QoS- and revenue-aware resource allocation mechanisms in multiclass IP networks

Abstract

Zhang esittelee väitöskirjassaan uuden kuormanjakomekanismin, jolla voidaan muodostaa eriteltyjä palveluita klusteripohjaisissa palvelimissa. Mekanismin toimintaa on simuloitu erilaisissa tilanteissa ja se toimii hyvin myös korkeasti kuormitetuissa palvelimissa. Zhang on kehittänyt myös asteittaisen (kaksitasoisen) www-klusterirakenteen, jonka asteittaisuus on määriteltävissä erilaisissa käyttökohteissa.Zhang analysoi työssään myös sitä, miten palvelimen resurssit voidaan optimaalisesti eritellä palveluluokkien kesken niin, että voidaan maksimoida palvelulaatusopimuksen tuotot, kun esimerkiksi tarjotaan sähköisen kaupankäynnin palvelua. Tähän liittyen Zhang toteutti työssään optimaalisen resurssienjakosuunnitelman tietyillä palvelinresursseilla ja kiinteällä hinnoittelulla.

The substantial changes transforming the Internet from a communication and browsing infrastructure to a medium for conducting personal business and e-commerce are making Quality of Service (QoS) an increasingly critical issue. However, QoS of networks by itself is not sufficient to support end-to-end QoS. To avoid high priority network traffic being dropped at the server, Web servers should have mechanisms and policies for delivering end-to-end QoS. More importantly, in the future multiclass Internet, each class of customers may have to pay their service providers for the received level of QoS based on the Service-Level-Agreements negotiated and committed between them.In this dissertation, we first propose a novel Arrival-Related Dynamic Partitioning mechanism for enabling differentiated services in cluster-based Web server systems, which works well even when the Web cluster system is heavily loaded and does not have enough server resources to be allocated. Then, a scalable Web cluster architecture -- the two-level cluster architecture, is proposed for implementing scalable service differentiation in cluster-based Web server systems, whose scalability is theoretically determined only by the scalability of its layer-4 switch.Next, we link the issue of resource partitioning scheme with the pricing strategy in a Service-Level-Agreement (SLA) and address the problem of maximizing the SLA revenue obtained in a IP network node under a given amount of resources by optimally allocating the resources among the supported service classes. First, the revenue-aware resource allocation schemes, which can achieve the maximization of the SLA revenue obtained in a multiclass-supported network node under a given amount of network resources and linear pricing strategy, are proposed. Then we derive a novel upper bound on mean packet delay of GPS-based (Generalized Processor Sharing based) Fair Queueing (FQ) algorithms under the probabilistic traffic model of Poisson arrival and any general packet length distribution, which is much simpler and tighter than the known ones by M. Hawa et al. and fits a class of GPS-based FQ algorithms including Weighted Fair Queuing (WFQ), Self-Clocked Fair Queuing (SCFQ) and Starting Potential-based Fair Queuing (SPFQ). Furthermore, based on this novel upper delay bound, the suboptimal resource allocation scheme is presented for maximizing the SLA revenue attained under a constrained amount of network resources and flat pricing strategy in a network node which deploys a GPS-based packetized fair queueing algorithm.In the end, the problem of maximizing the SLA revenue attained for the hosting of an e-commerce Web site upon a cluster-based Web server system by optimally partitioning the server resources among all supported service classes is analyzed. The optimal resource partitioning scheme is derived, which can implement the maximization of the SLA revenue obtained for the hosting of an e-commerce site under a given amount of server resources and linear pricing strategy. Moreover, the suboptimal resource partitioning scheme is also proposed for achieving the highest SLA revenue in the hosting of an e-commerce site under a given amount of server resources and flat pricing strategy.