Deadline Miss Rate (DMR) Overview

• Deadline Miss Rate (DMR) is a metric that quantifies the fraction of jobs or flows missing their deadlines in systems with timing constraints.
• It is computed using techniques such as simulation, Markov chain analysis, and fluid limit methods, providing actionable insights into system performance.
• DMR plays a pivotal role in optimizing scheduling policies and resource management in datacenters, wireless networks, and real-time embedded systems.

Deadline Miss Rate (DMR) quantifies the fraction of jobs, flows, packets, or transfers that fail to complete before their prescribed deadlines in scheduling, networking, and real-time systems. DMR is central to the analysis of systems with hard and soft timing constraints and appears in metrics, constraints, and optimization objectives across queueing theory, network scheduling, datacenter traffic management, wireless communications, and real-time embedded computing.

1. Formal Definitions and Variants

DMR is generally defined as the proportion of deadline-constrained tasks that exceed their individual deadlines. The standard mathematical form is:

$$\mathrm{DMR} = \frac{\text{number of jobs/flows that miss their deadline}}{\text{total number of deadline-constrained jobs/flows}}$$

Precise formalizations include:

• For flows with deadline $D_i$, arrival $A_i$, and flow completion time $FCT_i$ (Noormohammadpour et al., 2017):

$$\mathrm{DMR} = \frac{\bigl|\{i:\mathrm{FCT}_i > D_i\}\bigr|}{N_{\mathrm{dead}}}$$

• For slotted queueing systems, the empirical DMR across $N$ tasks:

$\mathrm{DMR}_N = \frac{1}{N}\sum_{k=1}^N \mathbf{1}\{\text{job %%%%4%%%% misses deadline}\}$

• Asymptotically for periodic or regenerative processes, DMR may be given by the stationary probability of deadline-failure events under the system’s equilibrium distribution (Chen et al., 2024).

Context-specific DMR definitions include:

• Packet drop rate in deadline-constrained wireless scheduling (Fountoulakis et al., 2018).
• Mis-delivery fraction in (soft or hard) real-time networks, coflows, or point-to-multipoint transfers, with possible distinction between DMR among admitted jobs and DMR among all arrivals (Luu et al., 2022).
• In some admission-control or resource reservation systems, DMR is implicitly defined by the fraction of traffic that cannot be admitted without risking deadline violation (Noormohammadpour et al., 2017).

2. DMR Computation and Measurement Methodologies

DMR measurement depends on the underlying system model:

• Queueing and Scheduling Simulation: For datacenter flows or LTE uplink scheduling, DMR is computed via trace- or simulation-based experiments: jobs or flows are generated, deadlines and completion times are tracked, and DMR is computed as the empirical fraction of deadline-missed events (Noormohammadpour et al., 2017, Haferkamp et al., 2016).
• Markov Chain Models: Analytical studies often utilize DTMCs to obtain the long-run fraction of jobs in deadline-missed states, e.g., by determining stationary distributions over deadline-hit and deadline-miss classes (Chen et al., 2024, Palopoli et al., 2016, Nomikos et al., 2022).
• Lyapunov Drift and Fluid Limit Techniques: In large-scale or stochastic systems, fluid limit approximations, Lyapunov drift analysis, or Skorokhod-mapped stochastic processes yield DMR as a function of system parameters (arrival, service distributions, patience/deadline laws) (Atar et al., 2013).
• Admission-controlled Systems: In systems guaranteeing zero deadline-misses among admitted traffic, system-wide DMR is indirectly inferred as

$$\mathrm{DMR_{total}} = 1 - \frac{\text{admitted volume}}{\text{offered volume}}$$

since all admitted requests are scheduled to complete before the deadline (Noormohammadpour et al., 2017).

Table: DMR Formulations in Representative Models

Context	DMR Formula	Source
Datacenter flows	$$\mathrm{DMR} = \frac{|\{i: FCT_i > D_i\}|}{N_\mathrm{dead}}$$	(Noormohammadpour et al., 2017)
Wireless packet queue	Long-run avg. per-slot drops	(Fountoulakis et al., 2018)
Markov chain (soft RT)	$\mathrm{DMR} = \sum_{s \in S_\mathrm{miss}} \pi_s$	(Chen et al., 2024)
Admission control (link)	DMR = $D_i$0 = 1 - $D_i$1	(Kreidl, 9 Sep 2025)

3. DMR in System Design and Policy Comparison

DMR is a key criterion for evaluating and comparing scheduling disciplines, admission policies, or resource allocation algorithms:

• Flow Scheduling in Datacenter Networks: Comparative studies show that policies such as Shortest Remaining Processing Time (SRPT) and deadline-aware variants of Earliest Deadline First (EDF) achieve the lowest DMR under load, while fair-sharing and FCFS result in markedly higher miss rates, particularly under heavy-tailed workloads or high utilization (Noormohammadpour et al., 2017).
• Wireless/Cellular Scheduling: Dynamic transmit-power control and real-time admission policies are often tuned to enforce hard DMR constraints (e.g., DMR$D_i$2 for some $D_i$3), with system states and drift-plus-penalty analyses ensuring these constraints are provably met (Ewaisha et al., 2016, Fountoulakis et al., 2018).
• Spatio-temporal DNN Scheduling: Modern GPU inference schedulers report DMR at per-priority levels (e.g., 0% for high priority, $D_i$42% for low priority as in DARIS (Babaei et al., 8 Apr 2025)), and configure task staging, admission, and oversubscription to empirically minimize observed miss rates.

DMR is often complementary to metrics such as lateness (for soft deadlines) or flow-completion time, and system policies are evaluated along the Pareto frontier between DMR and secondary objectives (throughput, lateness, utility).

4. DMR as an Optimization Constraint and Trade-off Parameter

Numerous optimization problems are posed with DMR as a primary constraint or objective:

• Constraint Enforcement: Real-time and deadline-driven wireless scheduling defines explicit DMR thresholds (e.g., each RT user must achieve on-time delivery $D_i$5 DMR $D_i$6), with algorithmic solutions guaranteeing feasibility (Ewaisha et al., 2016).
• Objective Function Component: Multi-class or profit-utility optimal controllers (e.g., admission control for elastic/inelastic loads) formulate an objective as a combination of negative DMR and utility accrued, with tunable weighting factors allowing explicit DMR vs. profit trade-offs (Kreidl, 9 Sep 2025).
• Admission and Resource Sizing: In hard-deadline traffic, the admission threshold or minimum reservation rate is set to ensure that scheduled jobs never violate deadlines, effecting DMR=0 on admitted traffic and relegating all DMR to rejected load (Noormohammadpour et al., 2017).

Policy tuning is frequently governed by parameters (weightings, minimum rates, buffer sizes) that permit the operator to trade increased DMR for improved secondary objectives, or vice versa.

5. Analytical Techniques for DMR Evaluation

Analytical derivation of DMR employs a diversity of tools matched to system structure:

• Markov Chains for Backlog Analysis: For periodic or reservation-based real-time models, DMR aligns with the (possibly infinite) DTMC’s stationary probability of exceeding backlog thresholds that yield deadline misses (Palopoli et al., 2016).
• Fluid Limit and Skorokhod Problems: For highly loaded G/G/1+G queues with abandonment or deadline-miss reneging, the DMR is characterized by Skorokhod maps with explicit closed-form solutions for the limiting (steady-state) fraction of missed deadlines (Atar et al., 2013). The steady-state DMR is $D_i$7 when the arrival rate $D_i$8 exceeds service rate $D_i$9.
• Probabilistic Deadline Failure (WCDFP): In fixed-priority schedulers under execution time randomness, the definition of DMR coincides with the supreme over busy-period release patterns of the worst-case miss probability (WCDFP), computed via convolutions of job execution distributions and busy-period analysis across finite busy-period starting points (Liu et al., 28 Aug 2025).

These models enable efficient numerical, sometimes closed-form, evaluation of equilibrium or worst-case DMR, with extensions to include soft deadlines, stochastic service, and periodic or bursty arrivals.

6. DMR in Diverse Application Contexts

DMR is reported across a spectrum of system scenarios:

• Datacenter and Network Scheduling: Per-flow DMR as a user- or flow-centered service quality metric, with DMR curves reported as a function of load, flow size distribution, and traffic mix (Noormohammadpour et al., 2017).
• Wireless/Uplink Scheduling: Average drop/DMR per user and system-wide, under channel fading, power constraints, and retransmissions (Ewaisha et al., 2016, Fountoulakis et al., 2018, Nomikos et al., 2022). Analyses account for both classic (collision) and MPR-capable random access networks.
• Real-Time Embedded and Cyber-Physical Systems: DMR per task under periodic scheduling, reservation-based servers, and probabilistic execution, with experimental, simulation, and analytic (Markov) support (Palopoli et al., 2016, Haferkamp et al., 2016, Chen et al., 2024).
• GPU Inference and Spatio-Temporal Resource Partitioning: Empirical DMR for high- and low-priority DNN inference jobs under multi-level scheduling and oversubscription, guiding task-admission and context migration policy (Babaei et al., 8 Apr 2025).
• Admission-Controlled and Resource-Isolated Systems: Indirect DMR, defined as one minus fraction of total offered traffic admitted under conservative greed maximization, when admitted jobs are guaranteed not to violate deadlines (Noormohammadpour et al., 2017).

7. Significance, Limitations, and Research Directions

DMR serves as a primary objectively quantifiable metric for timeliness in real-time, deadline-constrained, and service-level-oriented systems. It guides both policy comparison and system sizing. However, care must be taken in its application: in admission-controlled systems, observed DMR may be zero by construction among admitted traffic but nonzero for the total offered load, requiring careful accounting in system-level evaluations (Noormohammadpour et al., 2017). Soft-deadline models may require additional metrics such as average lateness or tail lateness to complement DMR, especially when late-completed jobs retain partial utility (Noormohammadpour et al., 2017).

Analytical modeling of DMR is challenging in the presence of correlated arrivals, burstiness, or complex service interactions, often necessitating approximation (Markov aggregation, fluid limits, probability bounds) and simulation for high-fidelity estimation. Recent advances address the worst-case probabilistic deadline miss (WCDFP), efficiently identifying busy-period intervals and producing safe and sometimes tight DMR bounds for fixed-priority preemptive scheduling with stochastic execution (Liu et al., 28 Aug 2025). These techniques provide actionable DMR guarantees in increasingly complex, service-level-driven settings.