Title Page i

Attila Kertész, Gábor Kecskeméti, Ivona BrandićOpenalex Unknown 2011 · DOI · OpenAlex

Abstract

This paper addresses the challenges of managing Service Level Agreements (SLAs) in heterogeneous distributed systems. It covers topics including parallel and distributed processing, resource management, load balancing, and grid applications. The authors propose an autonomic architecture for SLA-based Service Virtualization (SSV) to automate negotiation, brokering, and on-demand deployment.

Introduction

While Cloud and Grid computing offer massive scalability, managing these environments to meet specific business guarantees—non-functional requirements like performance or availability—is notoriously difficult. These guarantees are usually codified in Service Level Agreements (SLAs).

Most existing Cloud middleware lacks a way to localize failures or automatically react when an SLA is at risk of being violated. In loosely coupled systems, it is often unclear whether a failure occurred at the negotiation stage, the brokering stage, or the physical resource level. Without an "autonomic" (self-managing) approach, keeping these systems stable requires constant, costly human oversight. This paper introduces SLA-based Service Virtualization (SSV), a three-layered autonomic architecture designed to ensure SLAs are met without manual intervention.

The SSV Architecture

The proposed SSV architecture consists of three primary layers governed by an Autonomic Manager (AM), which implements a MAPE (Monitor, Analyze, Plan, Execute) control loop.

The three-layered SSV architecture consisting of Meta-Negotiator, Meta-Broker, and Automatic Service Deployer.
  1. Meta-Negotiator (MN): Acts as the interface between the user and the system, negotiating SLA terms and translating them into technical requirements.
  2. Meta-Broker (MB): Selects the most suitable broker or resource provider based on historical performance and current availability.
  3. Automatic Service Deployer (ASD): Uses virtualization to install, deploy, or replicate services on-demand when existing resources cannot fulfill an SLA.

Sensors throughout the system feed data to the AM, which uses case-based reasoning to identify failure sources. If a violation is detected, the system triggers reactive actions, such as renegotiating terms or spinning up new virtual machine (VM) instances.

Dynamic SLA Enforcement Workflow

The core strength of the SSV architecture lies in its ability to dynamically adapt to changing conditions. When a user submits a request, the system evaluates the entire chain of delivery rather than just looking for a static resource.

Sequence diagram illustrating the autonomic workflow for SLA negotiation, brokering, and on-demand service deployment.

As shown in the sequence diagram, if QoS requirements cannot be met by existing brokers, the system can trigger renegotiation or switch to a target broker. If a service is unavailable, the Automatic Service Deployer can spin up a new instance to fulfill the request, ensuring the SLA is maintained even under unexpected load.

Experimental Evaluation

The researchers evaluated their architecture using a simulation-based approach combining CloudSim and GridSim. They used real-world traces from the Grid Workloads Archive (GWA) to simulate realistic background noise and resource contention.

Key results demonstrated that the SSV architecture significantly outperformed traditional Grid brokering:

  • SLA Attainment: The system successfully avoided violations by triggering on-demand deployments when grid background loads exceeded 113%.
  • Deployment Latency: Services for a biochemical case study (TINKER) could be ready in 6.9 to 8.3 seconds on average.
  • Scalability: Using a "service duplication" strategy significantly reduced overall execution time compared to static resource allocation.

Conclusion and Future Work

The SSV architecture demonstrates that coupling SLAs directly to infrastructure layers allows for precise failure localization and targeted recovery. While on-demand deployment introduces some latency overhead, it is essential for avoiding costly SLA violations during unexpected demand spikes. Future work will focus on moving from reactive actions to proactive actions, using predictive modeling to anticipate failures before they occur.

Summary updated

Abstract

The following topics are dealt with: parallel processing; distributed processing; network-based processing; scheduling; resource management; load balancing; performance estimation; programming abstraction; tools; Web computing; grid application; distributed bioinformatics; nuclear fusion; peer-to-peer environment; and security.
BibTeX
@inproceedings{kertesz2011title,
  author = {Attila Kertész and Gábor Kecskeméti and Ivona Brandić},
  title = {{Title Page i}},
  booktitle = {Openalex Unknown},
  year = {2011},
  doi = {10.1109/pdp.2011.1},
}
Figure