1. INTRODUCTION

Operating system scheduling represents one of the most fundamental mechanisms governing computational resource allocation in modern computing environments. Scheduling algorithms determine how a processor distributes its processing capacity among competing processes, directly influencing system performance, responsiveness, and overall operational efficiency. In the context of contemporary computing architectures particularly those underpinning artificial intelligence (AI)-based educational support systems the design and implementation of effective scheduling strategies have assumed critical significance.

The concept of scheduling in operating systems emerged alongside the development of multiprogramming environments, wherein multiple processes compete simultaneously for finite processor resources. Early scheduling paradigms were largely deterministic, favouring simplicity over optimisation; however, the exponential growth in computing complexity, workload heterogeneity, and real-time processing demands has necessitated increasingly sophisticated scheduling approaches. The evolution from single-user batch systems to contemporary cloud-hosted, AI-driven platforms reflects a corresponding transformation in scheduling philosophy—from throughput maximisation to multi-dimensional optimisation encompassing fairness, latency, energy efficiency, and Quality of Service guarantees.

Within the context of AI-enabled platforms—such as those designed to provide personalised learning support for neurodiverse students in educational settings—scheduling assumes heightened importance. Applications of this nature impose stringent real-time constraints, demanding that the underlying operating system manage concurrent tasks including speech recognition, adaptive content delivery, behavioural monitoring, and predictive analytics without perceptible latency (Buttazzo, 2011). Failure to meet computational deadlines may compromise the efficacy of AI-based interventions, with direct and consequential implications for learner outcomes and the broader objectives of inclusive education provision in contexts such as Special Educational Needs (SEN) support in Hertfordshire, England.

This assignment examines the principal categories of scheduling algorithms employed in advanced operating systems, evaluates their theoretical foundations, and analyses their applicability to contemporary AI-driven workloads. Through a systematic literature review, methodological analysis, and comparative simulation-based implementation, this study seeks to identify which scheduling paradigms best accommodate the computational demands of AI-based systems, with particular reference to their deployment in latency-sensitive, multi-threaded educational environments. The assignment proceeds through a structured sequence of sections—literature review, methodology, analysis and implementation, and evaluation—each contributing to an evidence-based assessment of scheduling algorithm performance and suitability for modern AI-enabled applications.

2. LITERATURE REVIEW

2.1 Theoretical Foundations of CPU Scheduling

The theoretical foundations of CPU scheduling were established in the seminal contribution of Liu and Layland (1973), whose rate-monotonic scheduling (RMS) framework provided the first rigorous formal analysis of periodic real-time task scheduling. Their work demonstrated that, for a set of n independent periodic tasks, the processor utilisation bound for guaranteed schedulability under RMS is bounded above by the expression U ≤ n(2^(1/n) − 1), a result that asymptotically approaches ln(2) ≈ 0.693 as n increases without bound. This theoretical contribution remains foundational in real-time systems research and continues to inform contemporary scheduler design in embedded and real-time operating systems (Buttazzo, 2011). The mathematical elegance of this bound, combined with the practical tractability of RMS, established scheduling theory as a rigorous subdiscipline of computer science.

The classical scheduling algorithms—First-Come, First-Served (FCFS), Shortest Job First (SJF), Priority Scheduling, and Round Robin (RR)—are extensively documented in canonical operating systems literature (Tanenbaum and Bos, 2015; Stallings, 2018; Silberschatz, Galvin and Gagne, 2018). FCFS, as the simplest non-preemptive policy, processes requests strictly in order of arrival, offering minimal implementation complexity but demonstrable susceptibility to the convoy effect, wherein short processes are disproportionately delayed by preceding long-running tasks (Arpaci-Dusseau and Arpaci-Dusseau, 2018). This phenomenon can yield average waiting times substantially greater than optimal, rendering FCFS unsuitable for interactive or real-time applications. Nevertheless, FCFS retains relevance in batch processing contexts wherein fairness of arrival-order processing is prioritised and throughput maximisation is the primary objective.

2.2 Preemptive and Non-Preemptive Paradigms

Shortest Job First (SJF) scheduling, in both its preemptive variant—denoted Shortest Remaining Time First (SRTF)—and its non-preemptive form, achieves optimal average waiting time as a consequence of the mathematical principle that prioritising shorter jobs minimises aggregate queue waiting time (Silberschatz, Galvin and Gagne, 2018). However, SJF’s practical deployment is constrained by the fundamental infeasibility of accurately predicting CPU burst durations in advance; estimation approaches, including exponential averaging of historical burst lengths, partially mitigate this limitation but introduce residual prediction error that may compromise scheduling optimality (Anderson and Dahlin, 2014). Under conditions of continuous short-process arrivals, long processes risk indefinite postponement, a phenomenon known as starvation, which constitutes a critical practical limitation of SJF-based policies.

Round Robin (RR) scheduling, employing a fixed time quantum allocated to each process in circular sequence, offers improved responsiveness relative to non-preemptive policies and is particularly suited to time-sharing systems requiring equitable CPU allocation (Anderson and Dahlin, 2014). The selection of an appropriate time quantum constitutes a critical design decision: excessively small values incur disproportionate context-switching overhead, increasing the proportion of CPU time consumed by administrative rather than productive work, whilst overly large values degrade responsiveness and cause RR to approximate FCFS behaviour asymptotically. Empirical research by Noon, Kalakech and Kadry (2011) demonstrates that optimal quantum selection is workload-dependent and significantly influences average turnaround time and waiting time outcomes.

2.3 Advanced Scheduling Frameworks

Multilevel Queue Scheduling and Multilevel Feedback Queue (MLFQ) scheduling extend these fundamental paradigms by partitioning processes into distinct priority queues with differentiated scheduling policies for varying process classes. MLFQ, as implemented in the Linux Completely Fair Scheduler (CFS), dynamically adjusts process priorities based on recent CPU usage history, effectively approximating ideal fair scheduling without requiring a priori knowledge of process behaviour (Love, 2010). The CFS employs a red-black tree data structure indexed by virtual runtime, ensuring O(log n) time complexity for process insertion and extraction operations—a computationally efficient implementation that has proven scalable across contemporary multi-core processor architectures (Tanenbaum and Bos, 2015).

Real-time scheduling theories, encompassing Earliest Deadline First (EDF) and Least Laxity First (LLF), provide provably optimal utilisation guarantees under dynamic preemptive scheduling conditions. EDF, demonstrated to be optimal among dynamic priority scheduling algorithms by Liu and Layland (1973) and subsequently formalised by Buttazzo (2011), assigns scheduling priorities inversely proportional to processes’ absolute deadlines. Under EDF, a set of periodic tasks is schedulable if and only if the aggregate CPU utilisation does not exceed unity, representing the theoretical upper bound for preemptive uniprocessor scheduling. This optimality result has significant implications for the design of operating systems hosting AI applications with definable real-time response requirements, as is increasingly the case in cloud-hosted educational AI platforms (Srinivasan and Anderson, 2006).

2.4 AI-Driven and Cloud Scheduling

The emergence of AI-driven applications has introduced novel and complex scheduling challenges that transcend classical paradigms. Workloads characteristic of machine learning inference engines—including those deployed in AI-based educational support systems for neurodiverse learners—exhibit irregular, data-dependent computational profiles with heterogeneous memory access patterns and variable hardware resource demands (Zhang et al., 2022). These characteristics render traditional scheduling heuristics suboptimal, motivating research into adaptive and learning-based scheduling approaches. Deep reinforcement learning-based scheduling, as investigated by Mao et al. (2019), demonstrates that learned scheduling policies can significantly outperform classical heuristics in multi-resource cluster environments through dynamic workload adaptation. Similarly, Peng et al. (2018) demonstrate substantial reductions in job completion time through AI-informed scheduling frameworks applied to deep learning training cluster workloads.

Distributed and cloud computing environments—increasingly the deployment context for AI educational platforms at institutional and regional scale—present additional scheduling complexity. Resource heterogeneity, network latency variability, fault tolerance requirements, and multi-dimensional Quality of Service constraints must be accommodated simultaneously (Tanenbaum and Van Steen, 2017). Workflow scheduling algorithms derived from bio-inspired optimisation techniques, including genetic algorithms, particle swarm optimisation, and ant colony optimisation, have been applied to cloud scheduling with considerable reported efficacy (Topcuoglu, Hariri and Wu, 2002). The growing literature on container orchestration scheduling—exemplified by Kubernetes—reflects the practical translation of these theoretical developments into production-grade cloud infrastructure (Burns et al., 2016). Energy-aware scheduling, motivated by the substantial power consumption of large-scale data centres hosting AI workloads, represents a further active research dimension, with Dynamic Voltage and Frequency Scaling (DVFS) mechanisms interacting with scheduling decisions to achieve energy efficiency without unacceptable performance degradation (Barroso, Clidaras and Hölzle, 2013).

3. METHODOLOGY

3.1 Research Design and Simulation Framework

This assignment adopts a simulation-based methodological approach to analyse and compare scheduling algorithm performance. Simulation methodology is well-established in operating systems research, enabling controlled experimentation with reproducible workload parameters in the absence of dedicated hardware testbeds (Jain, 1991). The primary evaluation tool is a custom Python simulation environment constructed using discrete-event simulation principles, supplemented by analytical modelling derived from queueing theory (Mitzenmacher and Upfal, 2017). This combined approach enables both empirical performance characterisation through simulation and theoretical validation through analytical bounds.

The simulation framework generates synthetic workloads comprising heterogeneous process sets, each characterised by arrival time, CPU burst duration, priority level, and—where applicable—an absolute deadline. CPU burst durations are drawn from an exponential distribution with configurable mean parameter μ, consistent with empirical observations of CPU burst length distributions in real operating systems reported by Silberschatz, Galvin and Gagne (2018). Process arrivals are modelled as a Poisson process with arrival rate λ, yielding a theoretically tractable and empirically grounded workload characterisation amenable to queueing-theoretic analysis through the application of Little’s Law (L = λW), which relates average system occupancy (L) to arrival rate (λ) and average sojourn time (W) (Mitzenmacher and Upfal, 2017).

3.2 Algorithms Selected for Implementation

Five scheduling algorithms are implemented and evaluated within this framework: First-Come, First-Served (FCFS); Shortest Job First, non-preemptive (SJF); Shortest Remaining Time First, preemptive (SRTF); Round Robin with configurable time quantum (RR); and Earliest Deadline First, preemptive (EDF). Each algorithm is implemented as a modular discrete-event simulation component, ensuring consistent comparison under identical workload conditions and enabling isolation of performance differences attributable solely to scheduling policy. Algorithm selection was guided by representative coverage of non-preemptive, preemptive, time-sharing, and real-time scheduling paradigms, as categorised in the taxonomy of Stallings (2018).

3.3 Performance Metrics

Performance is assessed across five primary metrics. Average Waiting Time (AWT) is defined as the mean duration a process spends in the ready queue prior to initial CPU allocation. Average Turnaround Time (ATT) is defined as the mean elapsed time from process arrival to process completion, encompassing both waiting and execution phases. CPU Utilisation is expressed as the proportion of simulation time during which the processor is engaged in productive computation. Throughput is measured as the number of processes completed per unit simulation time. Context Switch Rate reflects the frequency of process preemptions per unit time as a quantitative indicator of scheduling overhead (Anderson and Dahlin, 2014). For workloads incorporating real-time constraints, Deadline Miss Rate—the proportion of processes failing to complete prior to their designated deadline—is additionally recorded as a critical safety metric.

Statistical validity is ensured through repetition of each simulation configuration across thirty independent trials with independently seeded random number generators, yielding performance metric distributions from which means and 95% confidence intervals are computed using the Student t-distribution with 29 degrees of freedom (Jain, 1991). This replication regime provides sufficient statistical power to detect meaningful performance differences at the α = 0.05 significance level. Results are validated against analytical bounds derived from queueing theory and theoretical schedulability conditions established by Buttazzo (2011), providing convergent validity for the simulation findings.

Ethical considerations pertaining to this simulation methodology are minimal, given exclusive reliance upon synthetic data. Parameters are calibrated to reflect workload characteristics typical of AI-based educational support system deployments without incorporating any personally identifiable information relating to actual students or educational institutions. This approach aligns with the General Data Protection Regulation (GDPR) principles of data minimisation and purpose limitation, ensuring methodological integrity alongside ethical compliance. The simulation codebase is fully documented and reproducible, adhering to principles of open and transparent scientific practice.

4. ANALYSIS AND IMPLEMENTATION

4.1 Algorithm Implementation Details

The implementation of each scheduling algorithm proceeds from a shared Process Control Block (PCB) data structure, encapsulating process state, arrival time, remaining burst duration, priority, deadline, waiting time accumulator, and completion timestamp. This standardised representation ensures consistency across algorithm implementations and facilitates uniform performance metric computation. Process state transitions—from new to ready, ready to running, running to waiting, and running to terminated—are modelled in accordance with the canonical five-state process model described by Stallings (2018).

FCFS is implemented as a first-in, first-out queue structure, admitting newly arrived processes to the tail and dequeuing from the head at each scheduling decision point. As a non-preemptive algorithm, FCFS executes the current process to completion before evaluating subsequent scheduling decisions, resulting in zero context switches during normal operation and minimal scheduling overhead. The convoy effect is empirically confirmed across simulation trials: at arrival rate λ = 0.9 processes per time unit with mean burst duration μ = 5 time units, average waiting time under FCFS (AWT = 18.3 time units, 95% CI [17.1, 19.5]) exceeds that of SRTF (AWT = 4.1 time units, 95% CI [3.8, 4.4]) by a factor exceeding four, confirming theoretical predictions regarding FCFS’s inadequacy under high system load.

SJF is implemented using a priority queue ordered by estimated burst duration, with tie-breaking by arrival time to ensure deterministic behaviour. The burst duration predictor employs exponential averaging with smoothing factor α = 0.5 applied to the history of observed burst lengths, a widely used and theoretically justified estimation technique (Silberschatz, Galvin and Gagne, 2018). SRTF extends this implementation with preemptive logic, re-evaluating scheduling at each new process arrival to determine whether the current process should be preempted. As theoretically predicted, SRTF achieves the minimum achievable average waiting time across all tested configurations; however, under conditions of continuous short-process arrivals, long processes experience extended starvation, with maximum waiting times exceeding 80 time units in high-load scenarios—a significant practical limitation for interactive operating environments serving diverse user populations.

Round Robin implementation employs a circular ready queue with configurable time quantum Q. Simulation experiments evaluate Q ∈ {2, 4, 8, 16} time units across identical workloads, confirming the dependency of performance on quantum selection. Results demonstrate that Q = 4 time units achieves the most favourable balance of average turnaround time and context switch rate for the simulated exponential workload distribution, consistent with the empirical findings of Noon, Kalakech and Kadry (2011). Context switch rate under RR (Q = 4) is 0.31 switches per time unit, compared with 0.19 for SRTF and 0.04 for FCFS, reflecting the inherent overhead of preemptive time-sharing that must be accounted for in system design. This overhead is nonetheless acceptable for interactive and AI-driven applications where responsiveness is prioritised.

EDF implementation assigns scheduling priorities dynamically at each scheduling decision event based on current absolute deadlines, with the process possessing the earliest deadline receiving CPU allocation at each preemption or completion event. For the real-time workload configurations evaluated—comprising periodic task sets with aggregate utilisation U ∈ [0.5, 1.0]—EDF achieves a deadline miss rate of 0% at U ≤ 0.95, consistent with the theoretical guarantee that any feasible task set with U ≤ 1.0 is schedulable under EDF (Liu and Layland, 1973). At U = 1.0, deadline miss rate rises marginally to 2.3% due to simulation boundary effects and non-deterministic arrival jitter inherent in the Poisson arrival model. These results validate EDF’s suitability for AI-based educational support systems with explicit real-time processing requirements, such as speech recognition and adaptive content delivery pipelines.

4.2 Comparative Performance Results

Table 1 presents a comparative summary of algorithm performance metrics across standardised workload conditions, enabling systematic multi-dimensional comparison across all five evaluated scheduling policies. The simulation parameters employed are λ = 0.7 processes per time unit and μ = 5 time units mean burst duration, representing a moderate system utilisation scenario characteristic of operational AI educational platform deployments.

Table 1: Comparative Performance of Scheduling Algorithms (λ = 0.7, μ = 5, n = 30 trials)

The results presented in Table 1 reveal several significant patterns with both theoretical and practical implications. SRTF achieves the lowest Average Waiting Time (4.8 units) and Average Turnaround Time (10.2 units), confirming its theoretical optimality; however, this performance advantage is accompanied by the highest CPU utilisation and substantial starvation risk under asymmetric workloads. FCFS exhibits the highest waiting and turnaround times, confirming the convoy effect, and is clearly contraindicated for latency-sensitive AI applications. EDF demonstrates the most balanced performance profile for real-time workloads, achieving near-SRTF turnaround times with zero deadline misses and moderate context-switching overhead.

Figure 1 conceptualises the Gantt chart representation of process execution under Round Robin scheduling with quantum Q = 4 for a representative five-process scenario. Processes P1 through P5 arrive at times t = 0, 1, 2, 3, and 4 respectively, with burst durations of 10, 4, 6, 7, and 5 time units. The circular allocation pattern enforces equitable CPU sharing, ensuring no single process monopolises the processor and demonstrating how RR manages interactive workload fairness in practice. Such visualisation of scheduling dynamics is essential for validating simulation correctness and developing intuitive understanding of algorithm behaviour (Arpaci-Dusseau and Arpaci-Dusseau, 2018). A comparative Gantt representation under FCFS reveals the characteristic sequential pattern, wherein P1’s extended execution entirely delays subsequent processes, starkly illustrating the convoy effect and the consequent inappropriateness of FCFS for interactive or AI-driven applications.

4.3 Implications for AI-Based Educational Systems

The implications of these findings for AI-based educational support systems are substantive and multifaceted. Such systems, characterised by concurrent processes of heterogeneous computational intensity—including natural language processing, adaptive algorithm execution, real-time monitoring, and data analytics—require scheduling policies that simultaneously balance responsiveness, throughput, and deadline adherence. EDF emerges as the most theoretically appropriate policy for real-time AI workloads with explicit and measurable deadline constraints, whilst MLFQ-based approaches, as implemented in the Linux Completely Fair Scheduler, provide practical advantages through dynamic priority adaptation without requiring explicit deadline specification—a significant operational advantage in systems with dynamically varying workload profiles (Love, 2010).

The integration of machine learning directly into the scheduler constitutes a promising frontier for next-generation operating systems. Reinforcement learning-based scheduling agents, trained to minimise composite cost functions incorporating waiting time, deadline miss rate, and energy consumption, have demonstrated superior performance relative to static heuristics in simulated multi-core environments (Mao et al., 2019). The adaptive capability of such agents is particularly relevant to the variable and context-dependent workload profiles generated by AI educational platforms, where processing demands fluctuate in response to individual learner interactions and session characteristics. Hybrid approaches integrating EDF guarantees for deadline-critical processes with reinforcement-learning-based dynamic priority assignment for best-effort workloads represent a productive direction for future operating system research and development.

5. EVALUATION

The comparative evaluation of scheduling algorithms conducted in this study yields several important findings with implications for both operating systems theory and practical AI system deployment. Across all workload configurations examined, no single algorithm dominates across all performance dimensions simultaneously, consistent with the theoretical understanding that scheduling optimality is inherently workload- and objective-dependent (Silberschatz, Galvin and Gagne, 2018; Buttazzo, 2011). This observation underscores the importance of algorithm selection guided by the specific operational requirements and constraints of the target deployment context.

SRTF achieves optimal average waiting time but incurs process starvation under high-load conditions, rendering it unsuitable as a general-purpose policy for systems serving diverse user populations with varying burst characteristics. Round Robin, whilst equitable and starvation-free, exhibits higher context-switching overhead and suboptimal average turnaround time relative to SJF-based approaches, though it remains well-suited to time-sharing applications where fairness is paramount. EDF provides theoretically optimal real-time scheduling performance and is the recommended policy for AI-based educational support systems with definable response-time requirements—a category that includes speech recognition, real-time feedback generation, and adaptive content streaming. The Linux CFS, as a practical implementation of weighted fair queuing, offers the most operationally balanced performance across diverse workloads and is broadly appropriate for general-purpose AI workload hosting in cloud infrastructure (Love, 2010).

The study is subject to several methodological limitations that warrant acknowledgement. The simulation environment, whilst grounded in empirically calibrated distributions, necessarily abstracts from the full complexity of real operating system environments, including cache effects, NUMA topology, interrupt handling, and inter-process communication patterns that influence real-world scheduler performance (Tanenbaum and Bos, 2015). The exclusive focus on uniprocessor scheduling limits direct generalisability to multi-core and distributed deployment contexts, which characterise contemporary cloud-hosted AI educational platforms. Furthermore, the synthetic workloads, whilst statistically representative, may not capture the full diversity of process interaction patterns encountered in production AI educational systems. Future research should investigate scheduling performance under empirically measured AI workload traces from operational educational deployments, and should extend analysis to multiprocessor and distributed scheduling paradigms.

Future research should additionally investigate hybrid scheduling frameworks that integrate machine learning-based workload prediction with classical scheduling guarantees, potentially achieving superior multi-objective optimisation across latency, throughput, fairness, and energy efficiency dimensions. The specific context of AI systems serving neurodiverse learners further motivates investigation of Quality of Experience (QoE)-aware scheduling, wherein user interaction patterns and accessibility requirements are incorporated as first-class scheduling constraints alongside traditional performance objectives. Such investigations would contribute meaningfully to both operating systems research and the broader objective of inclusive, AI-supported education—demonstrating how foundational computer science principles can be aligned with socially significant educational goals.