an “M” processor system is given in where is considered as initial work contribution to .
Proof. By Lemma 2 it is proved that the maximum waiting time for the task is ; therefore, has to wait for to finish execution and also for the other , executing on “” processors.
To prove consider, consider:(i).
The same is considered as in Lemma 2:(ii).
Consider Figure 5; assume that the task executes on one of the processors, and the other “” executes on “” processors. Task will start to execute on a processor, which becomes free first. During , assume that one processor becomes free; therefore, it will be the minimum time as shown in (13). To generalize, for to tasks,
Lemma 4. Consider a task ; if , then “M” tasks will contribute “M” initial work under pessimistic condition.
Proof. To prove consider the time duration as below:(i).
The tasks in execution before in “M” processors will be either or because task is not yet released.
Under pessimistic condition, it is considered that “M” tasks are in execution during because all have greater computation time compared to . Since, , there are “M” tasks contributing initial work to , as shown in Figure 6. Therefore, the initial work contribution to task , if , is given in
Lemma 5. For , “” tasks will contribute a worst case of “” initial work, and “” tasks will contribute the remaining initial work.
Proof . To prove consider the time duration as below:(i).
Tasks in execution before in “” processors will be either or . Since a pessimistic condition is analyzed, there is only “” task for the above condition, contributing the initial work and “” task having greater computation time will contribute the remaining initial work. Therefore,
Therefore, (16) shows the initial work for the task if .
5. RMA-SJF Algorithm
With known computation times, the aim is to design a work conserving task system to utilize the available processing capacity using RMA-SJF priority scheme. A task set is derived, which satisfies RMA-SJF from the known computation time; that is, the higher priority task possesses lesser computation time and lesser interrelease time. It is named as work area analysis (WAA), which is as follows.
The known computation timings are arranged in nondecreasing order: . It is considered that, if there are “” processors in the system, then there must be minimum of “” tasks for the proposed algorithm to work. Therefore, the task set can be divided into three categories according to priority, as given below:(i)higher priority tasks to ;(ii)intermediate tasks to ;(iii)lower priority tasks to .
Case 1 (higher priority tasks ( to )). Consider From (17), the interrelease times for each higher priority tasks are found using Lemmas 1, 2, and 3.
Case 2 (intermediate tasks ( to )). Considering (9), is obtained from (15) and (16) depending on the kth value of the task for intermediate tasks: From (19) to (21), interrelease times of intermediate tasks are found.
Case 3 (lower priority tasks ( to )). Consider
Initially, the interrelease times are derived for tasks to as shown from (22) to (25). In (23), is subtracted because it does not form the initial work. It is the computation time to be analyzed for schedulability. For the least priority task, in (26); the interrelease time is to be found. It is the only unknown in (27); it is found using condition for schedulability. The condition for a set of tasks to be feasibly scheduled on a multiprocessor system is that its system load should be less than or equal to .
6. Global Scheduling
For randomly generated computation times, interrelease times for tasks are derived using WAA analysis. The derived task set possesses the RMA-SJF priority scheme; that is, higher priority tasks will have lesser computation time and lesser interrelease time. The task set is analyzed for global, partitioned, and semipartitioned scheduling strategies and success ratio and effective processor utilization of tasks are analyzed.
Initially, task set is analyzed for global scheduling where tasks are assigned dynamically to the available processors.
The pseudocode for global scheduling is as shown in Pseudocode 1.
| ||||||||||||||||||||||||
7. Partitioned Scheduling
In partitioned scheduling, tasks are first assigned to specific processors and executed without migrations. The pseudocode for partitioned scheduling is as shown in Pseudocode 2.
| ||||||||||||||||||||||||
8. Semipartitioned Scheduling
In semipartitioned scheduling, some tasks are global and others are partitioned. The pseudocode for partitioned scheduling is as shown in Pseudocode 3.
| ||||||||||||||||||||||||
9. Simulation and Performance Evaluation
Simulation work is carried out for various load values. Figures 7, 8, and 9 show the total number of released tasks and total number of scheduled tasks for the load of 2.94 for partitioned, semipartitioned, and global scheduling, respectively. From the results, it is observed that the global scheduling utilises processors a little more efficiently when compared with partitioned and semipartitioned scheduling, thus increasing the schedulability. The success ratios of the three proposed scheduling methods are compared in Figure 10; it is inferred that the success ratio is comparatively more in global scheduling than in partitioned and semipartitioned scheduling. Effective processor utilization for partitioned, semipartitioned, and global scheduling is calculated for various loads and shown in Figure 11.
Global scheduling utilizes processor more efficiently when compared with partitioned and semipartitioned scheduling as analysed from Figure 11. Guan et al. [6] conducted simulation experiments empirically comparing the real-time performance of preemptive and nonpreemptive global fixed-priority scheduling, by which they obtained interesting results suggesting that, for a considerably part of applications on multiprocessor platforms, nonpreemptive scheduling is actually a better choice than preemptive scheduling regarding the real-time performance.
Therefore, obtained results are also compared with the success ratio for the same load by Thakor and Shah [15], and they show that the global RMA_SJF outperforms the global preemptive EDF (earliest deadline first) and D_EDF (deadline monotonic_earliest deadline first) in schedulability for the same load. Figure 12 shows success ratio analyzed with EDF, D_EDF, and RMA-SJF and Figure 13 shows the effective processor utilization is analyzed with EDF, D_EDF, and RMA-SJF. From the analysis, it is inferred that the RMA-SJF outperforms EDF and D_EDF in success ratio by effectively utilizing the processor.
10. Conclusion
In single processor scheduling, nonpreemptiveness leads to poor task responsiveness because higher priority tasks are blocked by lower priority tasks, but, in the case of multiprocessor environment, higher priority tasks still have chance to execute in available processors. Moreover, nonpreemptiveness enjoys benefits like lower implementation complexity and lower runtime overhead [6]. Our contribution considers nonpreemptive periodic tasks, scheduled using rate monotonic algorithm-shortest job first (RMA-SJF) on multiprocessor environment and modelled using modelling language UML-RT. In this newly proposed algorithm, the interrelease time for each task is derived from the known computation times and schedulability conditions. RMA-SJF is analysed for various scheduling principles such as global, semipartitioned, and partitioned scheduling for various system loads. Our result shows that the global scheduling utilizes processors little more efficiently when compared to partitioned and semipartitioned scheduling, thus improving schedulability. When compared with the success ratio and effective processor utilization for the same load in [15], RMA-SJF is analysed for global, semi-partitioned and partitioned scheduling strategies for various system loads.
Notations
| : | Problem area of the task , analyzed for schedulability; a necessary condition for the deadline miss to occur for is that the worst case work load in the problem area by all other tasks in the task set except is no less than |
| : | Set of all higher priority tasks of |
| : | Initial work for |
| : | Initial work for |
| : | Summation of all the task utilization of |
| : | Intermediate work |
| : | th release time of |
| : | th release time of and deadline for the task released at |
| : | Latest feasible start time for released at to start execution in order to meet its deadline |
| : | Number of intermediate tasks in |
| : | Work done by initial job, intermediate job, or final job in |
| : | Worst case latency of ; it is the maximum time lapse for a task to start executing |
| : | Total work done by other tasks in the problem area of |
| : | Final work |
| : | Set of all lower priority tasks of . |
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.