• No results found

Case Study for the AR model

Related work

2.3 Abort-and-Restart Model

2.3.5 Case Study for the AR model

There are two cases [69] which are implementations in both software and hardware. The case study for software that the authors used is the Generic

Avionics Platform (GAP) task-set to evaluate the AR model under RM and Earliest Deadline First (EDF) scheduling and then compare the result with other models such as non-preemptive, PCP and Stack Resource Policy (SRP). The other case is for hardware; the Analog Devices’ ADuC814 micro-controller is used for running the P-FRP compiled code with RM and EDF scheduling. The result of this case study is the number of tasks against the average number of aborts and the average number of aborts against system load.

There are two diagrams of the case study for software shown in Figures 2.29 and 2.30. Figures 2.31 and 2.32 are the result of the case study for the hardware. We directly cited all diagrams from the paper by Ras and Cheng [69].

Software — Generic Avionics Platform

The case study for software is the GAP [52, 60], is used in the simulation.

GAP models the functionality of an aircraft computer system and data han-dling that was created by Locke et al [61]. In the paper by Ras and Cheng [69], the authors list some of the avionics timing constraints, as below:

1. Navigation: The frequency of navigation is 20 Hz, which is based on the requirements of accuracy.

2. Display: The period is between 65 ms and 100 ms.

3. Ballistics Computation: The vehicle trajectory, altitude and attitude require 5 ms in period.

4. Sensor Control: The frequency of radar antenna search is 10 Hz. 1KHz or more is for electromagnetic surveillance equipment.

Ras and Cheng [69] used the theory of the GAP task-set which has sixteen periodic tasks and one sporadic. In the experiment, the authors assume that all tasks are periodic and have no release jitter. The motivation is to observe how much penalty is introduced by the AR model because they know that PCP and SRP are apparently more efficient. RM priority assignment is used

with the heavy and lighter resources usage that produces the two results shown in Figures 2.29 and 2.30. They also state that the performance of RM is based on the arrival pattern and synchronous release is the worst case for RM.

Figure 2.29: Heavy resource usage (long critical sections). (Cited from the paper [69])

Figure 2.29 shows the results of different policies (AR, SRP, PCP and non-preemptive) with RM priority ordering under heavy resource usage. The AR model has the worst performance, PCP and SRP perform well and the non-preemptive model is just better than the AR model. This diagram shows the result under heavy resource usage which means the system has long critical sections.

The next diagram, in Figure 2.30, is about the lighter resource usage in which the system has shorter critical sections. In this test, SRP is better in performance. PCP, AR and non-preemptive have the same performance as the test with heavy resource usage. In the paper [69], there are also two tests based on EDF scheduling which is a dynamic priority scheduling; we do not

discuss this approach in detail in this review.

According to the observation of the experiments, Ras and Cheng did not realise that the critical instant for P-FRP may not occur when all tasks are released at the same time. The result of the AR model could get worse because it is not based on the worst case. From our observation, the per-formance of the AR model has a lot of room for improvement because RM priority assignment is not optimal for the AR model.

Figure 2.30: Light resource usage (short critical sections). (Cited from the paper [69])

Hardware — Analog Devices’ ADuC814

Ras and Cheng also present an experiment with hardware implementation in their paper [69]. They run P-FRP compiled code on the hardware board which uses an Analog Devices’ ADuC814 micro-controller. The results are shown in Figures 2.31 and 2.32. This time we discuss both fixed and dynamic priority scheduling.

Figure 2.31: Number of tasks. (Cited from the paper [69])

In Figure 2.31, there is a diagram of the average number of aborts against the number of tasks with RM and EDF scheduling, and the utilisation is fixed.

The average number of aborts for RM scheduling is always higher than EDF scheduling. When the number of tasks gets higher, the average number of aborts decreases. The authors believe the reason is that the higher number of tasks makes the worst-case computation time get smaller for each task. It makes the possibility of aborts occurring lower.

Figure 2.32 shows the average number of aborts against the CPU load, and the utilisation is fixed. In this result, Ras and Cheng have less discussion about it. But we can see from the diagram that the average number of aborts for the RM is increasing when the CPU load is getting higher. On the other hand, the average number of aborts for EDF is decreasing after the CPU load is 0.85.

Based on our observation, EDF scheduling is much more efficient than fixed priority assignment with RM priority assignment for the AR model.

RM is not optimal for P-FRP but the number of aborts is an important

Figure 2.32: System load. (Cited from the paper [69])

property for schedulability analysis in P-FRP.