1. Efficient agent-based selection of DiffServ SLAs over MPLS networks Thanasis G. Papaioannou a,b, Stelios Sartzetakis a, and George D. Stamoulis a,b presented by Vasilios Siris a SPIE International Symposium on Information Technologies: Internet Performance and Control, Boston, U.S.A., November 2000 a Institute of Computer Science (ICS), Foundation for Research & Technology – Hellas (FORTH) P.O. Box 1385, GR 711 10, Heraklion, Crete, Greece b Department of Computer Science, University of Crete, Heraklion, Greece

2. 2 Outline l Introduction and problem definition l Our contribution l Overview of the architecture l The SLS selection process l Implementation issues l Theoretical and experimental assessment of the economic efficiency achieved. l Conclusions – future work

3. 3 Introduction and Problem Definition l QoS provision in Internet is (and will be) necessary. l Many QoS protocols and mechanisms have been proposed. è SLA negotiation algorithms have to be defined properly, in order for ISPs to sell such contracts that: n satisfy user needs, n improve network efficiency.

4. 4 Our Contribution l Development and assessment of an architecture for per-flow SLA negotiation, provision and control deployment in a DiffServ over MPLS network domain. l Development and assessment of an efficient SLS selection process. l Implementation of the whole system in an experimental testbed.

5. 5 Overview of the Architecture Policy Server User Agent SLS negotiation LSP QoS requests DiffServ over MPLS network domain Policy Directory Information Directory

6. 6 QoS Issues l Each QoS class has n the same performance characteristics over all paths, n a certain non-compliance risk r. l Non-compliance risk is defined as an a priori upper bound on the percentage of traffic that will not be treated in accordance to the SLS. n Computable by the network provider. n Understandable for the end-users. l User QoS requirements: 1.maximum acceptable non-compliance risk. 2.minimum acceptable QoS class.

7. 7 The Efficient SLS Selection Process l The user application places a QoS request. l This request reaches the Policy Server, which discovers the feasible SLSs [x = (h, ρ, β, QoS class, r)] and their associated expected charge. l The User Agent selects the SLS x that maximizes the net benefit of the user, i.e.

8. 8 Charging by the Most Congested Link n s i, t i characterize the operating point of a QoS class i. n p i is the price per unit of effective usage in a QoS class i. n T is the service duration. l This scheme is fair and provides users with the right incentives for resource usage. l The PS computes the expected charge according to [Courcoubetis-Siris]: è simple bound of effective bandwidth

9. 9 Optimization of Traffic Parameters l For a given peak rate h (or a shaping delay) there is an indifference curve of (ρ, β) pairs for which all the inserted traffic is conformant. l For charge minimization  Minimize H(t) over this curve. l Assumption: PS offers only one optimized contract x, per QoS class. Negotiation now simplified: selection of x reduces to selection of QoS class and r.

10. 10 The User Utility Function l The user utility for a SLS x: n U(QoS DF ) is a normalized user utility function of the QoS class. n W(x) expresses the willingness to pay of a user for the QoS class, when for the highest QoS class he is willing to pay W max. n r user is the upper bound of the noncompliance risk that the user can accept, while r netw is the noncompliance risk offered by the network. n The function f expresses the user satisfaction for low values of r netw relatively to r user.

11. 11 A Typical Function U(QoS DF ) Minimum Acceptable QoS class

12. 12 A Reasonable Function  (r user, r netw )

13. 13 Distribution and Exchange of Information l Each component stores information for which it has the proper incentive.

14. 14 Efficiency of Architecture l Performs traffic categorization in QoS classes, rather than CAC. l In the network ingress n Clear and effective SLS negotiation interface. n Per flow traffic classification, control of compliance with the SLSs, and DS assignment to the packets. l Processing of QoS requests and state storage only by the PS. l Simple expression of user QoS requirements. l Low computational overhead and minor additions to the existing infrastructure. l Scaleable architecture in large administrative domain, using multiple PS instances.

15. 15 Economic Efficiency l Each user performs individual optimization. n Is incentive compatibility maintained when users employ this negotiation process for SLS selection? I.e. is social welfare also promoted? l Experiments: Computation of social welfare when employing: 1.our negotiation approach. 2.equal sharing of the same total amount of resources. l Results: social welfare always better under our approach.

16. 16 Conclusions – Future Work l Defined and implemented n an architecture for control and categorization of the inserted traffic in a DiffServ over MPLS network domain. n an efficient SLS selection process, providing the right incentives to users. l The whole system can serve as n a framework for the employment of different resource allocation and charging policies for the improvement of the overall efficiency of the network. l Directions for future work: Extend the system n for negotiation of aggregate SLSs between DiffServ domains. n for end-to-end SLA deployment across multiple DiffServ domains.

17. 17 Support Slides

18. 18 Experimental Results è Our approach is better for all the charging curves and for all the amounts that the users are willing to pay.

19. 19 DiffServ over MPLS Network

20. 20 Formulae in Experiments