Introducing a Novel Feature in CloudSimPlus: Enhancing Load Balancer Efficiency based on AI

Technical Summary: Enhancing Load Balancer Efficiency in CloudSimPlus via AI

Problem Statement
Cloud computing environments are characterized by heterogeneity, dynamic resource provisioning, and diverse workloads, making load balancing a complex challenge. Existing studies often focus on the advantages and limitations of specific static load balancing techniques (e.g., Round Robin, Ant Colony Optimization) targeting single metrics, lacking practical tools that allow users to dynamically select the most appropriate strategy based on real-time conditions. Furthermore, while simulation frameworks like CloudSim and CloudSimPlus are essential for evaluating resource management strategies, they traditionally rely on fixed policies or require manual intervention to switch strategies. There is a need for an adaptive, user-oriented solution that can automatically adjust load balancing decisions based on runtime metrics such as response time, system overhead, and resource availability.

Methodology and Framework Design
The authors propose a novel feature for CloudSimPlus that integrates an AI-driven decision-making process to dynamically select load balancing strategies. The methodology involves a client-server architecture where the CloudSimPlus simulation environment (client) communicates with a pre-trained machine learning model hosted on a remote server (IaCloud Model Server).

1. Architecture: The core of the implementation is a new module, AdapterBroker, which extends the CloudSimPlus architecture. This module acts as a controller between the global CloudSimPlus components (specifically DatacenterBroker, VmAllocationPolicy, and CloudletScheduler) and the external AI model.
2. Communication: The AdapterBroker utilizes an IaCloud module to handle HTTP API requests. It sends current system metrics (e.g., response time, throughput, fault tolerance) to the server and receives a predicted optimal load balancing policy in return.
3. Operational Modes: The system supports two distinct scenarios:
   • Static Configuration: At initialization, the system reads user-defined metrics and thresholds from a configuration file. The model predicts a single load balancing technique which is applied for the duration of the simulation unless auto-adaptation is enabled.
   • Dynamic Adaptation: During runtime, a service monitoring module continuously collects statistical metrics via event classes. These metrics are sent to the pre-trained model, which dynamically re-evaluates and updates the load balancing policy in real-time to adapt to changing workload patterns and resource variations.
4. Policy Integration: The framework unifies multiple existing policies, including Round Robin, Join Idle Queue (JIQ), Power-Aware Load Balancing (PALB), and Ant Colony Optimization, treating them as independent instances that can be switched seamlessly by the AdapterBroker.

Key Contributions
The paper outlines six primary contributions:

• Dynamic Strategy Selection: Implementation of a mechanism using Decision Trees to automatically select the most suitable load balancing strategy based on the current state of the cloud environment, rather than relying on a fixed policy.
• Unified Policy Framework: Integration of diverse algorithms (Round Robin, JIQ, PALB, Ant Colony) into a single framework, allowing efficient switching based on system requirements.
• AI-Based Prediction: Utilization of a pre-trained Decision Tree model guided by runtime metrics to improve the accuracy and intelligence of policy selection.
• Continuous Runtime Adaptation: A system capable of monitoring the environment and updating policies in response to workload changes and resource availability.
• User-Oriented Objectives: Incorporation of user-defined goals (e.g., minimizing response time vs. reducing overhead) into the decision-making process to ensure the selected strategy aligns with specific performance targets.
• CloudSimPlus Extension: The technical extension of CloudSimPlus with this adaptive capability, enabling the simulation of intelligent, self-adjusting cloud environments.

Results
The authors conducted experiments to validate the approach, logging execution details including task IDs, assigned policies, start/finish times, and resource utilization. The results demonstrate the system's ability to switch policies dynamically during execution. For instance, the logs show tasks initially handled by the Join Idle Queue (JIQ) algorithm switching to Round Robin (RR) as the simulation progressed and conditions changed. The log files confirm that the system successfully updated its configuration (e.g., "Switched policy to: RoundRobinBroker at time 6405.692") based on the model's predictions. The results indicate that the system can effectively adapt to dynamic conditions, selecting relevant policies to manage tasks across different Virtual Machines (VMs) and Data Centers (DCs).

Significance and Claims
The paper claims that this feature significantly enhances the realism and flexibility of CloudSimPlus simulations. By moving away from static, pre-defined policies, the proposed approach allows researchers to evaluate adaptive and AI-driven load balancing solutions under controlled, repeatable conditions that mimic dynamic real-world environments. The authors assert that this tool facilitates future research by providing a framework where load balancing decisions are automatically adjusted based on real-time conditions and user objectives. They also note that the solution is designed to be portable and flexible, potentially supporting integration with other frameworks like OpenStack and Octavia via its API-based design.

Limitations
The authors modestly acknowledge several limitations:

• Latency: Reliance on a remote server for the pre-built model introduces communication latency, which could impact performance in time-sensitive scenarios.
• Configuration Complexity: Current metric configuration depends on user-defined files, which may reduce usability; a Graphical User Interface (GUI) is suggested as a future improvement.
• Scalability Testing: The solution has not yet been extensively evaluated under large-scale conditions involving high numbers of customers, tasks, and data centers, leaving its full scalability and robustness to be further assessed.