To Appear: Proc. IEEE Infocom'95, Boston, April 1995.
A Service with Bounded Degradation in Quality-of-Service
Jorg Liebeherr Dongwei Liao
Department of Computer Science
University of Virginia
Charlottesville, VA 22903
Many network applications that require Quality-of-
Service (QoS) support, such as transmission of digital
voice and video, tolerate a certain level of service degra-
dation. In this study, a novel network service, referred
to as service with bounded degradation (BD service),
is presented that can take advantage of the delay tol-
erance of applications. Dierent from previous propos-
als for similar services, the BD service can guarantee
hard lower bounds of the service degradation. This is
achieved by strictly limiting the amount of trac that
is subject to service degradation. To implement a BD
service, network trac is partitioned into components
with dierent delay requirements. Policing of the traf-
c components is achieved by multi-level leaky buckets.
A novel scheduler, referred to as EDD-BD (Earliest
Deadline Due - Bounded Degradation) scheduler, per-
forms switching of packets. Properties of the EDD-BD
scheduler are presented and analyzed.
Current computer network technology can provide veryhigh bandwidth up to several hundred megabits persecond. The dramatic increase in the capacity of com-munication links enables the design of packet-switchingnetworks which provide performance guarantees totrac that is carried by the network. The set of perfor-mance guarantees given to a network connection is re-ferred to as quality-of-service (QoS). Performance guar-antees are specied in terms of network bandwidth,network delay, network delay jitter, and loss probabil-ity. We refer to a packet-switching network that canprovide quality-of-service to connections as QoS net-work.
In recent years, several researchers have proposedarchitectures for connection-oriented QoS networks
This work is supported in part by the National Science Foun-
dation under Grant No. NCR-9309224.
[1, 2, 3]. In all proposals the concept of quality-of-service is quite similar. When requesting a new con-nection a client submits to the network a specicationof its trac and the requested QoS. The network ver-ies if the requested QoS can be guaranteed. If thenetwork is not able to support the QoS requirements itrejects the establishment of the connection, otherwisethe connection is established. After connection estab-lishment the network must support the admitted QoSuntil the connection is released.
Since many network clients do not require that ser-vice commitments be guaranteed for all packets of aconnection, a QoS network architecture oers severallevels of service commitments. At the highest level, aso-called deterministic service  guarantees that per-formance requirements are met for all packets at alltimes. Since this type of service commitment mustsupport the requested QoS even during possibly rareperiods of congestion, the amount of network resourcesthat must be allocated for a deterministic service typi-cally exceeds by far the average resource requirements.At the lowest level of service commitments, the networkdoes not guarantee any QoS requirements at all. Thislevel of service is commonly referred to as best-eortservice. The network does not reserve any resourcesfor a connection with a best-eort commitment.1 Also,best-eort trac can be subject to any form of servicedegradation, that is, packet delays can be arbitrarilylong or packets may be dropped.
Most network clients which require performanceguarantees can tolerate a certain level of service degra-dation. As an example, transmission of digitized voicedoes not need reliable delivery, but can tolerate a lim-ited number of packet losses. However, the amountof lost voice packets must be within application spe-cic bounds; otherwise, the voice stream will be notintelligible at the receiver . Therefore, for voice ap-plications neither the highest nor the lowest level of
1Note that the network may, however, reserve certain re-
sources, e.g., link bandwidth, for the class of all best-eort
performance guarantees is ideal. The highest level ofperformance guarantees does not permit any packetlosses at all, and thus, does not allow the network totake advantage of the loss tolerance of voice applica-tions. On the other hand, a best-eort service is notable to support the minimal service requirement forpacket voice applications, that is, intelligibility of voiceat the receiver.
All architectures for QoS networks proposed in theliterature oer a service with commitments that areweaker than a deterministic service, but stronger thana best eort service. The statistical channel in theTenet protocol suite  permits network clients to spec-ify probabilistic bounds on the percentage of delayedor lost packets. Probabilistically bounded QoS guaran-tees, similar to the statistical channel, have been stud-ied in [3, 5]. The architecture proposed in  proposesa so-called predictive service which adapts QoS guar-antees to the load conditions in the network.
A drawback of all existing proposals for QoS ser-vice commitments which are neither deterministic, norbest-eort, is that they cannot guarantee hard lowerbounds on the degree of service degradation as com-pared to a deterministic service. Note that any prob-abilistic service guarantees are by denition not veri-able in a nite time interval. Therefore, a probabilisticQoS guarantee can be satised even if the QoS guaran-tee is violated in a time interval of nite, but otherwisearbitrary, length. On the other hand, a service whichadapts the QoS guarantees to the currently prevailingnetwork load may degrade to a best eort service with-out violating any service commitments.
In this study, we propose a new network service thatenables a specication of degraded quality-of-servicecommitments for a xed portion of trac from a con-nection. The advantage of our scheme over previousproposals for similar services is that the degree of ser-vice degradation can be veried at all times. Sinceservice degradation is bounded to a specied portionof the trac we refer to our scheme as service withbounded degradation (BD service). For the implemen-tation of the new service we require network clients tocharacterize the amount of trac which can receive alower grade of service. Also, a network client must ex-plicitly specify the performance bounds for the portionof trac that can be subject to service degradation.
There are a number of problems that must be solvedto make a service with bounded degradation feasible.One problem is illustrated in the following example.Assume a connection wishes to transmit packetizedvoice over a QoS network. The voice connection mayrequest a delay bound of 100 ms for 90% of its traf-
c, and a delay bound of 200 ms for 10% of its trac.Naively, the network could separate the trac streaminto two components, and provide delay bounds of 100ms and 200 ms to the respective components. In thiscase, it is feasible that packets from the high delay com-ponent will catch up with, and possibly pass, packetsfrom the low delay component. As a result, packetsmay arrive at the receiver in a dierent order than theywere transmitted. To overcome this out-of-sequenceproblem we propose an extension to traditional packetscheduling mechanisms which guarantees in sequencedelivery of packets, yet, maintains dierent levels ofperformance commitments to dierent portions of thetrac.
The remaining paper is structured as follows. Westate our assumptions on the network trac in Sec-tion 2. In Section 3 we present a modied packet sched-uler which provides dierent delay bounds to dierentportions of the trac from a single connection, yet,maintains the sequence of the packet stream. In Sec-tion 4 we investigate properties of the modied packetscheduler. These properties are used in Section 5 wherewe obtain analytical results on the BD service. Theseresults are applied in an example in Section 6. In Sec-tion 7 we conclude our results.
2 Trac Model
In a QoS architecture, a network client submits witha request for a new connection (a) a set of parameterswhich characterize the trac of the connection, and(b) the required QoS guarantees. Once the connec-tion is established, that is, admission control functionshave determined that the performance guarantees canbe maintained for the trac as characterized by thenetwork client, so-called policing functions at the en-trance to the network enforce that all trac conformsto the specied characterization.
Throughout this study, we assume that trac froma connection is characterized by two parameters (b; T )where b denotes the maximum number of packets sub-mitted to the network at any point in time (burst fac-tor), and 1=T denotes the maximum average rate atwhich packets are generated (rate factor). We assumethat packets are xed-sized.
If a network client requests the establishment of aconnection with bounded service degradation (BD con-nection) we require that the client characterizes theportion of trac which may be subject to service degra-dation. That is, a client characterizes the connectiontrac with two sets of parameters (bp; Tp) and (bs; Ts).The sets describe the primary component and the sec-