Quality of Service in Internet

Quality of Service in InternetIntroduction1. With the increasing importance of Internet services, it is imperative for companiesrelying on web-based technology to offer predictable, consistent, as well asdifferentiated quality of service to their consumers. Internet services are commonlyhosted using clustered architectures where a number of machines, rather than a singleserver, work together in a distributed and parallel manner to serve requests to allinterested clients. Implementing service quality guarantees scalable in such adistributed setting is a difficult challenge. Traditional approaches to solving this QoSchallenge treat the problem as a capacity planning matter and rely on over-provisioningthe resources and on physically partitioning groups of cluster nodes for different classesof service. Unfortunately, the necessity to handle enormous and unpredictablefluctuations in load results in these techniques suffering from high cost and low resourceutilization. Moreover, such a static approach does not offer much flexibility, requiringhardware reconfiguration for any modification of the QoS policy or change in therecurring load pattern. Furthermore, although these hardware-based approaches cancertainly improve the quality of the service, they cannot provide QoS guarantees unlesseach of the partitions are always kept from being overloaded. As a result, softwarebasedapproaches have been suggested that embed the QoS logic into the internalsFig1: System Model for Internet Servicesof the operating system, distributed middleware or application code running on thecluster. Operating System techniques have been shown to provide a tight control on theutilization of resources (e.g., disk bandwidth or processor usage) while techniques that

2are closer to the application layer are able to satisfy QoS requirements that are moreimportant to the clients. However the majority of existing software approaches offerguarantees within the scope of a single machine or for an individual application, and failto provide global service quality throughout the site. Most current Internet sites arecomposed of a myriad of different hardware and software platforms which areconstantly evolving and changing. The largest drawback to software-based approachesis the high cost and complexity of reprogramming, maintaining, and extending theentirety of the complex software system such that it can provide QoS guarantees for allhosted services. Moreover, many components of a service are commonly third-party,proprietary software (e.g., Commercial Databases, Application Servers, etc) for whichthe source code is not available. Approaches that rely on embedding QoS supportdirectly into each application are difficult to implement in these cases. When theapplications themselves have embedded QoS support, it is often a mechanism that isunique to a particular application which makes enabling QoS and interoperability verydifficult. Therefore, QoS should be provided through the use of traffic shaping andadmission control techniques at the entrance of Internet sites.WHAT IS QUALITY OF SERVICE2. Quality of Service (QoS) is a set of service requirements to be met by thenetwork while transporting a packet stream from a source to a destination (unicast ormulticast). The common QoS parameters used in Internet services are• bandwidth: the average usable and available bandwidth over the link at anytime;• delay: the average end-to-end delay caused at network level at any time;• delay jitter: the average difference of the various delay times over the link;• packet loss probability: the average probability of packet loss over the link overa length of time.Different applications have different requirements for bandwidth, delay, and packet lossprobability. Therefore, QoS must also be a measurable level of service delivered tonetwork users. Such QoS can be provided by network service providers in terms of anagreement, known as a Service Level Agreement (SLA) made between network usersand service providers. QoS features provide the ability to manage traffic intelligentlyacross a network. Some applications require bandwidth and delay guarantees, referredto as quantitative applications, while others are more qualitative. For example, voiceapplications have stringent delay requirements and can tolerate minimal packet loss.Conversely, an FTP file transfer may be insentitive to delays but very sensitive to packetdrops. To accommodate these differences, QoS generally assigns traffic flows to one oftwo categories:• Guaranteed service (quantitive) reserves a designated amount of bandwidthfrom end to end and can guarantee a specified delay tolerance for the exclusive

3use of an application or even aggregated sessions. This type of service is bestfor applications that require a specific amount of bandwidth with long duration e.g.a video conference call may require at least 64 kbps for 30 minutes.• Differentitated service (qualitative) provides simple prioritisation. Applicationsare detected at the ingress and assigned SLAs. It is the SLA that drive whichQoS mechanisms the router will use, for example, which queue will be used toplace traffic, and, if congestion requires packets to be dropped, which droppriority will be designated.Integrated Services (IntServ)3. Its fundamental philosophy is that routers need to reserve resources in order toprovide quantifiable QoS for specific traffic flows. RSVP (Resource ReservationProtocol) serves as a signaling protocol for application to reserve network resources.Fig2: RSVP signalingRSVP adopts a receiver-initiated reservation style which is designed for a multicastenvironment and accommodates heterogenous receiver service needs. The flow sourcesends a PATH message to the intended flow receiver(s), specifying the characteristic ofthe traffic. As the PATH message propagates towards the receiver(s), each networkrouter along the way records path characteristics such as available bandwidth. Uponreceiving a PATH message, the receiver responds with a RESV message to requestresources along the path recorded in the PATH message in reverse order from thesender to the receiver. Intermediate routers can accept or reject the request of theRESV message. If the request is accepted, link bandwidth and buffer space areallocated for the flow, and the flow-specific state information is installed in the routers.Reservations can be shared along branches of the multicast delivery trees. RSVP takesthe soft state approach, which regards the flow-specific reservation state at routers ascached information that is installed temporarily and should be periodically refreshed bythe end hosts. State that is not refreshed is removed after a timeout period. If the route

4changes, the refresh messages automatically install the necessary state along the newroute. The soft state approach helps RSVP to minimize the complexity of connectionsetup and improves robustness, but it can lead to increased flow setup times andmessage overhead. The IntServ architecture adds two service classes to the existingbest-effort model, guaranteed service and controlled load service. Guaranteed serviceprovides an upper bound on end-to-end queuing delay. This service model is aimed tosupport applications with hard real-time requirements. Controlled-load service providesa quality of service similar to best-effort service in an underutilized network, with almostno loss and delay. It is aimed to share the aggregate bandwidth among multiple trafficstreams in a controlled way under overload condition. By using per-flow resourcereservation, IntServ can deliver fine-grained QoS guarantees. However, introducingflow-specific state in the routers represents a fundamental change to the currentInternet architecture. Particularly in the Internet backbone, where a hundred thousandflows may be present, this may be difficult to manage, as a router may need to maintaina separate queue for each flow.Differentiated Services (DiffServ)4. To address some of the problems associated with IntServ, differentiated services(DiffServ) has been proposed which is a per-aggregate-class based servicediscrimination framework using packet tagging. Packet tagging uses bits in the packetheader to mark a packet for preferential treatment. DiffServ defines a byte as the DSfield, of which six bits make up the DSCP (Differentiated Service Code Point) field, andthe remaining two bits are unused. DiffServ uses DSCP to select the per-hop behavior(PHB) a packet experiences at each node. A PHB is an externally observable packetforwarding treatment which is usually specified in a relative format compared to otherPHBs, such as relative weight for sharing bandwidth or relative priority for dropping. Themapping of DSCPs to PHBs at each node is not fixed. Before a packet enters a DiffServdomain, its DSCP field is marked by the end-host or the first-hop router according to theservice quality the packet is required and entitled to receive. Within the Diff-Serv domain,each router only needs to look at DSCP to decide the proper treatment for the packet.No complex classification or per-flow state is needed. DiffServ has two important designprinciples, namely pushing complexity to the network boundary and the separation ofpolicy and supporting mechanisms. The network boundary refers to application hosts,leaf (or first hop) routers, and edge routers. A network boundary has relative smallnumber of flows, whereas core router may have a larger number of flows, it shouldperform fast and simple operations. The differentiation of network boundary and corerouters is vital for the scalability of Diff- Serv. The separation of control policy andsupporting mechanisms allows these to evolve independently. DiffServ only definesseveral per-hop packet forwarding behaviors (PHBs) as the basic building blocks forQoS provisioning, and leaves the control policy as an issue for further work. The controlpolicy can be changed as needed, but the supporting PHBs should be kept relativelystable. The separation of these two components is key for the flexibility of DiffServ.

5Data Path Mechanisms5. Data path mechanisms are the basic building blocks on which Internet QoS isbuilt. They implement the actions that routers need to take on individual packets, inorder to enforce different levels of service. Control path mechanisms are concerned withconfiguration of network nodes with respect to which packets get special treatment whatkind of rules are to be applied to the use of resources. They are closely related, but theyaddress rather different performance issues. Queue management controls the length ofpacket queues by dropping or marking packets when necessary or appropriate, whilescheduling determines which packet to send next and is used primarily to manage theallocation of bandwidth among flows.Fig3: Basic Data Path OperationBasic Packet Forwarding Operation6. As a packet is received, a packet classifier determines which flow or class itbelongs to based on the content of some portion of the packet header according tocertain specified rules. There are two types of classification viz General classificationand Bit-pattern Classification. General classification performs a transport-levelsignature-matching based on a tuple in the packet header. It is a processing-intensiveoperation. This function is needed at any IntServ-capable router. In DiffServ, it isreferred as multifield (MF) classification, and it is needed only at network boundary. Bitpattern Classification sorts packet based on only one field in the packet header. It ismuch simpler and faster than general classification. In DiffServ, it is referred as behavioraggregate (BA) classification which is based only on DS field. It is used at network corerouters. After classification, the packet is passed to a logical instance of a trafficconditioner which may contain a meter, marker, shaper, and dropper. A marker markscertain field in the packet, such as DS field, to label the packet type for differentialtreatment later. A meter is used to measure the temporal properties of the traffic streamagainst a traffic profile. It decides that the packet is in profile or out of profile, and then itpasses the state information to other traffic conditioning elements. Out of profile packetsmay be dropped, remarked for a different service, or held in a shaper temporarily until

6they become in profile. In profile packets are put in different service queues for furtherprocessing. A shaper is to delay some or all of packets in a packet stream in order tobring the stream into compliance with its traffic profile. It usually has a finite buffer, andpackets may be discarded if there is insufficient buffer space to hold the delayedpackets. A dropper can be implemented as a special case of a shaper by setting shaperbuffer size to zero packets. It just drops out-of-profile packet.Queue Management7. One goal of Internet QoS is to control packet loss. It is achieved mainly throughqueue management. Packets get lost for two reasons: damaged in transit or droppedwhen network congested. Loss due to damage is rare, so packet loss is often a signal ofnetwork congestion. To control and avoid network congestion, we need somemechanisms both at network end-points and at intermediate routers. At network endpoints,we depend on the TCP protocol which uses adaptive algorithms such as slowstart, additive increase and multiplicative decrease. Inside routers, queue managementis used. Our goals are to achieve high throughput and low delay. The effectiveness canbe measured by network power, which is the ratio of throughput to delay. The bufferspace in the network is designed to absorb short term data bursts rather than becontinuously occupied. Limiting the queue size can help to reduce the packet delaybound. Traditionally, packets are dropped only when the queue is full.Scheduling8. Packet delay control is an important goal of Internet QoS. Packet delay has threeparts: propagation, transmission, and queuing delay. Propagation delay is given by thedistance, the medium and the speed of light, roughly 5 s/km. The per-hop transmissiondelay is given by the packet size divided by the link bandwidth. The queuing delay is thewaiting time that a packet spends in a queue before it is transmitted.Fig4: Policy Architecture

7This delay is determined mainly by the scheduling policy. Besides delay control, linksharing is another important goal of scheduling. The aggregate bandwidth of a link canbe shared among multiple entities, such as different organizations, multiple protocols(TCP, UDP), or multiple services (FTP, telnet, real-time streams). An overloaded linkshould be shared in a controlled way, while an idle link can be used in any proportion.Although providing delay guarantee and rate guarantee, are crucial for scheduling,scheduling needs to be kept simple since it needs to be performed at packet arrivalrates. There is variety of scheduling disciplines. First Come First Serve (FCFS) is thesimplest scheduling policy. It has no flow or class differentiation, no delay or rateguarantee. Priority scheduling provides a separate queue for each priority class.Basically, it is a multiple-queue FCFS scheduling discipline with the higher priorityqueue being served first.Control Path Mechanisms9. Admission control implements the decision algorithm that a router or host uses todetermine whether a new traffic stream can be admitted without impacting QoSassurances granted earlier. As each traffic stream needs certain amount of networkresources (link bandwidth and router buffer space) for transferring data from source todestination, admission control is used to control the network resource allocation. Thereare three basic approaches for admission control: deterministic, statistic, andmeasurement-based. The first two use a prior estimation, while the later one is basedon the current measurement of some criteria parameters. The deterministic approachuses a worst-case calculation which disallows any QoS violation. It is acceptable forsmooth traffic flows, but it is inefficient for burst flows and leads to a lower resourceutilization. Both statistical and measurement-based approach allows a small probabilityof occasional QoS violation to achieve high resource utilization.10. Policy Control specifies the regulation of access to network resources andservices based on administrative criteria. Policies control which users, applications, orhosts should have access to which resources and services and under what conditions.Instead of configuring individual network devices, ISPs and corporate administratorswould like to regulate the network through policy infrastructure, which provide supportsfor allowing administrative intentions to be translated into differential packet treatment oftraffic flows. The policy server has access to a policy database as well as authorizationand accounting databases. Each policy entry specifies a rule of “if certain conditionhappens, then take certain action”. A human network operator working at amanagement console would use a GUI management application which interfaces to thepolicy server through a set of Policy API (PAPI). This allows the operator to update andmonitor policy changes in the policy database. The policy server consists of a centralpolicy controller (CPC) and a set of policy decision points (PDP). PDPs are responsiblefor determining which actions are applicable to which packets. The CPC is to ensureglobal consistency between decisions made by the PDPs. The enforcement andexecution of policy actions are done by policy enforcement points (PEP). PEPs aretypically co-located with packet forward components, such as border routers. PDPs

8interact with PEPs via Common Open Policy Service (COPS). PDPs push configurationinformation down to the PEP’s as well as respond to queries from the PEP’s.Bandwidth Brokers11. A bandwidth broker is a logical resource management entity that allocates intradomainresources and arranges inter-domain agreements. A bandwidth broker for eachdomain can be configured with organizational policies and controls the operations ofedge routers. In the view of policy framework, a bandwidth broker includes the functionof PDP and policy database, while edge routers serve as PEPs. In its inter-domain role,a bandwidth broker negotiates with its neighbor domains, sets up bilateral agreementwith each of them, and sends the appropriate configuration parameters to the domain’sedge routers. Bilateral agreement means that a bandwidth broke only needs tocoordinate with its adjacent domains. End-to-end QoS is provided by the concatenationof these bilateral agreements across domains, together with adequate intra-domainresource allocation. Within a domain, a bandwidth broker performs resource allocationthrough admission control.Routing and QoSFig5: Bandwidth Broker12. Traffic engineering arranges traffic flows so that congestion caused by unevennetwork utilization can be avoided. Constraint-based routing is an extension to thebasic topology-based routing. It routes packets based on multiple constraints.Constrains include network topology (shortest path), network resource availabilityinformation (mainly link available bandwidth), flow QoS requirements, and policyconstraints. Constraint-based routing can help to provide better performance andimprove network utilization. But it is much more complex, may consume more networkresource and may lead to potential routing instability.

913. MPLS offers an alternative to IP-level QOS. An MPLS packet has a header thatis sandwiched between the link layer header and the network layer header. When apacket enters an MPLS domain, it is assigned an MPLS label which specifies the paththe packet is to take while inside the MPLS domain. Throughout the interior of theMPLS domain, each MPLS router switches the packet to the outgoing interface basedonly on its MPLS label. At the same time, the packet gets marked with a new label priorto transmission. The COS field is used to choose the correct service queue on theoutgoing interface. At the egress to the MPLS domain, the MPLS header is removedand the packet is sent on its way using normal IP routing. MPLS is a label-basedmessage forwarding mechanism. By using labels, it can set up explicit routes within anMPLS domain. A packet’s forwarding path is completely determined by its MPLS label.If a packet crosses all MPLS domains, an end-to-end explicit path can be establishedfor the packet. Label also serves as a faster and efficient method for packetclassification and forwarding. MPLS also to route multiple network layer protocols withinthe same network and can be used as an efficient tunneling mechanism to implementtraffic engineering.End Host Support for QoS14. There is an increasing need for server QoS mechanisms to provide overloadprotection and to enable tiered (or differentiated) service support. A request manager isused to intercept all requests. It classifies the requests and places the requests in theappropriate queue. To ensure that server is not overloaded, admission control is usedby request manager. Request classification and admission control can be based on avariety of policies. After the requests are put on the appropriate queue, a schedulingprocess determines the order in which the requests are to be served, and feeds thechosen requests to the server. Similar to the scheduling in network routers, differentpolicies can be adopted, such as strict priority, weighted fair queuing, or earliestdeadline first.Fig6: Server QoS

10Application Adaptation15. Instead of indicating requirements to the network via resource reservationmechanisms, applications can adapt their rate to the changing delays, packet loss andavailable bandwidth in the network. Play out delay compensation allows applications tofunction with the lowest overall delay even as the network delay changes. A play outbuffer is a queue for arriving packets emptied at the playback rate. It converts avariable-delay into a fixed delay, and it has to be large enough to compensate for thedelay jitter. For loss adaptation, several techniques have been explored, such asredundant transmission, interleaving, and forward error correction. Bandwidthadaptation depends on the media. For audio, applications can change the audio codec,while video applications can adjust the frame rate, image size and quality.Conclusions16. Scalability is a fundamental requirement for any Internet QoS scheme. It is anissue that impacts both data path and control path, including flow and queuemanagement, network device configuration, accounting and billing, authorization,monitoring, and policy enforcement. Aggregation is one common solution to scalingproblems, but it comes at the price of looser guarantees and coarser monitoring andcontrol. By separating of control policy from data forwarding mechanisms, we can havea set of relative stable mechanisms on top of which Internet QoS can be built. At thesame time, we can easily adjust or add any control policy whenever needed with asimple re-configuration or remapping from policy to mechanisms. All supportingmechanisms can be kept unchanged. Since the Internet comprises multipleadministrative domains, inter-domain QoS and intra-domain QoS can be designedseparately. At the inter-domain level, pure bilateral agreement based on trafficaggregate is used. Within each domain, variety of different choices can be adopted. Theconcatenation of domain-to-domain data forwarding provides end-to-end QoS delivery.Although important technical progress has been made, much work needs to be donebefore QoS mechanisms will be widely deployed. In general, pricing and billing are theprimary issues that need to be resolved. Once services are billed for, customers willneed to be able to monitor that the purchased service is meeting specifications. Whileadaptive applications have been built to respond to changes in network performance,integrating a mechanism to adapt to price changes over time adds yet anotherdimension. Incorporating wireless transmission into the Internet adds additional QoSissues, such as mobile setup, limited bandwidth, and hand over.