Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

599 
https://internationalpubls.com 

Design of an Iterative Method for MapReduce Scheduling Using Deep 

Reinforcement Learning and Anomaly Detection 

 

Mr. Aihtesham Kazi1, Dr. Dinesh Chaudhari2 

1Bajaj Institute of Technology, Department of Computer Engineering, Wardha India. 

 E-mail: kazi_aihtesham@rediffmail.com. 

2Jawaharlal Darda Institute of Engineering and Technology, Department of Computer Science & Engineering, Yavatmal 

India. E-mail: dinesh_chaudhari@jdiet.ac.in 

 

Article History: 

Received: 23-04-2024 

Revised: 24-06-2024 

Accepted: 28-06-2024 

Abstract:  

Due to high complexities within distributed computing environments, there's a 

critical need for advanced scheduling frameworks that are capable of optimizing 

MapReduce systems. Current approaches have static policies that limit their 

capability to adapt to changing system dynamics and workload variations for 

different cloud scenarios. To overcome these issues, this study introduces a 

robust MapReduce framework empowered by intelligent scheduling 

algorithms, tailored to enhance system efficiency and resilience levels. The 

framework introduces three novel scheduling models: Deep Reinforcement 

Learning for Dynamic Job Scheduling (DRDJS), Anomaly Detection-driven 

Adaptive Scheduling (ADAS), and Cluster-based Job Categorization and 

Scheduling (CJCS). DRDJS utilizes deep reinforcement learning to 

dynamically generate optimal scheduling policies based on multiple metrics that 

include historical job data, system metrics, and workload characteristics. This 

adaptive approach leads to significant reductions in job completion times, 

outperforming static scheduling methods by up to 30%. Next, ADAS leverages 

anomaly detection to prioritize critical tasks and efficiently allocate resources 

in response to anomalies, resulting in a notable improve scheduling speed by up 

to 40%. Finally, CJCS employs clustering algorithms to categorize jobs based 

on their resource requirements and execution profiles, enabling more accurate 

resource allocation and reducing average job completion times by up to 25%. 

By integrating these methods into a unified scheduling framework, the proposed 

solution addresses the variability in job characteristics and system performance, 

thereby enhancing the overall throughput and stability of MapReduce 

operations.  

Keywords: MapReduce, Deep Reinforcement Learning, Anomaly Detection, Resource 

Allocation, Distributed Computing 

1. Introduction 

In the field of distributed computing, the management of computational resources on a large-scale 

network is of prime importance. The architecture of MapReduce has emerged as a very important 

architecture for processing huge amounts of data by breaking the tasks into small fragments, which are 

processed in a parallel mode over a distributed network. Despite its wide use, the standard scheduling 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

600 
https://internationalpubls.com 

mechanisms employed by MapReduce are often not dynamic enough to adjust according to the 

changing conditions of the environment and workload demands, thus rendering many inefficiencies in 

the utilization of resources and processing delays. Recent advances in machine learning and data 

analytics have opened new avenues to enhance the adaptability and efficiency of job scheduling in 

distributed systems. More precisely, the integration of intelligent scheduling algorithms with the 

MapReduce framework will dramatically improve the drawbacks of conventional static scheduling 

methods. This paper presents a sophisticated scheduling framework that uses Deep Reinforcement 

Learning for Dynamic Job Scheduling (DRDJS) and Anomaly Detection-driven Adaptive Scheduling 

(ADAS) to dynamically adjust the allocation of resources in real timestamp with regard to system 

performance and workload characteristics. 

The need for this is that existing static scheduling algorithms have inherent limitations designed under 

conditions of predictable and uniform workloads, conditions that seldom prevail in a real-world 

scenario. These conventional methods tend not to respond effectively to the spontaneous variations in 

the characteristics of the jobs and the system status, which often lead to suboptimal distribution of 

resources and protracted delays in the completion of jobs. To make up for these challenges, our 

proposed framework includes a DRDJS that utilizes deep reinforcement learning to learn continuously 

from system states and job execution feedback toward optimizing scheduling decisions in real-time. 

This method improves not only the efficiency of resource allocation but also the reduction of the 

completion timestamp of various jobs, hence providing better system throughput. Complementing this 

approach, ADAS applies anomaly detection techniques to real-time system logs and performance 

metrics for identifying and prioritizing critical tasks requiring immediate attention. This adaptive 

method ensures that system performance does not degrade when unexpected deviations happen; that 

is, it stabilizes the environment of processing and reduces potential delays. 

More importantly, the integration of Cluster-based Job Categorization and Scheduling (CJCS) within 

our framework makes use of clustering algorithms to group jobs with similar characteristics. Such 

categorization aids the creation of tailored scheduling policies optimized for specific clusters and 

thereby enhances the precision of resource allocation, further reducing processing delays. The 

convergence of such path-breaking innovative scheduling methods within one integrated framework 

marks a great stride toward the achievement of a resilient and efficient distributed computing 

environment. This introduction gives way to the in-depth discussion of the methodologies, 

implementation, and impactful results of the proposed scheduling framework that, when considered 

collectively, enhance the robustness and adaptability of the MapReduce architectures. 

Motivation & Contribution 

The exponential growth in data generation and the ever-increasing complexity of computational tasks 

have significantly pressured traditional data processing frameworks, particularly in the context of 

distributed computing environments like MapReduce. Although MapReduce has been instrumental in 

handling large-scale data processing by decomposing tasks into smaller, manageable batches executed 

across various nodes, its efficiency heavily relies on the underlying job scheduling mechanisms. 

Traditional job scheduling strategies typically adopt a static approach, predetermining job execution 

without considering the dynamic nature of the runtime environment or workload variability. Such 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

601 
https://internationalpubls.com 

static scheduling often results in inefficient resource utilization, longer job completion times, and 

increased latency, particularly under fluctuating network conditions and diverse job requirements. This 

scenario underscores a critical need for a more adaptive scheduling framework that can intelligently 

respond to real-time changes in the system state and workload characteristics, thereby optimizing 

resource allocation, reducing processing times, and improving overall system throughput. 

In response to these challenges, this paper contributes a robust scheduling framework specifically 

designed for MapReduce environments, which integrates advanced machine learning techniques to 

enhance the adaptability and efficiency of job scheduling. The core of our contribution lies in the 

development and implementation of three innovative scheduling strategies: Deep Reinforcement 

Learning for Dynamic Job Scheduling (DRDJS), Anomaly Detection-driven Adaptive Scheduling 

(ADAS), and Cluster-based Job Categorization and Scheduling (CJCS). DRDJS leverages deep 

reinforcement learning algorithms to form a self-learning scheduling system that dynamically adapts 

its strategies based on continuous feedback from the system’s performance and the job execution 

outcomes. This method significantly reduces job completion times by constantly optimizing the 

decision-making process in real-time, thereby ensuring that the scheduling decisions are always 

aligned with the current state of the system and job queue characteristics. Meanwhile, ADAS 

introduces anomaly detection to the scheduling process, enabling the system to proactively adjust its 

resource allocation strategies in response to anomalies detected in system performance or job execution 

patterns. This proactive approach not only mitigates the impact of unexpected system behaviors but 

also ensures the prioritization of critical jobs, thereby maintaining system stability and reducing the 

potential for job slowdowns. Lastly, CJCS utilizes clustering algorithms to analyze and categorize jobs 

based on their resource demands and execution characteristics. By grouping similar jobs into clusters 

and developing optimized scheduling policies for each cluster, this method facilitates more precise 

resource allocation, reducing resource wastage and enhancing the efficiency of job processing across 

the distributed network. 

The integration of these three methods forms a comprehensive, intelligent framework that not only 

addresses the limitations of traditional static scheduling but also sets a new standard for job scheduling 

in distributed computing environments. The proposed framework's ability to adapt to changing 

conditions and its proactive approach to job and resource management significantly improve the 

performance and reliability of MapReduce operations. Empirical results obtained from extensive 

simulations and real-world deployments indicate that our framework can reduce job completion times 

and resource consumption significantly compared to existing scheduling approaches, thereby 

validating the effectiveness of our contributions. This work not only provides a practical solution to 

the challenges faced in distributed computing job scheduling but also opens new avenues for research 

in intelligent, adaptive systems that can revolutionize the efficiency of large-scale data processing 

architectures. Through the detailed exposition of each method's development and integration, this 

paper aims to inspire further innovations in the field and provide a scalable, efficient solution adaptable 

to various distributed computing scenarios. 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

602 
https://internationalpubls.com 

2. Literature Review 

The landscape of MapReduce task scheduling has seen significant advancements in recent years, 

driven by the ever-increasing demand for efficient processing of big data in various computing 

environments. In this section, we discuss papers that contribute to this domain, exploring the methods 

used, findings, results, and limitations of each study process. The primary objective of MapReduce 

task scheduling is to optimize resource utilization, reduce job completion times, and enhance overall 

system performance. Achieving these goals requires innovative approaches that address the unique 

challenges posed by heterogeneous environments, dynamic workloads, and diverse application 

requirements. From traditional heuristic algorithms to cutting-edge machine learning techniques, 

researchers have explored a wide range of methodologies to tackle these challenges. 

Reference Method Used Findings Results Limitations 

[1] Pair Jobs for 

Optimization 

Joint scheduling of 

overlapping MapReduce 

phases 

Improved 

makespan 

optimization 

Limited to map and 

shuffle phases 

[2] Hungarian Algorithm MapReduce task scheduling 

in heterogeneous geo-

distributed data centers 

Enhanced total 

tardiness 

Assumes fixed data 

center locations 

[3] Ant Colony 

Optimization 

Optimizing MapReduce task 

scheduling on virtualized 

heterogeneous environments 

Improved resource 

utilization 

Might face scalability 

issues 

[4] Trust-Aware 

Framework 

Trust-based scheduling 

framework for big data 

processing with MapReduce 

Enhanced big data 

security 

Assumes trust levels 

can be accurately 

determined 

[5] Topology-Aware 

Assignment 

Joint optimization of 

MapReduce scheduling and 

network policy in 

hierarchical data centers 

Improved network 

policy integration 

Limited to hierarchical 

data centers 

[6] New Scheduling 

Algorithms 

Scheduling algorithms for 

improving performance and 

resource utilization in 

Hadoop YARN clusters 

Enhanced resource 

management 

Scalability concerns 

with large clusters 

[7] Approximation 

Algorithm 

Regularization-based coflow 

scheduling in optical circuit 

switches 

Improved coflow 

scheduling 

Limited to optical 

circuit switches 

[8] Heuristic Algorithms Cost-efficient workflow 

scheduling algorithm for 

applications with deadline 

constraints on heterogeneous 

clouds 

Enhanced schedule 

length 

Might not handle 

dynamic workload 

changes efficiently 

[9] Convex Optimization Coflow scheduling in data 

centers: routing and 

bandwidth allocation 

Optimized coflow 

scheduling 

Limited to convex 

optimization scenarios 

[10] Divisible Task 

Scheduling 

Scheduling-guided 

automatic processing of 

massive hyperspectral image 

classification on cloud 

computing architectures 

Improved 

hyperspectral image 

classification 

Might face challenges 

with large-scale image 

datasets 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

603 
https://internationalpubls.com 

[11] Sampling-Based 

Learning 

A case for sampling-based 

learning techniques in 

coflow scheduling 

Enhanced 

scalability 

Assumes accurate 

sampling for learning 

[12] Heuristic Energy utilization task 

scheduling for MapReduce 

in heterogeneous clusters 

Improved energy 

consumption 

Might not handle 

fluctuating energy 

demands efficiently 

[13] Flexible Heuristic A flexible heuristic to 

schedule distributed analytic 

applications in compute 

clusters 

Enhanced 

scheduling 

flexibility 

Might not scale well 

with large cluster sizes 

[14] Optimization Joint reducer placement and 

coflow bandwidth 

scheduling for computing 

clusters 

Improved 

computing cluster 

optimization 

Limited to specific 

cluster configurations 

[15] GPGPU AEML: an acceleration 

engine for multi-GPU load-

balancing in distributed 

heterogeneous environment 

Enhanced load 

balancing 

Might face challenges 

with heterogeneous 

GPU configurations 

[16] Replication-Based 

Query Management 

Replication-based query 

management for resource 

allocation using Hadoop and 

MapReduce over big data 

Improved resource 

allocation 

Limited to specific big 

data scenarios 

[17] Job Scheduler Cooperative job scheduling 

and data allocation in data-

intensive parallel computing 

clusters 

Enhanced data 

locality 

Limited to parallel 

computing clusters 

[18] Task Scheduling Shadow: exploiting the 

power of choice for efficient 

shuffling in MapReduce 

Improved shuffling 

efficiency 

Might not handle 

highly dynamic 

environments 

efficiently 

[19] Bottleneck-Aware 

Scheduling 

Bottleneck-aware non-

clairvoyant coflow 

scheduling with FAI 

Enhanced coflow 

completion time 

Limited to certain 

network topologies 

[20] Data Affinity Task scheduling for Spark 

applications with data 

affinity on heterogeneous 

clusters 

Improved task 

scheduling 

efficiency 

Assumes accurate data 

affinity information 

[21] Dynamic Scheduling Optimizing internal overlaps 

by self-adjusting resource 

allocation in multi-stage 

computing systems 

Reduced makespan Might not handle 

highly dynamic 

workloads efficiently 

[22] Online Identification Scheduling coflows by 

online identification in data 

center network 

Improved coflow 

scheduling 

adaptability 

Limited to specific 

network architectures 

[23] Green Parallel 

Computing 

Towards greening 

MapReduce clusters 

considering both 

computation energy and 

cooling energy 

Enhanced energy 

efficiency 

Limited to specific 

cluster cooling 

infrastructures 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

604 
https://internationalpubls.com 

[24] Complexity-Aware 

Scheduling 

A complexity aware 

scheduler with dynamic slot 

allocation for cloud video 

transcoding 

Improved load 

balance operation 

Might not handle 

highly dynamic 

transcoding demands 

efficiently 

[25] Ant Colony 

Optimization 

Handling non-local 

executions to improve 

MapReduce performance 

using ant colony 

optimization 

Enhanced 

MapReduce 

performance 

Limited to specific 

execution scenarios 

Table 1. Review of Existing Methods used for Map Reduce based Task Scheduling Operations 

Table 1 reveals a diverse landscape of MapReduce task scheduling techniques, each offering unique 

insights and contributions to the field. One common theme among these studies is the emphasis on 

optimizing resource allocation and job scheduling to improve system efficiency and performance. 

However, the methodologies employed vary significantly, reflecting the complex nature of the task 

scheduling problem and the diversity of application scenarios. 

Several papers focus on leveraging optimization algorithms such as ant colony optimization, 

Hungarian algorithm, and convex optimization to achieve optimal task scheduling and resource 

utilization. For example, [3] proposes an ant colony optimization approach to optimize MapReduce 

task scheduling in virtualized heterogeneous environments, demonstrating improved resource 

utilization compared to traditional scheduling methods. Similarly, [9] utilizes convex optimization 

techniques for co-flow scheduling in data centers, effectively balancing bandwidth allocation and 

minimizing co-flow completion timestamp levels. 

In addition to optimization algorithms, heuristic approaches play a crucial role in addressing the 

practical challenges of task scheduling in MapReduce environments. Papers such as [8] and [13] 

propose flexible heuristic algorithms to schedule distributed analytic applications and workflow tasks 

in heterogeneous clouds, respectively. These heuristic methods offer practical solutions for real-world 

scenarios where optimal solutions may be computationally expensive or impractical to achieve. 

Furthermore, machine learning techniques have emerged as powerful tools for improving MapReduce 

task scheduling. Papers such as [11] advocate for the use of sampling-based learning techniques to 

predict coflow sizes and optimize flow scheduling, demonstrating enhanced scalability and 

performance compared to traditional methods. Similarly, [15] introduces AEML, an acceleration 

engine for multi-GPU load balancing, which leverages machine learning algorithms to dynamically 

balance workloads across heterogeneous environments, leading to improved system efficiency. 

Despite the significant advancements highlighted in these studies, several challenges and limitations 

remain. One recurring limitation is the scalability of proposed solutions, particularly in large-scale 

distributed environments with dynamic workloads. While many papers demonstrate promising results 

in controlled experimental settings, scaling these solutions to production-level systems with thousands 

of nodes and complex application requirements remains a significant challenge. 

Another common limitation is the assumption of idealized conditions or fixed system parameters, 

which may not hold true in real-world deployments. For example, [2] assumes fixed data center 

locations in its scheduling algorithm, overlooking the dynamic nature of geo-distributed environments. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

605 
https://internationalpubls.com 

Similarly, [18] proposes a shuffling optimization technique based on the assumption of homogeneous 

task distributions, which may not accurately reflect the variability of real-world workloads. 

Moreover, the trade-off between optimization objectives, such as resource utilization, job completion 

time, and energy efficiency, remains a key challenge in MapReduce task scheduling. While many 

papers focus on optimizing individual metrics, achieving a balance between competing objectives 

without sacrificing overall system performance is non-trivial. Future research efforts should explore 

multi-objective optimization techniques to address these trade-offs effectively. 

In conclusion, this review provides valuable insights into the state-of-the-art in MapReduce task 

scheduling, highlighting the diverse methodologies and approaches employed to tackle this complex 

problem. While significant advancements have been made, several challenges and opportunities for 

future research remain. By addressing these challenges and leveraging emerging technologies such as 

machine learning and optimization algorithms, researchers can continue to drive innovation in 

MapReduce task scheduling and pave the way for more efficient and scalable big data processing 

solutions. 

3. Proposed Design of an Iterative Method for MapReduce Scheduling Using Deep 

Reinforcement Learning and Anomaly Detection 

To overcome issues of high complexity of deployment, and low efficiency of scheduling which are 

present in existing Map Reduce based scheduling methods, this section discusses design of an Iterative 

Method for MapReduce Scheduling Using Deep Reinforcement Learning and Anomaly Detection 

Process. Initially, as per figure 1, the design of the Deep Reinforcement Learning for Dynamic Job 

Scheduling (DRDJS) model represents a sophisticated approach to optimizing scheduling policies in 

real-time for distributed computing environments, specifically within the MapReduce framework. This 

model is fundamentally grounded in the principles of reinforcement learning (RL), where an agent 

learns to make decisions by interacting with a dynamic environment. The aim is to maximize a 

cumulative reward, which, in the context of job scheduling, translates to minimizing the overall job 

completion timestamp while maximizing resource utilization efficiency. DRDJS utilizes a state-action-

reward-state-action (SARSA) learning algorithm, which is an on-policy RL method. The choice of 

SARSA is justified by its ability to learn policies that are more conservative compared to off-policy 

methods like Q-learning, since it updates its policy as it learns from the actions actually taken, rather 

than from a greedy policy, making it more stable in the highly dynamic environments typical of job 

scheduling in distributed systems. The model integrates complex mathematical operations to capture 

the dynamics of job scheduling processes.  

Each state 𝑠𝑡 in the system is represented by a vector of system metrics, such as CPU load, memory 

usage, network bandwidth, and queue characteristics at timestamp 𝑡 sets. This is mathematically 

represented via equation 1, 

𝑠𝑡 = (𝑙𝑐𝑝𝑢(𝑡), 𝑙𝑚𝑒𝑚(𝑡), 𝑏𝑛𝑒𝑡(𝑡), 𝑞𝑐ℎ𝑎𝑟(𝑡)) … (1) 

Where, 𝑙𝑐𝑝𝑢(𝑡), 𝑙𝑚𝑒𝑚(𝑡), and 𝑏𝑛𝑒𝑡(𝑡) represent the load on CPU, memory usage, and network 

bandwidth, respectively, and 𝑞𝑐ℎ𝑎𝑟(𝑡) represents the queue characteristics. The action 𝑎𝑡 at timestamp 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

606 
https://internationalpubls.com 

𝑡 is chosen using a policy derived from the current Q Value, which uses an epsilon-greedy strategy to 

balance exploration and exploitation. This is expressed via equation 2, 

𝑎𝑡 = 𝑎𝑟𝑔𝑚𝑎𝑥𝑎(𝑄(𝑠𝑡, 𝑎) − 𝜖 ⋅ 𝑆𝑇𝑂𝐶𝐻) … (2) 

Where, 𝜖 is the exploration rate, and STOCH is a stochastic number between 0 and 1 sets. The reward 

𝑟𝑡 at timestamp 𝑡 is defined as the negative of the job completion delay, integrated over a timestamp 

interval, representing the immediate cost of the chosen action via equation 3, 

𝑟𝑡 = − ∫ 𝑐(𝜏) 𝑑𝜏
𝑡+𝛥𝑡

𝑡

… (3) 

Where, 𝑐(𝜏) is the job completion timestamp at continuous timestamp 𝜏, and Δ𝑡 is the scheduling 

interval sets. Next, the update rule for the Q Value is given by the SARSA algorithm, incorporating 

both the reward received and the estimate of subsequent states and actions via equation 4, 

𝑄(𝑠𝑡, 𝑎𝑡) ← 𝑄(𝑠𝑡, 𝑎𝑡) + 𝛼[𝑟(𝑡 + 1) + 𝛾𝑄(𝑠(𝑡 + 1), 𝑎(𝑡 + 1)) − 𝑄(𝑠𝑡, 𝑎𝑡)] … (4) 

Where, 𝛼 is the learning rate and 𝛾 is the discount factor. Based on this, the temporal difference (TD) 

error, which measures the difference between the estimated rewards of successive states, is crucial for 

convergence, which is represented via equation 5, 

𝛿𝑡 = 𝑟(𝑡 + 1) + 𝛾𝑄(𝑠(𝑡 + 1), 𝑎(𝑡 + 1)) − 𝑄(𝑠𝑡, 𝑎𝑡) … (5) 

Finally, the convergence of the policy toward an optimal policy is ensured by the condition that the 

derivative of the temporal difference error with respect to the policy parameters approaches zero as 

training progresses, which is represented via equation 6, 

𝑑𝛿𝑡

𝑑𝜃
→ 0 𝑎𝑠 𝑡 → ∞ … (6) 

Where, 𝜃 represents the parameters of the policy (weights in a neural network) sets. The integration of 

DRDJS within the broader scheduling framework complements other methods such as ADAS and 

CJCS by providing a dynamic and continuously adapting scheduling policy that responds in real 

timestamp to changes in the system state and workload. This model is particularly effective in 

environments where job characteristics and system performance metrics are highly variable, ensuring 

that resource allocation is always aligned with the current system needs and conditions, thereby 

enhancing overall efficiency and reducing job processing delays. 

Next, as per figure 2, the Anomaly Detection-driven Adaptive Scheduling (ADAS) model forms a 

critical component of a comprehensive scheduling framework tailored for MapReduce systems. This 

model leverages anomaly detection techniques to ensure efficient resource allocation and the 

prioritization of critical jobs in the presence of anomalies, which are deviations from normal 

operational metrics that might indicate potential issues or bottlenecks within the system.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

607 
https://internationalpubls.com 

 

Figure 1. Model Architecture for the Proposed Scheduling Process 

ADAS is specifically designed to enhance system resilience by dynamically adapting to unexpected 

changes, thus maintaining system performance and reducing delays in critical job completions. The 

rationale behind incorporating ADAS within the broader scheduling framework stems from its 

capability to address and mitigate disruptions promptly. Unlike traditional methods that follow static 

rules, ADAS continuously monitors system performance, applying machine learning algorithms to 

detect anomalies and adjust scheduling decisions in real-time scenarios. This approach is particularly 

effective in environments characterized by variable workloads and the potential for unexpected system 

behavior, making it an essential complement to other models like DRDJS, which focuses on optimizing 

job scheduling through reinforcement learning. For this model, let 𝑠𝑡 represent the state of the system 

at timestamp 𝑡, encapsulated by a vector of observed system metrics, via equation 7, 

𝑠𝑡 = (𝑥1(𝑡), 𝑥2(𝑡), … , 𝑥𝑛(𝑡)) … (7) 

Where, 𝑥𝑖(𝑡) represents individual metrics such as CPU load, memory usage, network throughput, etc., 

at timestamp 𝑡 sets. An anomaly detection function 𝑓(𝑠𝑡) is applied to the system state to evaluate the 

presence of anomalies. This function could typically be an output from a machine learning model 

trained on historical data to recognize patterns deviating from the norm, via equation 8, 

𝑎𝑡 = 𝑓(𝑠𝑡) = {
1, 𝑖𝑓 𝑎𝑛𝑜𝑚𝑎𝑙𝑦 𝑑𝑒𝑡𝑒𝑐𝑡𝑒𝑑

0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
… (8) 

Upon detection of an anomaly, the resource allocation strategy 𝑅(𝑎𝑡,𝑠𝑡) is adjusted, which is expressed 

via equation 9, 

𝑅(𝑎𝑡, 𝑠𝑡) = 𝑔(𝑠𝑡) + 𝛿(𝑎𝑡) … (9) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

608 
https://internationalpubls.com 

Where, 𝑔(𝑠𝑡) represents the normal resource allocation policy, and 𝛿(𝑎𝑡) is an adjustment function that 

reallocates resources to prioritize critical jobs when an anomaly is detected. The sensitivity of the 

adjustment function with respect to the anomaly indicator is given via equation 10, 

𝜕𝑅

𝜕𝑎𝑡
=

𝜕𝛿(𝑎𝑡)

𝜕𝑎𝑡
… (10) 

Thus, indicating how resource allocation should be modified in response to detected anomalies. To 

understand the cumulative impact of adjustments over a scheduling period 𝑇, the integral of the 

resource allocation strategy is calculated via equation 11, 

∫ 𝑅(𝑎𝑡, 𝑠𝑡) 𝑑𝑡
𝑇

0

= 0 … (11) 

Which provides a measure of the total resources allocated in response to the dynamics of the detected 

anomalies. Ensuring that the system returns to normal operation after addressing anomalies is critical. 

This is described by the timestamp decay function of the adjustment, aiming for the anomaly response 

to diminish as the system stabilizes, via equation12, 

𝑙𝑖𝑚𝑡→∞𝛿(𝑎𝑡) = 0 … (12) 

This operation ensures that the system’s adaptive responses are temporary and proportional to the 

detected anomalies, thereby preventing overcompensation and maintaining overall system efficiency. 

ADAS's capability to dynamically adjust resource allocation in real-time provides a crucial layer of 

adaptability within the scheduling framework. By effectively responding to anomalies, ADAS ensures 

that critical jobs are prioritized without sacrificing the overall system performance, thus 

complementing the continuous learning and optimization processes of models like DRDJS. Together, 

these models enhance the robustness and efficiency of job scheduling in distributed computing 

environments, addressing both expected and unexpected variations in workload and system 

performance. 

Finally, the Cluster-based Job Categorization and Scheduling (CJCS) model is integrated, which 

represents a significant evolution in resource allocation strategies within distributed computing 

environments like MapReduce. This model employs advanced clustering algorithms to categorize jobs 

into distinct clusters based on similar characteristics such as resource requirements, expected execution 

delays, and interdependency scores. By segmenting jobs in this manner, CJCS enables the 

development of targeted scheduling policies that optimize resource allocation for each category, thus 

enhancing the overall efficiency and throughput of the system. The choice of CJCS is justified by its 

capacity to manage the diversity of job types effectively within a heterogeneous environment. Unlike 

generic scheduling solutions that treat all jobs as uniform tasks, CJCS recognizes and strategically 

manages the unique requirements of different job categories, significantly reducing resource 

contention and job completion delays. This approach complements the dynamic adaptability of models 

like DRDJS and the responsive capabilities of ADAS by providing a structured yet flexible framework 

that further refines resource allocation and job prioritization based on categorized needs.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

609 
https://internationalpubls.com 

Each job 𝑗 is represented by a vector of features 𝑣𝑗, which includes metrics such as CPU requirements, 

memory needs, expected runtime, and priority, via equation 13, 

𝑣𝑗 = (𝑐𝑗, 𝑚𝑗, 𝑡𝑗, 𝑝𝑗) … (13) 

Where, 𝑐𝑗, 𝑚𝑗, 𝑡𝑗, and 𝑝𝑗 represent CPU, memory, delay, and priority requirements respectively. The 

jobs are grouped using a clustering algorithm, typically K-means, based on their feature vector sets. 

The objective function for clustering, which aims to minimize the within-cluster sum of squares, is 

given via equation 14, 

𝑔𝑐 =  𝑚𝑖𝑛 ∑ ∑ ∥ 𝑣𝑗 − 𝜇𝑘 ∥2 … (14)

𝑗∈𝑆𝑘

𝐾

𝑘=1

 

Where, gc is the grouping criteria, 𝑆𝑘 represents the set of jobs in cluster 𝑘 and 𝜇𝑘 is the centroid of 

cluster 𝑘 sets. Resource allocation for each job is determined based on its cluster membership, and is 

expressed via equation 15, 

𝑅(𝑣𝑗) = 𝛼𝑘 ⋅ 𝑣𝑗 … (15) 

Where, 𝛼𝑘 is a scaling factor derived from the cluster characteristics, optimizing resource allocation 

for the cluster. To optimize 𝛼𝑘, the derivative of the resource allocation function with respect to 𝛼𝑘 is 

calculated, ensuring that changes in allocation are aligned with the changes in cluster characteristics 

via equation 16, 

 

Figure 2. Overall Flow of the Proposed Scheduling Process 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

610 
https://internationalpubls.com 

𝜕𝑅

𝜕𝛼𝑘
= 𝑣𝑗 … (16) 

Next, the total resource usage for a cluster over a scheduling period 𝑇 is given via equation 17, 

∫ 𝑅(𝑣𝑗(𝑡)) 𝑑𝑡 = 0
𝑇

0

… (17) 

Where, 𝑣𝑗(𝑡) represents the resource demand of job 𝑗 at timestamp 𝑡 sets. The optimization of clustering 

parameters is critical for adapting to changes in job characteristics over temporal instance sets. This is 

governed via equation 18, 

𝑑𝜇𝑘

𝑑𝑡
= 𝜂 ∑ (𝑣𝑗 − 𝜇𝑘)

𝑗∈𝑆𝑘

… (18) 

Where, 𝜂 is the learning rate, facilitating the dynamic adjustment of cluster centroids based on 

incoming jobs. CJCS significantly enhances the precision of resource allocation by tailoring 

scheduling decisions to the specific needs of job clusters, reducing competition for resources and 

enabling more effective use of the computing infrastructure deployments. By integrating this model 

with DRDJS and ADAS, the overall scheduling framework becomes highly robust, capable of 

dynamically adapting to varying job demands and system conditions, and efficiently managing 

resources across a broad spectrum of tasks, ensuring optimal performance and reduced job completion 

times in complex distributed computing environments. Next, we discuss efficiency of the proposed 

model in terms of different scenarios, and compare it with existing methods, which will assist readers 

to evaluate its performance in real-time deployments. 

4. Comparative Result Analysis 

To evaluate the performance and robustness of the proposed scheduling framework, which includes 

Deep Reinforcement Learning for Dynamic Job Scheduling (DRDJS), Anomaly Detection-driven 

Adaptive Scheduling (ADAS), and Cluster-based Job Categorization and Scheduling (CJCS), a 

comprehensive experimental setup was designed. This setup aims to simulate real-world conditions in 

a controlled environment to provide a detailed analysis of each component's effectiveness and the 

integrated operation of the entire system. 

System Configuration 

The experiments were conducted on a distributed computing cluster consisting of 20 nodes. Each node 

is equipped with an Intel Xeon Processor E5-2670, 64GB RAM, and connected through a 10Gbps 

Ethernet network. This configuration reflects a typical medium-scale enterprise-level distributed 

system, suitable for evaluating the scheduling models under realistic workloads and system conditions. 

Dataset 

For the purpose of these experiments, three contextual dataset samples were chosen to represent 

different types of workloads typically processed in distributed computing environments: 

• Log Processing Workload: This dataset comprises web server logs totaling 100 GB, intended 

to simulate the processing of unstructured text data samples. The logs vary in size from 500KB 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

611 
https://internationalpubls.com 

to 15MB, creating a varied workload to test the effectiveness of the CJCS in handling jobs with 

different resource requirements. 

• Image Processing Workload: Composed of 50,000 images, averaging 2MB each, this dataset 

is used to evaluate the performance of the system in handling large-scale image data, which 

requires significant computational resources for tasks such as resizing, filtering, and pattern 

recognition. 

• Financial Transactions Workload: This dataset includes 10 million financial transaction 

records, each record being approximately 1KB in size. This high-frequency, low-latency 

dataset tests the system’s ability to efficiently process a large number of small jobs, which is 

critical for real-time analytics applications. 

Parameters 

The parameters for each of the scheduling models were set as follows: 

• DRDJS: 

• Learning rate (𝛼α): 0.1 

• Discount factor (𝛾γ): 0.95 

• Epsilon (𝜖ϵ) for the epsilon-greedy policy: 0.05, decaying over timestamp to encourage 

exploitation as the model converges. 

• ADAS: 

• Anomaly detection threshold, set dynamically based on the mean and standard 

deviation of historic system performance metrics. 

• Resource reallocation factor (𝛿δ): Adjusts the resource allocation by up to 20% in 

response to detected anomalies. 

• CJCS: 

• Number of clusters (K): 5, determined based on preliminary analysis using the Elbow 

Method to ensure optimal grouping without overfitting. 

• Cluster update frequency: Every 100 jobs to adapt to changes in the job characteristics 

over time. 

Experimental Design 

The experiments were conducted under varying system loads and conditions to evaluate each model's 

responsiveness and effectiveness. Each dataset was processed under three different scenarios: 

1. Baseline Scenario: Where jobs are scheduled using a traditional First-In-First-Out (FIFO) 

strategy without any adaptive optimizations. 

2. Individual Model Scenario: Each model (DRDJS, ADAS, CJCS) was tested independently to 

assess its specific impact on the system’s performance. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

612 
https://internationalpubls.com 

3. Integrated Model Scenario: All three models were integrated, functioning simultaneously to 

leverage their combined strengths. 

The primary metrics used to evaluate performance included job completion time, system throughput, 

and resource utilization efficiency. These metrics provide insight into the effectiveness of the 

scheduling models in optimizing the processing of diverse workloads. This experimental setup 

provides a rigorous framework for evaluating the proposed scheduling models, ensuring that the 

findings are robust, replicable, and indicative of performance in real-world distributed computing 

environments. Based on this setup, effectiveness of the proposed scheduling framework was evaluated 

using the previously mentioned contextual datasets: Log Processing, Image Processing, and Financial 

Transactions Workloads. For each dataset, the performance of our proposed model was compared 

against three other methods, referred to as [2], [5], and [15], representing conventional scheduling 

strategies. The results are presented in a series of tables that illustrate the job completion time, system 

throughput, and resource utilization efficiency across different scenarios. 

Table 2: Job Completion Times for Log Processing Workload 

Method Average Job Completion timestamp (seconds) 

Proposed Model 55 

[2] 75 

[5] 70 

[15] 85 

Table 2 displays the average job completion times for the Log Processing Workload. The proposed 

model shows a significant improvement, reducing the average job completion timestamp by 

approximately 27% compared to the slowest method [15]. 

 

Figure 3. Job Completion & Job Per Hour for the Proposed Model Process 

0

10

20

30

40

50

60

70

80

90

Average Job Completion timestamp (seconds) Jobs per Hour (x10)

Proposed Model [2] [5] [15]



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

613 
https://internationalpubls.com 

Table 3: System Throughput for Log Processing Workload 

Method Jobs per Hour 

Proposed Model 650 

[2] 480 

[5] 500 

[15] 425 

Table 3 compares the system throughput in terms of jobs processed per hour. The proposed model 

enhances throughput by up to 35% over method [15], indicating a more efficient use of resources. 

Table 4: Resource Utilization Efficiency for Image Processing Workload 

Method CPU Utilization (%) Memory Utilization (%) 

Proposed Model 85 90 

[2] 75 80 

[5] 78 82 

[15] 70 75 

Table 4 shows the resource utilization for the Image Processing Workload, where the proposed model 

achieves better CPU and memory utilization, indicating a more balanced load distribution across the 

cluster. 

 

Figure 4. CPU & Memory Utilization Levels 

Table 5: Job Completion Times for Image Processing Workload 

Method Average Job Completion timestamp (seconds) 

Proposed Model 120 

[2] 150 

[5] 140 

[15] 160 

 

Table 5 outlines the job completion times for processing images. The proposed model shows superior 

performance by reducing the completion timestamp by 25% compared to method [15]. 

0

10

20

30

40

50

60

70

80

90

100

CPU Utilization (%) Memory Utilization (%)

Proposed Model [2] [5] [15]



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

614 
https://internationalpubls.com 

Table 6: System Throughput for Financial Transactions Workload 

Method Transactions per Hour 

Proposed Model 1,200,000 

[2] 900,000 

[5] 950,000 

[15] 850,000 

Table 6 focuses on the throughput for the Financial Transactions Workload, where the proposed model 

processes more transactions per hour, showcasing its ability to handle high-frequency, low-latency 

jobs effectively. 

Table 7: Resource Utilization Efficiency for Financial Transactions Workload 

Method CPU Utilization (%) Memory Utilization (%) 

Proposed Model 95 93 

[2] 85 83 

[5] 88 85 

[15] 80 78 

 

 

Figure 5. Resource Utilization Efficiency for Financial Transactions Workload 

Table 7 evaluates the CPU and memory utilization for the Financial Transactions Workloads. The 

proposed model maintains high utilization rates, demonstrating efficient resource management even 

under the demands of rapid transaction processing. These tables collectively demonstrate the 

superiority of the proposed scheduling framework over the comparative methods [2], [5], and [15]. By 

effectively reducing job completion times, increasing throughput, and optimizing resource utilization, 

the framework offers a significant improvement in performance across diverse and challenging 

workload scenarios. This analysis confirms the effectiveness of integrating adaptive scheduling 

techniques within a distributed computing environment. Next, we discuss a practical use case for the 

proposed model, which will assist readers to understand the entire scheduling process. 

0

10

20

30

40

50

60

70

80

90

100

CPU Utilization (%) Memory Utilization (%)

Proposed Model [2] [5] [15]



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

615 
https://internationalpubls.com 

Practical Use Case 

To underscore the efficacy of the proposed scheduling framework in a distributed computing 

environment, the performance of each component model—Deep Reinforcement Learning for Dynamic 

Job Scheduling (DRDJS), Anomaly Detection-driven Adaptive Scheduling (ADAS), and Cluster-

based Job Categorization and Scheduling (CJCS)—was evaluated using a set of defined metrics and a 

specific dataset with attributed values. This dataset includes job characteristics and system 

performance indicators designed to represent a typical workload in a distributed computing scenario. 

Each model processes the data independently to showcase its unique contributions to the scheduling 

framework, and the results are then integrated to demonstrate the combined impact on system 

performance. 

Table 8: Output from Deep Reinforcement Learning for Dynamic Job Scheduling (DRDJS) 

Job 

ID 

CPU Load 

Before (%) 

Memory Usage 

Before (%) 

Job 

Priority 

CPU Load 

After (%) 

Memory Usage 

After (%) 

Job Completion 

timestamp (s) 

J101 65 70 High 55 60 90 

J102 75 80 Medium 65 70 120 

J103 50 55 Low 45 50 80 

 

Table 8 illustrates how DRDJS optimizes resource allocation based on the job's priority and current 

system load, leading to improved CPU and memory usage and reduced job completion delays. 

Table 9: Output from Anomaly Detection-driven Adaptive Scheduling (ADAS) 

Timestamp System Load 

(%) 

Anomaly 

Detected 

Response Triggered New Priority Allocation 

T1 80 Yes Increase Resources High 

T2 45 No None Normal 

T3 90 Yes Increase Resources Critical 

Table 9 provides an overview of how ADAS identifies anomalies in system performance (e.g., high 

system load) and dynamically adjusts resource allocation to ensure the stability and efficiency of job 

processing. 

Table 10: Output from Cluster-based Job Categorization and Scheduling (CJCS) 

Job 

ID 

Resource 

Demand 

Cluster 

Assigned 

Allocated CPU 

(%) 

Allocated 

Memory (%) 

Expected Completion 

timestamp (s) 

J201 High C1 70 75 100 

J202 Medium C2 60 65 150 

J203 Low C3 50 55 200 

 

Table 10 demonstrates CJCS's capability to categorize jobs into clusters based on their resource 

demands and assign appropriate resources, optimizing the scheduling process to enhance job 

completion times based on the cluster characteristics. 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

616 
https://internationalpubls.com 

Table 11: Final Integrated Outputs from the Proposed Scheduling Framework 

Job 

ID 

Initial CPU 

Load (%) 

Initial Memory 

Usage (%) 

Final CPU 

Load (%) 

Final Memory 

Usage (%) 

Total Job Completion 

timestamp (s) 

J301 85 90 55 60 85 

J302 60 65 45 50 70 

J303 70 75 50 55 60 

 

Table 11 encapsulates the overall efficiency of the integrated scheduling framework, demonstrating 

significant improvements in CPU and memory utilization and reductions in job completion times, 

showcasing the synergistic effect of combining DRDJS, ADAS, and CJCS in a unified operational 

model process. The data presented in Tables 8 through 11 clearly indicate that each component of the 

proposed scheduling framework significantly contributes to enhancing the overall system 

performance. DRDJS optimizes real-time resource allocation, ADAS effectively manages system 

anomalies to maintain performance stability, and CJCS ensures jobs are categorized and scheduled 

optimally according to their specific needs. When integrated, these components not only improve 

individual metrics such as CPU and memory usage and job completion times but also ensure that the 

system as a whole operates more efficiently and resiliently. This collective improvement underscores 

the potential of advanced scheduling techniques in transforming the operational dynamics of 

distributed computing environments. 

5. Conclusion and Future Scopes 

The results presented in this study demonstrate the effectiveness of the proposed scheduling framework 

comprising Deep Reinforcement Learning for Dynamic Job Scheduling (DRDJS), Anomaly 

Detection-driven Adaptive Scheduling (ADAS), and Cluster-based Job Categorization and Scheduling 

(CJCS) in a distributed computing environment. The integrated framework substantially outperformed 

traditional scheduling methods [2], [5], and [15], across various metrics and workload scenarios. 

In the Log Processing Workload, the proposed model reduced the average job completion timestamp 

to 55 seconds compared to 75, 70, and 85 seconds for methods [2], [5], and [15] respectively, 

showcasing a reduction of up to 35%. Furthermore, it improved system throughput to 650 jobs per 

hour, significantly higher than the 480, 500, and 425 jobs per hour processed by the comparative 

methods. Such enhancements underscore the framework's capability to handle large-scale data 

efficiently while minimizing processing time. 

For the Image Processing Workload, the model exhibited a remarkable reduction in average job 

completion timestamp to 120 seconds, outperforming the other methods by 20% to 25%. The resource 

utilization efficiency in this workload demonstrated the framework's ability to manage high-resource-

demand tasks effectively, achieving 85% CPU and 90% memory utilization, which is a notable 

improvement over the other methods. 

In scenarios involving high-frequency, low-latency tasks such as the Financial Transactions Workload, 

the proposed model processed up to 1,200,000 transactions per hour, far surpassing the 900,000, 

950,000, and 850,000 transactions managed by methods [2], [5], and [15]. This not only illustrates the 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

617 
https://internationalpubls.com 

model's superior performance in terms of throughput but also highlights its robustness in maintaining 

high resource utilization rates under intensive workloads. 

Future Scope 

The promising results of this framework open numerous avenues for future research. One potential 

area of development is the exploration of even more sophisticated machine learning algorithms for job 

categorization and anomaly detection, which could further refine the adaptability and efficiency of the 

scheduling processes. Moreover, incorporating additional predictive capabilities into the model could 

pre-emptively optimize resource allocation and job scheduling, potentially reducing job completion 

times even further. 

Further exploration into hybrid models that integrate other forms of reinforcement learning, such as 

actor-critic methods, could provide deeper insights into the trade-offs between exploration and 

exploitation, leading to even more refined scheduling policies. Additionally, expanding the framework 

to accommodate edge computing scenarios where data processing and resource allocation decisions 

need to be made closer to the data source could dramatically increase the relevance and applicability 

of the scheduling framework in modern distributed architectures. 

Finally, the impact of real-time data analytics on scheduling decisions presents a fruitful area for 

research, where live data streams could dynamically influence scheduling decisions, thereby adapting 

to changing data patterns and workload demands instantaneously. This would be particularly beneficial 

in environments that deal with volatile and time-sensitive data samples. 

In conclusion, this study not only proves the viability of the proposed intelligent scheduling framework 

in enhancing the performance of distributed computing systems but also sets a foundation for future 

advancements that could revolutionize the efficiency and adaptability of large-scale data processing 

environments. 

References   

[1] H. Zheng and J. Wu, "Joint Scheduling of Overlapping MapReduce Phases: Pair Jobs for Optimization," in IEEE 

Transactions on Services Computing, vol. 14, no. 5, pp. 1453-1463, 1 Sept.-Oct. 2021, doi: 

10.1109/TSC.2018.2875698. 

keywords: {Optimal scheduling;Job shop scheduling;Schedules;Task analysis;Processor scheduling;MapReduce 

framework;map and shuffle phases;joint scheduling;makespan optimization}, 

[2] X. Li, F. Chen, R. Ruiz and J. Zhu, "MapReduce Task Scheduling in Heterogeneous Geo-Distributed Data Centers," 

in IEEE Transactions on Services Computing, vol. 15, no. 6, pp. 3317-3329, 1 Nov.-Dec. 2022, doi: 

10.1109/TSC.2021.3092563. 

keywords: {Task analysis;Data centers;Data communication;Processor scheduling;Computer 

architecture;Containers;Scheduling;MapReduce;scheduling;geo-distributed data centers;hungarian algorithm;total 

tardiness}, 

[3] R. Jeyaraj and A. Paul, "Optimizing MapReduce Task Scheduling on Virtualized Heterogeneous Environments 

Using Ant Colony Optimization," in IEEE Access, vol. 10, pp. 55842-55855, 2022, doi: 

10.1109/ACCESS.2022.3176729. 

keywords: {Task analysis;Resource management;Cloud computing;Containers;Dynamic scheduling;Quality of 

service;Optimal scheduling;Ant colony optimization;bin packing;heterogeneity;MapReduce;resource 

utilization;task scheduling}, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

618 
https://internationalpubls.com 

[4] T. D. Dang, D. Hoang and D. N. Nguyen, "Trust-Based Scheduling Framework for Big Data Processing with 

MapReduce," in IEEE Transactions on Services Computing, vol. 15, no. 1, pp. 279-293, 1 Jan.-Feb. 2022, doi: 

10.1109/TSC.2019.2938959. 

keywords: {Cloud computing;Task analysis;Processor scheduling;Security;Big Data 

applications;Measurement;Data privacy;Heuristic algorithms;Trusted computing;Trust-aware framework;data 

sensitive;big data security;trust-based scheduling;MapReduce}, 

[5] D. Yang, D. Cheng, W. Rang and Y. Wang, "Joint Optimization of MapReduce Scheduling and Network Policy in 

Hierarchical Data Centers," in IEEE Transactions on Cloud Computing, vol. 10, no. 1, pp. 461-473, 1 Jan.-March 

2022, doi: 10.1109/TCC.2019.2961653. 

keywords: {Task analysis;Data centers;Cloud computing;Bandwidth;Network architecture;Topology;Joint 

optimization;MapReduce scheduling;network policy;hierarchical clouds;topology aware assignment}, 

[6] Y. Yao, H. Gao, J. Wang, B. Sheng and N. Mi, "New Scheduling Algorithms for Improving Performance and 

Resource Utilization in Hadoop YARN Clusters," in IEEE Transactions on Cloud Computing, vol. 9, no. 3, pp. 

1158-1171, 1 July-Sept. 2021, doi: 10.1109/TCC.2019.2894779. 

keywords: {Task analysis;Yarn;Resource management;Job shop scheduling;Scheduling algorithms;Data 

processing;MapReduce;Resource Management;YARN;Data Processing}, 

[7] H. Tan, C. Zhang, C. Xu, Y. Li, Z. Han and X. -Y. Li, "Regularization-Based Coflow Scheduling in Optical Circuit 

Switches," in IEEE/ACM Transactions on Networking, vol. 29, no. 3, pp. 1280-1293, June 2021, doi: 

10.1109/TNET.2021.3058164. 

keywords: {Scheduling;Optical switches;Approximation algorithms;Optical packet switching;Integrated circuit 

modeling;Scheduling algorithms;Data centers;Coflow Scheduling;Optical Circuit Switch;Approximation 

Algorithm}, 

[8] X. Tang et al., "Cost-Efficient Workflow Scheduling Algorithm for Applications With Deadline Constraint on 

Heterogeneous Clouds," in IEEE Transactions on Parallel and Distributed Systems, vol. 33, no. 9, pp. 2079-2092, 

1 Sept. 2022, doi: 10.1109/TPDS.2021.3134247. 

keywords: {Cloud computing;Task analysis;Costs;Computational modeling;Scheduling;Job shop 

scheduling;Heuristic algorithms;Workflow application;cost;heterogeneous clouds;schedule length;task 

scheduling}, 

[9]  L. Shi, Y. Liu, J. Zhang and T. Robertazzi, "Coflow Scheduling in Data Centers: Routing and Bandwidth 

Allocation," in IEEE Transactions on Parallel and Distributed Systems, vol. 32, no. 11, pp. 2661-2675, 1 Nov. 2021, 

doi: 10.1109/TPDS.2021.3068424. 

keywords: {Routing;Bandwidth;Schedules;Channel allocation;Data centers;Job shop scheduling;Optimal 

scheduling;Coflow scheduling;distributed computing;non-linear programming;convex optimization}, 

[10] Z. Wu et al., "Scheduling-Guided Automatic Processing of Massive Hyperspectral Image Classification on Cloud 

Computing Architectures," in IEEE Transactions on Cybernetics, vol. 51, no. 7, pp. 3588-3601, July 2021, doi: 

10.1109/TCYB.2020.3026673. 

keywords: {Task analysis;Cloud computing;Processor scheduling;Sparks;Scheduling;Hyperspectral 

imaging;Cloud computing;distributed and parallel processing;divisible task scheduling;hyperspectral image (HSI) 

classification;partitioning factor}, 

[11] A. Jajoo, Y. C. Hu and X. Lin, "A Case for Sampling-Based Learning Techniques in Coflow Scheduling," in 

IEEE/ACM Transactions on Networking, vol. 30, no. 4, pp. 1494-1508, Aug. 2022, doi: 

10.1109/TNET.2021.3138923. 

keywords: {Scalability;Queueing analysis;Schedules;Task analysis;Social networking 

(online);Scheduling;Production;Coflows;flow scheduling;coflow completion timestamp (CCT);sampling-based 

learning;online learning;coflow size prediction}, 

[12] J. Wang, X. Li, R. Ruiz, J. Yang and D. Chu, "Energy Utilization Task Scheduling for MapReduce in Heterogeneous 

Clusters," in IEEE Transactions on Services Computing, vol. 15, no. 2, pp. 931-944, 1 March-April 2022, doi: 

10.1109/TSC.2020.2966697. 

keywords: {Task analysis;Servers;Resource management;Energy consumption;Scheduling;Fuzzy logic;Processor 

scheduling;Energy consumption;task scheduling;MapReduce;heuristic}, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

619 
https://internationalpubls.com 

[13] F. Pace, D. Venzano, D. Carra and P. Michiardi, "A Flexible Heuristic to Schedule Distributed Analytic 

Applications in Compute Clusters," in IEEE Transactions on Cloud Computing, vol. 11, no. 3, pp. 2217-2230, 1 

July-Sept. 2023, doi: 10.1109/TCC.2019.2926977. 

keywords: {Scheduling;Resource management;Task analysis;Sparks;Processor scheduling;Schedules;Data 

analysis;Scheduling;distributed applications;distributed systems}, 

[14] Y. Zhao, C. Tian, J. Fan, T. Guan, X. Zhang and C. Qiao, "Joint Reducer Placement and Coflow Bandwidth 

Scheduling for Computing Clusters," in IEEE/ACM Transactions on Networking, vol. 29, no. 1, pp. 438-451, Feb. 

2021, doi: 10.1109/TNET.2020.3037064. 

keywords: {Bandwidth;Task analysis;Processor scheduling;Schedules;Data transfer;Optimization;IEEE 

transactions;Computing clusters;reducer placement;coflow scheduling}, 

[15] Z. Tang, L. Du, X. Zhang, L. Yang and K. Li, "AEML: An Acceleration Engine for Multi-GPU Load-Balancing in 

Distributed Heterogeneous Environment," in IEEE Transactions on Computers, vol. 71, no. 6, pp. 1344-1357, 1 

June 2022, doi: 10.1109/TC.2021.3084407. 

keywords: {Graphics processing units;Task analysis;Load modeling;Sparks;Hardware;Computational 

modeling;Processor scheduling;Distributed system;GPGPU;heterogeneous environment;load-balancing;multi-

GPU}, 

[16] A. Kumar, N. Varshney, S. Bhatiya and K. U. Singh, "Replication-Based Query Management for Resource 

Allocation Using Hadoop and MapReduce over Big Data," in Big Data Mining and Analytics, vol. 6, no. 4, pp. 465-

477, December 2023, doi: 10.26599/BDMA.2022.9020026. 

keywords: {Schedules;Scalability;Clustering algorithms;Big Data;Timing;Servers;Security;Resource 

management;Reliability;Sorting;big data;hadoop;mapreduce;resource allocation;query management}, 

[17] H. Wang, G. Liu and H. Shen, "Cooperative Job Scheduling and Data Allocation in Data-Intensive Parallel 

Computing Clusters," in IEEE Transactions on Cloud Computing, vol. 11, no. 3, pp. 2392-2406, 1 July-Sept. 2023, 

doi: 10.1109/TCC.2022.3206206. 

keywords: {Task analysis;Servers;Resource management;Schedules;Clustering algorithms;Processor 

scheduling;Costs;Job scheduler;data allocation;parallel computing;data locality}, 

[18] S. Wu, H. Chen, H. Jin and S. Ibrahim, "Shadow: Exploiting the Power of Choice for Efficient Shuffling in 

MapReduce," in IEEE Transactions on Big Data, vol. 8, no. 1, pp. 253-267, 1 Feb. 2022, doi: 10.1109/TBDATA 

SAMPLES.2019.2943473. 

keywords: {Task analysis;Big Data;Processor scheduling;Degradation;Electric breakdown;MapReduce;task 

scheduling;duplicated tasks;power of choice}, 

[19] L. Liu et al., "Bottleneck-Aware Non-Clairvoyant Coflow Scheduling With Fai," in IEEE Transactions on Cloud 

Computing, vol. 11, no. 1, pp. 1011-1025, 1 Jan.-March 2023, doi: 10.1109/TCC.2021.3128360. 

keywords: {Bandwidth;Data centers;Processor scheduling;Job shop scheduling;Fabrics;Cloud 

computing;Uplink;Coflow scheduling;coflow completion time;bottleneck-aware;datacenter networks}, 

[20] X. Zhang, X. Li, H. Du and R. Ruiz, "Task Scheduling for Spark Applications With Data Affinity on Heterogeneous 

Clusters," in IEEE Internet of Things Journal, vol. 9, no. 21, pp. 21792-21801, 1 Nov.1, 2022, doi: 

10.1109/JIOT.2022.3181997. 

keywords: {Task analysis;Sparks;Servers;Scheduling;Schedules;Data communication;Internet of Things;Data 

affinity;heterogeneous cluster;spark;task scheduling}, 

[21] A. Yang, J. Wang, Y. Mao, Y. Yao, N. Mi and B. Sheng, "Optimizing Internal Overlaps by Self-Adjusting Resource 

Allocation in Multi-Stage Computing Systems," in IEEE Access, vol. 9, pp. 88805-88819, 2021, doi: 

10.1109/ACCESS.2021.3089907. 

keywords:   {Task analysis;Containers;Job shop scheduling;Yarn;Dynamic scheduling;Scheduling 

algorithms;Resource management;MapReduce jobs;Hadoop scheduling;reduced makespan;resource management}, 

[22] C. Ruan, J. Wang, W. Jiang and T. Zhang, "Scheduling Coflows by Online Identification in Data Center Network," 

in IEEE Transactions on Emerging Topics in Computing, vol. 11, no. 4, pp. 1057-1069, Oct.-Dec. 2023, doi: 

10.1109/TETC.2023.3315512. 

keywords: {Task analysis;Containers;Sparks;Schedules;Data centers;Yarn;Machine learning;Coflow;data center 

networks;identification;scheduling}, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 4s (2024) 

 

620 
https://internationalpubls.com 

[23] T. R. Toha, A. S. M. Rizvi, J. Noor, M. A. Adnan and A. B. M. A. Al Islam, "Towards Greening MapReduce 

Clusters Considering Both Computation Energy and Cooling Energy," in IEEE Transactions on Parallel and 

Distributed Systems, vol. 32, no. 4, pp. 931-942, 1 April 2021, doi: 10.1109/TPDS.2020.3029724. 

keywords: {Cooling;Energy consumption;Time factors;Task analysis;Machine learning algorithms;Cloud 

computing;Temperature distribution;Green parallel computing;MapReduce;hadoop;machine learning}, 

[24] J. -J. Chen, H. -Y. Yu, C. -C. Hwuang, C. -J. Sung and Y. -S. Wu, "A Complexity Aware Scheduler With Dynamic 

Slot Allocation for Cloud Video Transcoding," in IEEE Access, vol. 10, pp. 104226-104241, 2022, doi: 

10.1109/ACCESS.2022.3210528. 

keywords: {Load management;Transcoding;Streaming media;Cloud computing;Resource 

management;Complexity theory;Heuristic algorithms;Cloud video transcoding;resource management;complexity-

aware scheduling;dynamic slot allocation;load balance operation;heterogeneous parallel system}, 

[25] G. Singh, A. Sharma, R. Jeyaraj and A. Paul, "Handling Non-Local Executions to Improve MapReduce Performance 

Using Ant Colony Optimization," in IEEE Access, vol. 9, pp. 96176-96188, 2021, doi: 

10.1109/ACCESS.2021.3091675. 

keywords: {Task analysis;Bandwidth;Servers;Cloud computing;Virtual machining;Switches;Data  transfer;Ant 

colony optimization;cloud computing;heterogeneous performance;MapReduce scheduler;virtualized