Worst-Case Deadline Failure Probability

• WCDFP is a worst-case measure that quantifies the maximum probability of deadline misses by maximizing over admissible release patterns in soft real-time systems.
• It encompasses various formulations including job-level, weakly hard constraints, and distributed system scenarios, emphasizing the role of event semantics and busy-period analysis.
• Analytical approaches for WCDFP range from direct miss-probability maximization and busy-period decomposition to conservative surrogate estimators, highlighting computational challenges and application-specific trade-offs.

Worst-Case Deadline Failure Probability (WCDFP) is a worst-case measure of deadline-miss risk under timing uncertainty. In one explicit formulation for a periodic soft real-time task, WCDFP is defined as $\sup_{j \in \mathbb{N}} DM(j)$, where $DM(j)$ is “the probability that the $j$-th job misses deadline” (Chen et al., 2024). In probabilistic fixed-priority preemptive scheduling for sporadic tasks, the corresponding task-level quantity is

$$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$

that is, the largest deadline-miss probability over all admissible release patterns and all jobs of task $\tau_i$ (Liu et al., 28 Aug 2025). A closely related formulation writes

$$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$

with $\mathcal{RT}_{i,j}^{\xi}$ the probabilistic response time under arrival scenario $\xi$ (Takahashi et al., 1 Dec 2025). Across the literature, however, the failure event need not always be a single-job deadline miss: it may also be a weakly hard constraint violation, a packet drop before end-to-end deadline, or failure to finish a distributed job before a target deadline. WCDFP is therefore best understood as a family of worst-case deadline-failure notions whose exact semantics depend on the system model and the event being bounded.

1. Definitions and event semantics

The most important distinction in WCDFP work is the level at which “failure” is defined. In fixed-priority probabilistic schedulability, failure is usually job-level: a job of task $\tau_i$ fails if its random response time exceeds $D_i$, and WCDFP takes the maximum such probability over jobs and admissible release patterns (Liu et al., 28 Aug 2025). In the revised-critical-instant formulation, the same object is expressed through processor-demand-based probabilistic response time, with arrival scenarios $DM(j)$0, service, carry-in, and aborted execution time contributing to

$DM(j)$1

and WCDFP is then the maximum of $DM(j)$2 over $DM(j)$3 and $DM(j)$4 (Takahashi et al., 1 Dec 2025).

In weakly hard systems, the failure event is higher level. The design-stage analysis for weakly hard real-time systems distinguishes the per-job miss event

$DM(j)$5

from the weakly hard violation event $DM(j)$6 and the system unsafe event

$DM(j)$7

The learned probability in that setting is approximately $DM(j)$8, not the probability of a single miss event (Lee et al., 2023).

A further distinction appears in the literature on long-run soft real-time behavior. There, WCDFP is explicitly contrasted with deadline miss rate (DMR): WCDFP is a supremum over per-job miss probabilities, whereas DMR is a long-run asymptotic miss frequency (Chen et al., 2024).

Context	Failure event	Representative quantity
Probabilistic fixed-priority scheduling	$DM(j)$9	$j$0
Periodic soft real-time task	$j$1	$j$2
Weakly hard real-time system	violation of weakly hard constraints	$j$3
Distributed convolution	job not finished by deadline $j$4	$j$5

This variation in event semantics is not incidental. It determines what is being upper-bounded, what uncertainty is maximized over, and whether the result is a direct WCDFP, a certified deadline-failure bound, or only a conservative surrogate.

2. Analytical forms across application domains

One major strand of work studies deadline-failure probability directly as a stochastic tail event. For coded distributed convolution, the relevant quantity is

$j$6

an explicit deadline failure probability. Its asymptotic behavior is characterized through

$j$7

with the exponent controlled by the worst-case recovery threshold $j$8. For coded convolution,

$j$9

and the tail coefficient is

$$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$0

This is deadline-failure analysis informed by worst-case straggler tolerance rather than a minimax worst-case-over-adversarial-timing probability (Dutta et al., 2017).

A different analytical path appears in hierarchical plans. There the central object is the makespan random variable $$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$1, and the relevant deadline quantities are

$$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$2

Exact evaluation is intractable in general: the paper proves that deciding whether the probability that a task tree satisfies a deadline exceeds a threshold is NP-hard, that $$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$3deadline-probability is $$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$4-complete, and that computing expected completion time of a mixed task network is NP-hard. The proposed deterministic approximation uses one-sided Kolmogorov bounds $$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$5, which imply

$$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$6

That is not adversarial WCDFP, but it is a certified conservative bound on deadline-failure probability under the paper’s probabilistic model (Cohen et al., 2015).

Multihop networking introduces a third interpretation. There the objective is to maximize the weighted number of packets delivered before hard end-to-end deadlines, not to compute an explicit WCDFP. Nevertheless, under i.i.d. Bernoulli or Binomial arrivals, the probability of a packet being dropped is at most $$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$7 if

$$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$8

and the proposed algorithms achieve $$\mathrm{WCDFP}_{i} = \max_{\xi \in \Xi}\max_{x \ge 1}\{P(R_{i,x}^{\xi}>D_i)\},$$9-type approximation guarantees to optimal timely throughput under stochastic traffic (Tsanikidis et al., 2023). This yields deadline-failure information indirectly, through excess miss fraction relative to optimal.

Taken together, these domains show that WCDFP-oriented reasoning can take at least three forms: direct miss-probability maximization, certified deadline-failure approximation, and deadline-tail or drop-probability analysis tied to worst-case structural parameters such as recovery threshold or minimum capacity.

3. Busy periods, critical instants, and safe upper bounds

A central development in recent WCDFP research is the rejection of the classical deterministic critical-instant assumption as a safe basis for probabilistic analysis. In deterministic fixed-priority scheduling, one busy period starting at the critical instant suffices. In probabilistic analysis, this is no longer safe, because the busy-period start time can depend on the realized execution-time pattern (Liu et al., 28 Aug 2025).

The unsafe nature of the classical assumption is demonstrated by a counterexample with two tasks:

• $\tau_i$0: $\tau_i$1, $\tau_i$2, $\tau_i$3
• $\tau_i$4: $\tau_i$5, $\tau_i$6

If both release synchronously at time $\tau_i$7, the miss probability of $\tau_i$8 is $\tau_i$9. Under another release pattern where $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$0 releases at time $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$1 and $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$2 releases at time $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$3, the miss probability becomes $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$4 (Liu et al., 28 Aug 2025). The implication is that probabilistic worst case may occur at a non-critical instant.

The busy-period-based WCDFP analysis addresses this by decomposing the miss event over all possible busy-period starting points. For task $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$5, the miss event is partitioned into subsets $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$6 corresponding to candidate busy-period starts, yielding

$$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$7

Each $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$8 is then bounded by the better of two quantities: $$\mathrm{WCDFP}_i \triangleq \max_{\xi} \max_{j} \mathbb{P}[\mathcal{RT}_{i,j}^{\xi} > D_i],$$9 where

$\mathcal{RT}_{i,j}^{\xi}$0

The final improvement replaces an $\mathcal{RT}_{i,j}^{\xi}$1 accumulation with

$\mathcal{RT}_{i,j}^{\xi}$2

leading to the upper bound

$\mathcal{RT}_{i,j}^{\xi}$3

This is a safe upper-bounding analysis rather than an exact computation (Liu et al., 28 Aug 2025).

A related development uses a revised critical instant. Instead of the classical simultaneous-release assumption, the revised upper bound is

$\mathcal{RT}_{i,j}^{\xi}$4

with

$\mathcal{RT}_{i,j}^{\xi}$5

The additional $\mathcal{RT}_{i,j}^{\xi}$6 term accounts for carry-in released before the analyzed job. Convolution-based methods then estimate this safe upper bound, and a Huffman-inspired merge order reduces aggregate convolution cost from roughly $\mathcal{RT}_{i,j}^{\xi}$7 to $\mathcal{RT}_{i,j}^{\xi}$8 while retaining safe-sided WCDFP estimates (Takahashi et al., 1 Dec 2025).

The conceptual consequence is straightforward: probabilistic deadline-failure analysis must account for carry-in and for multiple busy-period starts. Treating synchronous release as the universal worst case can underestimate the true worst-case miss probability.

4. Conservative surrogates and design-stage estimators

Not all work that is relevant to WCDFP computes WCDFP directly. A substantial design-stage literature instead constructs conservative probabilistic surrogates for safe and unsafe timing regions under uncertainty.

One example is the design-stage method for weakly hard real-time systems on multi-core platforms under BlackBerry QNX’s Adaptive Partitioning Scheduler (APS). The system model includes periodic and aperiodic tasks, WCET ranges

$\mathcal{RT}_{i,j}^{\xi}$9

uncertain context-switch costs $\xi$0, and $\xi$1-style weakly hard constraints. The method combines multi-objective search, via NSGA-II, over arrival sequences and scheduler parameters with polynomial logistic regression in WCET space. The search objectives are the deadline-miss magnitude

$\xi$2

and a heuristic consecutiveness objective

$\xi$3

The learned model is a second-order polynomial logistic response surface,

$\xi$4

Here $\xi$5 is stated to be “the probability of violating deadline constraints,” but, given the labels, it is more precisely the estimated probability that the sampled system configuration is unsafe with respect to weakly hard deadline constraints under the searched and sampled scenarios. The output is a conservative safe region in WCET space, not a formal worst-case deadline-failure probability (Lee et al., 2023).

An earlier design-stage method for real-time systems with WCET ranges uses a related SAFE workflow. It models tasks with $\xi$6, searches for worst-case arrival sequences using a steady-state genetic algorithm and SafeScheduler, and fits a second-order polynomial logistic regression over WCET variables. The probability

$\xi$7

is interpreted as probability of deadline misses, and a threshold is chosen so that all data instances classified as safe are actually safe in the input dataset. This again yields probabilistic safe WCET ranges rather than an explicit supremum over failure probabilities (Lee et al., 2020).

A different but related design perspective appears on DVS platforms with varying worst-case execution cycles (WCEC). There the paper does not define WCDFP; instead, it studies danger zones, task killing or suspension, and rapid update of scheduling functions when actual execution exceeds the assumed WCEC. The closest observable proxy to deadline-failure probability is the empirical killing rate during the transient after a WCEC increase (0809.1132).

These methods are relevant to WCDFP because they search for adverse timing scenarios and because their learned or empirical quantities are conservative in practice. They should not, however, be conflated with direct methods that compute or safely upper-bound a formally defined worst-case deadline-failure probability.

5. Metrics commonly conflated with WCDFP

A recurring source of confusion is the use of deadline-related metrics that are not WCDFP.

The clearest example is the deadline miss rate in the long run. For a periodic soft real-time task with relative deadline $\xi$8, period $\xi$9, dismiss point $\tau_i$0, and iid discrete execution times, the miss indicator $\tau_i$1 is used to define

$\tau_i$2

The long-run DMR is then defined through almost-sure convergence: $\tau_i$3 Under irreducibility and positive recurrence of the finite Markov chain model, the DMR equals stationary mass over miss states,

$\tau_i$4

This is a time-average miss frequency. It is not a supremum of per-job miss probabilities, and it is therefore not WCDFP (Chen et al., 2024).

Weakly hard miss-count bounds provide another non-equivalent metric. In distributed real-time embedded systems with transient soft errors, weakly hard schedulability is written as

$\tau_i$5

where $\tau_i$6 is the maximum number of deadline misses of task $\tau_i$7 in any $\tau_i$8 consecutive activations. Typical Worst-Case Analysis (TWCA) yields the upper bound

$\tau_i$9

This supports conservative ratios such as $D_i$0 or $D_i$1, which are upper bounds on miss density or worst-case failure frequency over a horizon. They are not probabilities unless additional stochastic assumptions are added (Liang et al., 2020).

The asymptotic failure exponent is also distinct from WCDFP. In coded convolution, the paper studies

$D_i$2

through the tail exponent

$D_i$3

This directly characterizes tail reliability and is highly informative for deadline-sensitive systems, but it is not a worst-case-over-admissible-scenarios probability in the minimax sense (Dutta et al., 2017).

The practical consequence is that deadline miss rate, weakly hard miss bounds, empirical kill rate, and asymptotic failure exponent can all support deadline-risk reasoning, but none is interchangeable with a formal WCDFP without an explicit statement of the event space, the uncertainty space, and the maximization domain.

6. Assumptions, hardness, and scope

WCDFP results are highly model-dependent. Representative exact or safe-upper-bound analyses for fixed-priority scheduling assume a discrete-time, single-processor setting, independent sporadic tasks, constrained deadlines $D_i$4, discrete execution-time distributions, iid copies of each task’s execution time, and abort-on-deadline-miss semantics (Liu et al., 28 Aug 2025). The revised-critical-instant convolution method assumes fully preemptive fixed-priority scheduling, independent task pWCET distributions, IID jobs per task, constrained deadlines, and immediate discard of tardy jobs (Takahashi et al., 1 Dec 2025).

Approximation methods impose their own structure. The hierarchical-plan approximation assumes independent primitive-task duration random variables, discrete distributions with finite support, and a task tree composed of sequence and parallel nodes (Cohen et al., 2015). The weakly hard design-stage estimator depends on the quality of the NSGA-II search over arrival sequences and context-switch settings, uses random sampling within WCET ranges rather than a formally justified execution-time law, optimizes heuristic lateness and consecutiveness objectives rather than direct miss probability, and has a quadratic learning step in the number of WCET-ranged tasks because of second-order polynomial terms (Lee et al., 2023).

Computational hardness also constrains what can reasonably be expected from WCDFP analysis. Exact probabilistic deadline computation is already NP-hard for sequential hierarchical plans, $D_i$5P-hard in counting form, and expected makespan of mixed sequence/parallel task networks is NP-hard (Cohen et al., 2015). This suggests that exact WCDFP is unlikely to be broadly tractable outside restricted models.

Deterministic service-guarantee frameworks illuminate the opposite end of the spectrum. State-based scheduling for worst-case services introduces necessary and sufficient schedulability conditions

$D_i$6

for a slot-timed constant-capacity server, along with exact characterizations of feasible one-slot decisions. This is not probabilistic WCDFP, but it provides deterministic no-miss guarantees and certificates of inevitable future service shortfall under adversarial arrivals (Xu et al., 2021). A plausible implication is that such deterministic frameworks can serve as conservative backbones for later probabilistic deadline-failure analyses.

WCDFP, in consequence, is not a single universally standardized object. It ranges from exact job-level worst-case miss probability, through safe upper bounds based on busy-period decomposition or revised critical instants, to conservative surrogates learned from simulation and search. The unifying theme is worst-case reasoning about deadline-related failure events under uncertainty; the dividing line is whether the final quantity is a formally defined probability maximized over admissible scenarios, a certified bound under an explicit stochastic model, or only a conservative proxy for unsafe timing behavior.