Definitions for Abstract Response-Time Analysis

In the following, we propose a set of definitions for the general framework for response-time analysis (RTA) of uni-processor scheduling of real-time tasks with arbitrary arrival models. We are going to introduce two main variables of the analysis: (a) interference, and (b) interfering workload.

a) Interference

Execution of a job may be postponed by the environment and/or the system due to different factors (preemption by higher-priority jobs, jitter, black-out periods in hierarchical scheduling, lack of budget, etc.), which we call interference. Besides, note that even the subsequent activation of a task can suffer from interference at the beginning of its busy interval (despite the fact that this job hasn’t even arrived at that moment!). Thus, it makes more sense (at least for the current busy-interval analysis) to think about interference of a job as any interference within the corresponding busy interval, and not just after the release of the job. Based on that rationale, assume a predicate that expresses whether a job j under consideration incurs interference at a given time t (in the context of the schedule under consideration). This will be used later to upper-bound job j's response time. Note that a concrete realization of the function may depend on the schedule, but here we do not require this for the sake of simplicity and generality.

b) Interfering Workload

In addition to interference, the analysis assumes that at any time t, we know an upper bound on the potential cumulative interference that can be incurred in the future by any job (i.e., the total remaining potential delays). Based on that, assume a function interfering_workload that indicates for any job j, at any time t, the amount of potential interference for job j that is introduced into the system at time t. This function will be later used to upper-bound the length of the busy window of a job. One example of workload function is the "total cost of jobs that arrive at time t and have higher-or-equal priority than job j". In some task models, this function expresses the amount of the potential interference on job j that "arrives" in the system at time t.

Next we introduce all the abstract notions required by the analysis.

Consider any type of job associated with any type of tasks...

... with arrival times and costs.

Consider any kind of processor state model.

Consider any arrival sequence ...

... and any schedule of this arrival sequence.

Let tsk be any task that is to be analyzed

Assume we are provided with abstract functions for interference and interfering workload.

In order to perform subsequent abstract Response Time Analyses (RTAs), it is essential that we have the capability to bound interference of a certain type (for example, interference that comes from other tasks' jobs). To achieve this, we define "conditional" interference as a conjunction of a predicate P : Job → instant → bool encoding the property of interest and interference.

In order to bound the response time of a job, we must consider (1) the cumulative conditional interference, ...

... (2) cumulative interference, ...

... and (3) cumulative interfering workload.

Next, we introduce a notion of absence of speculative execution. This notion is not directly related to Abstract RTA, but it is useful when proving the existence of bounded busy intervals. We say that the functions Interference and InterferingWorkload do not allow speculative execution if the cumulative interference never exceeds the cumulative interfering workload. Intuitively, one can interpret this definition as stating that it is not allowed to perform work before it is known whether it is actually needed (i.e., before the work appears as actual workload).

Definition of Busy Interval Further analysis will be based on the notion of a busy interval. The overall idea of the busy interval is to take into account the workload that cause a job under consideration to incur interference. In this section, we provide a definition of an abstract busy interval.

We say that time instant t is a quiet time for job j iff two conditions hold. First, the cumulative interference at time t must be equal to the cumulative interfering workload. Intuitively, this condition indicates that the potential interference seen so far has been fully "consumed" (i.e., there is no more higher-priority work or other kinds of delay pending). Second, job j cannot be pending at any time earlier than t and at time instant t (i.e., either it was pending earlier but is no longer pending now, or it was previously not pending and may or may not be released now). The second condition ensures that the busy window captures the execution of job j.

Based on the definition of quiet time, we say that an interval [t1, t2) is a (potentially unbounded) busy-interval prefix w.r.t. job j iff the interval (a) contains the arrival of job j, (b) starts with a quiet time and (c) remains non-quiet.

Next, we say that an interval [t1, t2) is a busy interval iff [t1, t2) is a busy-interval prefix and t2 is a quiet time.

Note that the busy interval, if it exists, is unique.

In this section, we introduce some assumptions about the busy interval that are fundamental to the analysis.

We say that a schedule is "work-conserving" (in the abstract sense) iff, for any job j from task tsk and at any time t within a busy interval, there are only two options: either (a) interference(j, t) holds or (b) job j is scheduled at time t.

Next, we say that busy intervals of task tsk are bounded by L iff, for any job j of task tsk, there exists a busy interval with length at most L. Note that the existence of such a bounded busy interval is not guaranteed if the schedule is overloaded with work. Therefore, in the later concrete analyses, we will have to introduce an additional condition that prevents overload.

Although we have defined the notion of cumulative (conditional) interference of a job, it cannot be used in a (static) response-time analysis because of the dynamic variability of job parameters. To address this issue, we define the notion of an interference bound. As a first step, we introduce a notion of an "interference bound function" IBF. An interference bound function is any function with a type duration → duration → work that bounds cumulative conditional interference of a job of task tsk (a precise definition will be presented below). Note that the function has two parameters. The second parameter is the length of an interval in which the interference is supposed to be bounded. It is quite intuitive; so, we will not explain it in more detail. However, the first parameter deserves more thoughtful explanation, which we provide next.

The first parameter of IBF allows one to organize a case analysis over a set of values that are known only during the computation. For example, the most common parameter is the relative arrival time A of a job (of a task under analysis). Strictly speaking, A is not known when computing a fixpoint; however, one can consider a set of A that covers all the relevant cases. There can be other valid properties such as "a time instant when a job under analysis has received enough service to become non-preemptive." To make the first parameter customizable, we introduce a predicate ParamSem : Job → instant → Prop that is used to assign meaning to the second parameter. More precisely, consider an expression IBF(X, delta), and assume that we instantiated ParamSem as some predicate P. Then, it is assumed that IBF(X, delta) bounds (conditional) interference of a job under analysis j ∈ tsk if P j X holds.

As mentioned, IBF must upper-bound the cumulative conditional interference. This is done to make further extensions of the base aRTA easier. The most general aRTA assumes that an IBF bounds all interference of a given job j. However, as we refine the model under analysis, we split the IBF into more and more parts. For example, assuming that tasks are sequential, it is possible to split IBF into two parts: (1) the part that upper-bounds the cumulative interference due to self-interference and (2) the part that upper-bounds the cumulative interference due to all other reasons. To avoid duplication, we parameterize the definition of an IBF by a predicate that encodes the type of interference that must be upper-bounded.

Next, let us define this reasoning formally. We say that the conditional interference is bounded by an "interference bound function" IBF iff for any job j of task tsk and its busy interval [t1, t2) the cumulative conditional interference incurred by j w.r.t. predicate Cond in the sub-interval [t1, t1 + Δ) does not exceed IBF(X, Δ), where X is a constant that satisfies a predefined predicate ParamSem. In other words, for a job j ∈ tsk, the term IBF(X, Δ) provides an upper-bound on the cumulative conditional interference (w.r.t. the predicate Cond) that j might experience in an interval of length Δ assuming that ParamSem j X holds.

Consider a job j of task tsk, a busy interval [t1, t2)>> of j, and an arbitrary interval [t1, t1 + Δ) ⊆ t1, t2)>>.

We require the IBF to bound the interference only within the interval t1, t1 + Δ).

Next, we require the IBF to bound the interference only until the job is completed, after which the function can behave arbitrarily.

And finally, the IBF function might depend not only on the length of the interval, but also on a constant X satisfying predicate Param.

As an important special case, we say that a job's interference is (unconditionally) bounded by IBF if it is conditionally bounded with Cond := fun j t ⇒ true.