Access Official Technical PDF

Share Summary:LinkedInEmail

In industrial automation, embedded controllers and other time-sensitive systems must deliver results within defined deadlines. Some applications require hard real-time guarantees, where every deadline must be met. Others can tolerate occasional misses, provided their likelihood is acceptably low and can be quantified. Estimating deadline miss probability gives engineers a way to assess this second category without treating every job as if it always experiences its longest possible execution time.

This analysis is relevant when job execution time can vary among known modes, each with a specified probability. A longer mode might represent additional processing, such as recovery work after a transient fault. Planning around only the longest execution time can be unnecessarily conservative, while ignoring timing risk is not an option. The practical challenge is to estimate the risk efficiently enough to evaluate realistic task sets.

Why deadline miss probability matters

In a soft real-time system, a deadline miss does not necessarily mean that the system has failed. The design requirement may instead be to keep misses sufficiently rare. Quantifying that probability can support decisions about task parameters and processor resources, and can help assess systems where recovery mechanisms add execution time only in particular circumstances.

The analysis considered here focuses on independent periodic or sporadic tasks scheduled on a uniprocessor using preemptive fixed priorities. Each task has a relative deadline, a minimum interarrival time, and a set of execution modes. A job selects one mode, with its associated worst-case execution time, according to a known probability distribution. The probability of each job’s mode is assumed independent of the modes selected by other jobs.

For a task under analysis, higher-priority jobs can interrupt its execution. The key question is whether the combined workload from those jobs and the task’s own job exceeds the available time before its deadline. An upper bound on the probability of that overload provides an upper bound on deadline-miss probability. The analysis considers relevant time points within the deadline window and uses the lowest overload bound among them.

Why straightforward convolution can be costly

A direct way to calculate the workload distribution is to combine, or convolve, the execution-time distributions of jobs that can run in the interval. For every possible combination of execution modes, the method adds the workloads and multiplies their probabilities. Combinations that produce the same workload can then be merged.

The difficulty is that the number of combinations can grow rapidly with the number of jobs. Even when many combinations ultimately represent the same total workload, a job-by-job calculation may have to construct and process those combinations first. This state-space growth makes traditional convolution expensive for longer analysis windows or larger task sets.

The central improvement in the multinomial approach is to recognize that the order of execution modes among jobs from the same task does not affect their total workload or joint probability. If a task releases several jobs, the analysis can count how many select each mode rather than enumerate every ordering. The multinomial distribution accounts for the number of ways those mode counts can occur and the probability of the resulting group.

Multinomial analysis reduces redundant states

Consider a task with two execution modes and several jobs in the interval. A job-by-job method distinguishes sequences such as “short, long, short” and “long, short, short” while building the state space. If the jobs are independent and share the same mode probabilities, both sequences have the same total workload and probability. The multinomial representation groups them into one equivalent class, weighted by the number of possible sequences.

After forming these task-level workload distributions, the analysis combines them across tasks to determine the probability that total workload exceeds the available interval. This avoids creating many intermediate states that differ only in the order of a task’s execution modes. It does not eliminate all computational growth, but it can make the calculation more practical than direct job-by-job enumeration.

Pruning and controlled approximation

Further reductions can improve efficiency:

  • Exact pruning: States that cannot contribute to the result can be discarded without changing the calculated probability. For example, a state may be known to remain within the available time, or to exceed it, as subsequent workload is considered.
  • Equivalence-class merging: States with equivalent workload properties can be grouped. The analysis can also merge classes using an approximation whose precision loss is bounded in advance.

These techniques offer a way to balance runtime and precision. Exact pruning preserves the result, while approximation through state merging trades some precision for reduced computation with a known bound on the introduced loss.

Analytical bounds offer a faster alternative

When a fast estimate is more important than a tight probability, concentration inequalities can provide analytical upper bounds. The approach applies Hoeffding’s and Bernstein’s inequalities to the sum of independent job execution times.

These methods use characteristics of the execution-time distributions to bound the chance that total workload exceeds the time available. Hoeffding’s inequality uses the possible minimum and maximum execution times. Bernstein’s inequality also uses variance, which can produce a more informative bound when that distribution information is available. If the expected workload is already at or above the interval limit, the stated safe bound is 1; otherwise, the inequalities provide an upper estimate.

Analytical bounds avoid enumerating workload combinations and can therefore be computed efficiently. Their trade-off is precision: an upper bound may be conservative, and its tightness is not guaranteed. They are useful when a rapid, safe overestimate is sufficient, but less suitable when the decision depends on a precise probability.

Choosing an approach for engineering analysis

The appropriate method depends on the purpose of the assessment. Multinomial analysis is suited to cases where a more precise probability estimate is important and the task and execution-mode structure makes grouping useful. Pruning can further reduce work, while bounded approximation offers an explicit way to exchange some precision for speed.

Analytical bounds are a reasonable option for early screening or other evaluations where a conservative estimate is acceptable. They can help identify whether a workload is clearly separated from the available timing budget, but a loose bound may not resolve a close design decision. Traditional convolution remains conceptually direct, yet its job-level state growth can limit its usefulness as the number of interfering jobs increases.

Conclusion

Deadline-miss analysis for soft real-time workloads is a balance between computational cost and confidence in the probability estimate. Grouping execution-mode combinations by task and using multinomial distributions can remove substantial redundant enumeration. Exact pruning preserves accuracy, bounded state merging makes the precision trade-off explicit, and Hoeffding or Bernstein bounds provide faster alternatives when overestimation is acceptable. For industrial automation engineers and system integrators, selecting among these methods means matching the analysis to the timing risk, workload structure, and precision required by the design decision.

Access Official Technical PDF

About The Author