Abstract
The allocation planning of resources in a web server cluster is accomplished nowadays by the administrator. Once the internet is quite dynamic, as far as the use of resources is concerned, such a task may be considered critical and inefficient if accomplished manually. Our approach benefits from the use of agents to learn from the environment and adjust automatically the behavior of the system to make a better use of the available resources. With this approach it is possible to help the administrator by minimizing your stress in moments of work overload. The conception, specification, adopted allocation planning strategy, modeling in Petri nets, implementation of this platform in the Java language are presented. Experimentations and simulations which prove the efficiency of the proposal are presented.
Introduction: Automating the Hotseat
Managing a web server cluster during a traffic spike is often described as a "stressful, exhausting job" for system administrators. When the internet throws an unpredictable load at a cluster, manual resource allocation is simply too slow and error-prone to maintain high availability. In their research, Marques et al. propose DARC (Dynamic Architecture for Reconfiguration of Web servers Clusters), a multi-agent system that automates resource allocation in web clusters. By integrating intelligent agents with a load-balancing platform, the system can self-reconfigure in real-time, slashing request rejection rates from nearly 10% to virtually zero.
The Problem: Manual Reconfiguration in a Volatile Web
The core challenge is the volatility of web traffic. Traditional clusters often rely on a human administrator to monitor load and manually move resources—such as virtual or physical servers—between different service classes.
The authors build upon an existing platform called WS-DSAC (Web Servers - Differentiated Services Admission Control). While WS-DSAC is effective at balancing loads and prioritizing traffic using "shared," "exclusive," and "saturated" work modes, it suffers from two major flaws: it requires constant human oversight, and if a cluster hits "saturated" mode without available "shared" resources, it simply rejects requests.
DARC aims to bridge this gap by creating a self-configuring system that learns from the environment and reallocates resources before the system hits a breaking point.
Multi-Agent System Architecture
The researchers designed a hierarchical multi-agent system (MAS) organized into three distinct levels to manage the complexity of real-time reconfiguration:
- Strategic Level (Management Layer): Managed by a Coordinator Agent that oversees the global state and long-term strategy.
- Tactical Level (Monitoring & Execution Sub-layers): Includes Monitoring Agents that track CPU load and Execution Agents (MaximumLoad, DynamicThreshold, and UpdateManager) that decide when to move a server or adjust a cluster's capacity.
- Operational Level (Communication Layer): Handled by Communication Agents that interface directly with the WS-DSAC platform to pull metrics like the Reactivity Coefficient (RC)—a measure of average task waiting time.
The system uses a logic-based strategy: if a cluster's load exceeds a boundary (e.g., 80% of its limit), the MaximumLoad Agent attempts to pull a server from a less-loaded cluster. If that is insufficient, the DynamicThreshold Agent proactively increases capacity thresholds to prevent the cluster from entering "saturated" mode.
Methodology and Implementation
To validate the design, the authors modeled the entire architecture using Colored Petri Nets (CPN), a graphical discrete-event modeling tool, before implementing the final system in Java.
The researchers utilized a combination of simulated traffic generated via CPNTools and a real-world experimental setup. The physical environment consisted of heterogeneous hardware (AMD Athlon 64 and Intel Pentium IV nodes) connected via a 100 Mb Ethernet switch running Apache Tomcat 5. They used Webserver Stress Tools 7 Professional Edition to emulate real-world HTTP request patterns.
Key Results and Insights
The results from both simulations and real-world experiments were stark:
- Rejection Rates: In a standard WS-DSAC setup, the average request rejection rate was roughly 3.93% to 9.8% during peak loads. With DARC enabled, this dropped to between 0% and 0.05%.
- Resource Utilization: DARC allowed clusters to handle higher Reactivity Coefficients (RC=510 vs RC=450) before failing, meaning the hardware was being pushed to its actual potential rather than being throttled by rigid software limits.
- Proactive Learning: The UpdateManager Agent proved vital by learning from past behavior. If it detected frequent, oscillating threshold updates, it adjusted thresholds more aggressively to stabilize the system.
Conclusion
While DARC is highly effective, the authors acknowledge constraints such as a reliance on the Coordinator Agent and the need for hard threshold caps to prevent server meltdown. However, the success of the 10% rule—where redirecting just 10% of resources from a healthy cluster often prevents saturation elsewhere—proves that intelligent, automated reconfiguration is a viable path forward for high-availability web services.
Reference: Marques, C., Oliveira, I., Barroso, G., & Serra, A. (2011). A Dynamic Architecture for Reconfiguration of Web Servers Clusters. International Journal of Computer Networks & Communications (IJCNC), 3(3), 113-131.
Summary updated