Service Level Agreement-based management (Tyche)

Introduction Service Level Agreements (SLAs) capture the agreed-upon guarantees between a service provider and its customer. In today's complex and dynamic computing environments, service providers have to constantly trade off between guaranteeing service levels and the service infrastructure requirements.

Related links » Data center automation research » IT discovery

It is thus important to determine which Service Level Agreements the service provider must enter into in the first place. The service provider must be cognizant of its objectives, customer utility and the effort involved when agreeing to SLAs. We term this the SLA determination problem. It involves a careful and deliberate decision-making about determining optimal SLAs.

Once a service provider has agreed to an SLA, it is imperative to ensure that the constraints imposed on service level metrics are being met. It is thus important to provision the service infrastructure in such a way that it is able to meet the desired goals. Also it is important to monitor the system thereafter to continually meet the constraints.

Target area: Performance, reliability, availability

	SLA Determination and Negotiation

The primary model in the SLA determination problem is an adaptation of the moral hazard models on the contract theory literature to the problem in question -- that is, the design of SLAs. The basic idea is that the provider, through some costly effort (investment, use of scarce resources), can increase the quality of the service, but that there is also a stochastic component to it. Bigger effort increases the probability of a good quality outcome, but this is still not a certain thing.

In the context of SLAs this makes sense: Better infrastructure will give in average better performance, but some random elements (extra demand, breakdown of a system, etc.) can still worsen the quality. The critical aspect of this situation is that the client can observe quality, but not effort. Therefore, the only way to induce a high level of effort is through a compensation system that is “steep,” so it pays more when observed quality is better.

The end goal of the work is to design an optimal menu of contract that meets a certain level of customer utility, maximizes the service provider utility and meets certain quality guarantees in terms of service levels. We also extend this notion to involve strategy-based negotiation around the menu of contracts.

	SLA Decomposition

SLA decomposition enables creating low-level policies from high-level goals. The intent here is to take high-level goals and derive or infer low-level policies that can then be validated to asses the system performance. Traditionally, design of systems and determination of thresholds of metrics for the purpose of monitoring has been done by domain experts. The SLA Decomposition problem is an attempt to automate the process of translating high-level requirements into lower level thresholds that may be used for sizing and/or monitoring.

We provide a general approach for calculating the bounds on system behavior given SLOs for the service. Our approach uses analytical models to capture the relationship between high level performance goals (e.g., response time of the overall system) and the refined goals for each component (e.g., average service time of each component).

	Team members

• Akhil Sahai (Principal Investigator)
• Yuan Chen
• Subu Iyer
• Zhongtang Cai (Intern, April-June 2006)
• Yathiraj Udupi (Intern, April-June 2006)
• Sofia Moroni (Intern, Feb-June 2007)
• Cristian Figueroa (Intern, June-Oct 2007)

	Collaborators

• Terence Kelly
• Alex Zhang
• Jerry Rolia
• Sharad Singhal
• Alejandro Joffre, Nicolas Figueroa (Universidad de Chile)
• Karsten Schwann, Vibhore Kumar (Georgia Institute of Technology)
• Calton Pu, Elba Team (Georgia Institute of Technology)
• Christopher Stewart (University of Rochester)

	Publications

Yathiraj Udupi, Akhil Sahai, Sharad Singhal. A Classification based approach to Policy Refinement. In the proceedings of IEEE/IFIP IM 2007. full version

Yuan Chen, Xue Liu, Subu Iyer, Akhil Sahai. SLA Decomposition: Translating Service Level Objectives to System Thresholds. In the proceedings of International Conference on Autonomic Computing (IEEE ICAC 2007). full version

Sofia Moroni, Alejandro Joffre, Nicolas Figueroa, Akhil Sahai, Yuan Chen, Subu Iyer. A Game-theoretic framework for Optimal SLA/Contract Creation. HPL Tech Report full version

Zhongtang Cai, Yuan Chen, Dejan Milojicic, Karsten Schwan. Automated Availability Management Driven By Business Policies. In the proceedings of IEEE/IFIP IM 2007. full version

Vibhore Kumar, Karsten Schwann, Subu Iyer, Yuan Chen, Akhil Sahai. A State Space Approach to SLA based Management. IEEE/IFIP NOMS 2008 (under submission). full version

» Download the latest Adobe Acrobat Reader