Report on the 5 th IFIP International Workshop

on Quality of Service (IWQOS'97)

M a y 21-23, 1997.
Center for T e l e c o m m u n i c a t i o n s Research
C o l u m b i a University, N e w York City
h t t p : //comet. ctr. c o l u m b i a , e d u / i w q o s 9 7 /
Oguz Angin, Andrew T. Campbell, Lai-Tee Cheok, Raymond R-F Liao,
Koon-Seng Lim and Klara Nahrstedt
workshop@ctr, columbia, edu
Abstract The 5th IFIP International Workshop on Quality of
Service was held at the Center for Telecommunications
This paper presents a summary of the fifth International Research, Columbia University, and is the latest in a
Workshop on Quality of Service (IWQOS) which was series of continuing workshops. The first workshop,
held at Columbia University in May 1997. The goal of held in May 1993 in Montreal, Canada, was supported
this three-day meeting was to foster interaction between by the European RACE project QOS TOPIC and the
researchers active in the area of Quality of Service Canadian CITR project Broadband Services. The
(QOS) research, to reflect on past experiences and second workshop, organized within the European
lessons learnt, and to discuss future QOS challenges. RACE Conference on Integrated Services and Networks
To reflect this goal, this year's workshop included a hot (IS&N), took place in September 1994 in Aachen,
program made up of (i) a keynote address on Germany. The third workshop was held in February
"Programming Telecommunications Networks"; (ii) 1995 in Brisbane, Australia, in conjunction with the
panels addressing "QOS for Distributed Object International IFIP Conference on Open Distributed
Computing Middleware - Fact or Fiction?" and Processing. In March 1996, the workshop was held in
"Reservations about Reservations"; (iii) a workshop Paris in conjunction with the IFIP/IEEE International
invited paper entitled, "Quality of Service - Where are Conference on Distributed Platforms.
we?" and (iv) ten technical sessions that included new
topics for IWQOS such as mobile communications, The theme for IWQOS'97 was "Building QOS into
QOS routing and QOS-based transport systems. This Distributed Systems". Implicit in the theme is the notion
report summarizes the technical program and captures that the QOS community should focus on discussing
the main themes and major areas of discussion that results from prototype implementations of their ideas.
emerged during IWQOS'97. We were naturally interested in assessing the impact of
ideas discussed at previous meetings on future products
as QOS ideas move from research to development.
1. Introduction While 1WQOS is interested irt experimental results, it
Over the past several years, there has been a remains a forum for the discussion of fresh and
considerable amount of research within the field of innovative ideas. As a result of this, authors were
Quality of Service (QOS). Much of the work has taken solicited to provide experimental research (long) papers
place within the context of QOS support for distributed and more speculative position (short) statements. The
multimedia systems, operating systems, transport technical program successfully reflected the organizers
subsystems, networks, devices and formal languages. desire to hear about experiment results, controversial
The objective of the International Workshop on Quality ideas, retrospectives and future directions.
of Service (IWQOS) is to bring together researchers,
developers and practitioners working in all these facets
of QOS research. While many conferences and
2. IWQOS'97 Program Outline
workshops offer technical sessions on the topic QOS, This year's workshop included an invited program,
none other than IWQOS, provides a single-track which comprised a keynote address, panels and a
workshop dedicated to the broad subject of QOS workshop invited paper. The keynote address given by
research. Aurel A. Lazar (Columbia University) was entitled

ACM SIGCOMM 100 Computer Communication Review

"Programming Telecommunications Networks". The Programmer Interfaces (APIs) provided by vendors in
invited panels addressed two aspects of delivering QOS this market would allow users to write basic
in distributed computing environments and the Internet. communication services and middleware components.
The first panel, chaired by Douglas Schmidt He considered that the second layer of the model
(Washington University), addressed the question: "QOS reflects a middleware service market where carriers,
for Distributed Object Computing Middleware - Fact or software developers and middleware service providers
Fiction?". The second workshop panel entitled offer middleware service products to customers who are
"Reservations about Reservations" and chaired by in the user service provisioning business. The APIs
Henning Schulzrinne (Columbia University), discussed provided in the middleware service market are suitable
the topic of QOS provision in the next generation for development of consumer level services. Finally, at
Internet. To complete the IWQOS'97 invited program, the highest layer he considered that the model reflects a
we had a reality check in the form of the workshop consumer services market where consumer service
invited paper by Ralf Steinmetz and Lars Wolf providers compete to bundle, integrate and customize
(Darmstadt University) on "Quality of Service - Where their wares in the most appealing form for mass market
are we?". consumption. Within each market there may exist
brokers whose role is to mediate the interaction and
A strong technical program that included twenty long
dealings between buyers and sellers who, because of
papers and twenty short papers complemented the
regulatory and business policies, cannot transact
IWQOS invited program, which were chosen from more
directly.
than seventy papers from nine countries.
The keynote speaker commented that this service model
falls somewhere between the Internet's peer-to-peer
3. Programming Telecommunications model and the Telephone Network's strict provider-
Networks customer model. In essence, it allows for cooperation
between any number of entities in the network for
Keynote Speaker: Aurel A. Lazar, Columbia University realizing a common service as well as the competition
Recent moves toward market deregulation and open among services for network resources. The
competition have sparked a wave of serious corresponding engineering model can be parameterized
introspection in the telecommunications service in such a way that the basic characteristics of the peer-
industry. Telecom providers and operators are now to-peer model as well as the characteristics of the
required to open up their primary revenue channels to provider and consumer model can be accommodated.
competing industries. In the keynote address, Aurel Within each layer (which models a particular market),
Lazar [Lazar,97] focussed on the problem of players are free to enter and buy, sell or re-bundle each
programmability of telecommunications networks for other's services. Across layers, the relationship reflects
new services. The speaker outlined an agenda for the traditional provider-customer model.
realizing an open programmable networking
environment based on the concept of a broadband Lazar argued that investigating such a model would
kernel that better reflects the service creation and help clarify some important issues facing the
deployment environment and economic concerns of telecommunications service industry as it deals with
future telecommunications systems. changes in service needs. An engineering model for
realizing the open service market model was then
The address began with an examination of the service presented as a vehicle for creating multimedia services
structure of two major global communication networks on broadband networks. An example of engineering
(i.e., the Telephone Network and the Internet) exploring some of the components of the open service model was
their relative strengths and weaknesses. Lazar proposed then presented from an implementation viewpoint.
a three-tiered open service model that reflects the
economic market structure of the future In an insightful presentation the speaker connected
telecommunications service industry. He considered abstractions with real-life implementation reinforcing
that the lowest layer of the model reflected a hardware major themes of the keynote. At one point, Lazar
market where numerous equipment manufacturers and presented a model of a schedulable region (which
vendors offer hardware and firmware solutions for represents the resource capacity of a switch multiplexer)
building the basic communication infrastructure. The as a live feed from one of the ATM switches in
customers of this market are typically network carriers, Columbia's broadband network. The schedulable region
third party software developers who specialize in was configured and managed using a QOS extended
developing software for service providers and a handful version of Ipislon's General Switch Management
of service providers themselves. The Application Protocol (GSMP) called qGSMP developed by the

ACM SIGCOMM ' , '~'!, 101 Computer Communication Review

COMET Group at Columbia. To bring the audience The next presentation, by Stephen Wade [Blair,97],
closer to the problem of engineering a solution, the Lancaster University, addressed the design of a
speaker incorporated Java applets to illustrate live feeds distributed systems platform and algorithms for mobile
from switches in the COMET laboratory. The audience computing environments. The platform, called Limbo,
clearly saw a representation of a schedulable region and aims to support the development of demanding mobile-
its operational points changing dynamically as the aware distributed applications in a heterogeneous
traffic through the switches in the testbed varied over networking environment. Limbo is based on the concept
time. Such illustrations helped to showcase several of tuple spaces, which has been extended for the mobile
important points of this spirited and informative computing environment to support QOS management.
keynote address. Limbo places emphasis on QOS monitoring and
adaptation. The speaker argued that the tuple space
paradigm was particularly useful in modeling adaptation
4. Technical Sessions to changes in network connectivity in mobile
IWQOS'97 was a truly interactive event. The general networking environments.
format of each of the ten technical sessions included The third talk of the session, by Badrinath
long and short papers followed by a thirty minute panel [Badrinath,97], Rutgers University, focused on
discussion on topics raised during the session and in architectural support for Internet cellular telephony. It
response to questions from audience. was evident to Badrinath that those who designed RSVP
The ten technical sessions comprised: mobile and the integrated service architecture had not looked at
communication, QOS routing, advanced reservation, mobility issues. In contrast, those that developed
traffic management, QOS and video systems, mobile-IP had not addressed QOS issues. The speaker
distributed object computing, QOS management, QOS- made an argument to unify some of these disparate
based transport protocols, QOS mapping and QOS pieces in support of cellular phone services. The
adaptation. proposed service was based on an IP network capable
of delivering packetized voice to moving users.
Badrinath calls the solution "RSVP+mobilelP+QOS".
4.1 Mobile Communications
Next, Andrew Campbell [Campbell,97], Columbia
Chair: Mahmoud Naghshineh, IBM University, discussed a number of QOS challenges for
The opening session of the workshop investigated the next generation mobile middleware. The speaker
feasibility of delivering continuous media with QOS reflected that the area of QOS and mobility was in its
guarantees in mobile networks. A key observation of the infancy. The wireless media systems project at
session was that providing QOS support in a wireless Columbia was attempting to shine some light on the
and mobile environment requires a fundamentally subject by building a prototype QOS-aware middleware
different approach from that found in wireline networks. platform for mobile multimedia networking called
The session included two long and four short papers. mobiware. The platform is programmable and runs on
mobile devices, base stations and mobile-capable
Presentations
switches. Mobiware includes a new active and adaptive
The first talk was given by Bharghavan [Lu,97], transport, QOS controlled handoff algorithms and an
University of Illinois at Urbana-Champaign. adaptive network QOS model.
Bharghavan proposed a number of solutions to the
The next speaker, Javier Gomez-Castellanos
problem of providing sustained QOS to mobile
[Gomez,97], Columbia University, presented results
applications. Limited and varying resources availability,
showing the effect of transmission errors on MPEG
stringent application requirements and user mobility
streams over wireless links. The speaker proposed
make providing QOS guarantees in mobile and wireless
several algorithms which improve the perceptible
environments challenging. He introduced an adaptive
quality of MPEG stream during periods of fast and slow
service model that enables the network and mobile
fading. Gomez-Castellanos suggested an algorithm
applications to renegotiate QOS depending on dynamic
based on a combination of Forward Error Correction
network conditions. Following this he described an
(FEC) and Automatic Repeat Request (ARQ) as a way
algorithmic framework that provides cost-effective
to minimize the impact of error characteristics on video
resource adaptation in the presence of resource
in this instance. He introduced a packet tagging
dynamics. Bharghavan concluded by arguing for a
technique that takes into account the particular
unified architecture for QOS adaptation.
semantics of MPEG flows and the relative importance

ACM SIGCOMM 102 Computer Communication Review

of different packets (e.g., GOP, headers, IPB, scalable Toward the end of this lively discussion, there was a
profiles) as they traverse a mobile network. question from the audience concerning the relationship
between pricing models and adaptive mobile
In the final presentation of the session, Steven Pope
multimedia environments. Badrinath explained that
[Pope,97], Olivetti and Oracle Research Laboratory,
clearly there would be a premium cost for mobile over
discussed QOS support for mobile computing
mobile incapable users - as exists today in cellular
environments. There is a demand for completely
telephony systems. The issue is what would the policy
portable computers (which Pope called walkstations) to
be when a user service is forced to degrade to a lower
access the network while traversing both indoor
quality; how do we cost that likelihood? Badrinath
wireless LAN networks and the outdoor mobile radio
speculated that while users currently pay a premium for
network infrastructure. The speaker introduced a traded
mobility they would naturally expect a discount from
handoff where connections are rebuilt during a handoff
the provider as a consequence of network initiated QOS
to the most appropriate service, taking into account the
degradation.
properties required by the application and locally
available, replicated or compatible services.
Panel Discussion
4.2 Traffic Management
Chair: Ed Knightly, Rice University
At the end of this session, the question of how to
provide QOS in wireless and mobile environments was The four long papers presented in the traffic
raised. The session chair presented two opposing views management session approached the QOS traffic
for consideration by the panel where the mobile management problem from various angles and proposed
network provided: widely differing solutions. The first paper considered
the worst case traffic pattern for source policing as an
• hard guarantees, in which case there was no need integral part of the traffic model. The second and third
for adaptation mechanisms in applications - as in papers focused on Markovian modeling techniques,
the ease of future wireless ATM networks; and while the fourth advocated a measurement-based
approach.
* no guarantees, in which case there was a strong
need for highly adaptive mechanisms in the Presentations
applications - as in the case of today's best effort
Philippe Oechslin [Oechslin,97], University College
mobile IP network.
London, presented the first talk of the traffic
Many of the panelists thought that QOS offered by management session on the topic of myths of on-off
future mobile multimedia networks would lie between sources. This was in relation to the worst case arrivals
the two extremes. The panel considered supporting hard of leaky bucket constrained sources. Simulation results
guarantees for a wide range of multimedia traffic from a set of independent connections limited by leaky
unrealistic given the nature of the wireless medium and bucket shapers and fed into a buffered multiplexer were
mobility requirements. Bhraghavan stated that there presented. Oechslin indicated that this scenario was
seems to be universal acceptance that an intermediate typical of an ATM switch or in a looser sense typical of
service be based around an adaptive resource an RSVP capable router. Results from the analysis
management model. This approach benefits by reducing found periodic traffic patterns resulted in poorer loss
handoff dropping probability and increasing the rates over the on-off or tri-state patterns models. The
utilization of the network. results invalidate the widespread belief that on-off
patterns are the worst case traffic of independent leaky
Mahmoud Nagshsineh followed up with another bucket constrained sources.
question to the panel: if there was agreement on an
adaptive resource management then where should the The second paper, given by Hoon Lee [Lee,97], Korea
adaptive algorithms reside: at the physical, MAC, Telecom, presented a cell access control scheme for
network, transport or application layers? If applied guaranteeing multiple classes of cell loss QOS
solely to application layer, this would result in low cost requirements in an output buffer of an ATM switch. Lee
and minimal complexity in the network. On the other proposed a class acceptance controller, which regulates
hand, if the network explicitly supports adaptation then the acceptance of the cells of QOS classes, based on the
this would result in high cost and increased network dynamic state of the queues. He considered decision
complexity. Weighing these arguments the panel agreed functions for the class acceptance controller with a view
that there was need for research into adaptation support to comparing their "effects to the QOS performance.
in the network and the end-system. Queueing analysis of the scheme derived a number of
performance measures. The implications of the work

ACM SIGCOMM 103 Computer Communication Review

were further illustrated using a number of numerical supporting new types of traffic where measurement data
experiments. is not available, or in planning backbone networks
where the requirement on the modeling accuracy is
The third paper, by Dietmar Becker [Becker,97],
relaxed due to the multiplexing of a large number of
Aachen University of Technology, reported on
flows.
queueing analysis for a partial buffer system with
discrete Markovian arrival processes. An evaluation of Lee indicated that in the backbone networks, resources
the performance of the partial buffer system with finite are conservatively dimensioned so that the control
capacity, deterministic service time and multiple algorithms are not sensitive to the inaccuracy caused by
sources was presented. A discrete Markovian arrival measurement algorithms. This led to a debate initiated
process modeled each source. The queueing system was by Maryni on whether traffic modeling is needed at all.
evaluated for several traffic compositions and different In future, he rhetorically argued, why use traffic
sizes of the shared buffer area. Becker considered a modeling techniques when network capacity may be
number of traffic compositions including VBR sources infinite? Since traffic modeling aims at increasing
with periodical or negative exponential correlation bandwidth utilization, with the abundance of bandwidth
functions and CBR traffic with fixed interarrival cell this may disappear like the technique for reducing the
emission. The probability distribution of the cell loss of number of transistors in circuit design. Oechslin added
each source was presented. that abundant bandwidth availability would be a reality
now if the right pricing model were in place to generate
The final talk of the session, on real-time estimation of
enough revenue to turn on more bandwidth.
the link capacity in multimedia networks, was presented
by Piergiulio Maryni [Maryni,97], DIST-University of
Genoa. Maryni suggested that simple but powerful 4.3 QOS Routing
abstractions that represented the capacity of multimedia
Chair: Mischa Schwartz, Columbia University
networks are needed. In order to guarantee QOS, the
link capacity must first be calculated, i.e. the total The basic goal of QOS routing is to select a path
number of calls of different types that can be admitted through the network that satisfies a set of QOS
on a single link at a given time. The speaker used the constraints while achieving some level of network
notion of a schedulable region to represent the link resource utilization. The first two long papers in this
capacity for an ATM multiplexer. Maryni presented a session presented differing approaches to achieving this
new approach for computing the schedulable region in goal. Following this, Scott Corson, Maryland
real-time which could be used as input to admission University, joined the panel discussion and raised a
controllers. The methodology relies on real-time QOS number of open issues in providing QOS support when
measurements to dynamically compute the size and the routing flows through mobile ad-hoc networks.
shape of the region. No assumptions about traffic Presentations
resource models or scheduler operations were needed
for its construction. The first talk given by Qingming Ma [Ma,97], Carnegie
Mellon University, addressed the issue of QOS routing
Panel Discussion for traffic with performance guarantees. The speaker
The panel discussion began by addressing the role of presented some initial results on QOS path selection for
traffic models in delivering QOS guarantees. The panel traffic flows requiring bandwidth, delay and jitter
recognized the difficulty in traffic modeling and the guarantees. For traffic that required bandwidth
complexity of meeting traffic characteristics and guarantees, it was found that several routing algorithms
adapting to a wide variety of user applications needs. that favored paths with fewer hops perform well. Ma
Most of them agreed that the measurement based argued that a modified version of the Bellman-Ford
approach could significantly relieve some of the shortest-path algorithm in polynomial time was
difficulty inherent in many traffic models proposed in sufficient for traffic with delay guarantees. He showed
the literature. that the problem of finding a path that can satisfy
bandwidth, delay, jitter, and/or buffer could be solved
Ed Knightly commented on the weakness of while at the same time deriving the bandwidth that has
measurement-based resource allocation, whose to be reserved to meet these constraints.
inaccuracy lies in the measurement procedure that
averages out the heterogeneous behavior of traffic S. Verma [Verma,97], University of Toronto, reported
flows. The panelists believed that Markovian modeling on QOS based routing in support of emerging multicast
and worst-case analysis can still play roles that multimedia communications. Verma proposed a routing
complement measurement-based approaches; that is, in metric that could be used in combination with heuristic

ACM SIGCOMM 104 Computer Communication Review

algorithms to find the multicast tree for guaranteed QOS session included two long and two short papers that
services. The optimum tree used in the formulation was reported on adaptive QOS driven video systems.
a delay-bounded minimum Steiner tree. Simulation
results showed a marked improvement in network Presentations
utilization expressed in terms of cost over other Sanjay Jha [Jha,97], University of Technology, Sydney,
proposed schemes such as QOS path first routing. proposed a set of playout management algorithms for
interactive video. The work examined problems
Panel Discussion
associated with display of live continuous media. Under
The panelists were asked by the audience to clarify the the assumption that the network cannot guarantee the
relationship between routing and resource reservation required bounds on delay and jitter and the operating
for end-to-end QOS guarantees. Ma stated that the first system scheduling is non-real time, there is a need to
goal is to discover a path and then reserve the resources accommodate the delay and jitter in the end systems in
along the path or reject the flow/call if resources are not order to maintain a desirable QOS. Jha proposed a
available. Other approaches handle routing and resource method of video playback that requires accurate
reservation at the same time - connection oriented estimation of display cycle time of video frames and the
systems couple routing and resource reservation. delay suffered by frames in packet networks.
Deterministic forecasting methods used in time series
The issue of state management was debated. An
analysis were applied to experimental data collected
objective of QOS routing is to distribute state
from video transmission.
information that accurately reflects available resources
on a particular path. Probabilistically resources may The second paper by Rajeev Koodli [Koodli,97]
not be available at a certain switch/router if the state is proposed the notion of noticeable loss, which directly
old. The panel discussed the implication of QOS state relates loss pattern to the perceived QOS for an
management and concluded that it needs to realistically application. Noticeable loss was used to evaluate
represent available resources. In the past, call setup resource management algorithms that provided QOS to
used crank-back techniques to resolve inconsistencies in individual adaptive applications. Simulation results
the QOS routing state representation. illtistrated the performance of the algorithm.
Scott Corson [Corson.97] described the generic D. Bourges Waldegg [Waldegg,97], ENST, presented a
character of routing in mobile ad-hoc networks. He temporal QOS based CPU scheduling model for
highlighted some of the differences between ad-hoc and multimedia. The model is based on a playout
wireline multihop networks and the impact these specification and a runtime application structure that
differences had on QOS-based delivery systems. Corson allows workahead processing and quality degradation
suggested that supporting QOS in ad-hoc networks was for delay during overload conditions. Policy is used by
an extreme challenge. Wireless QOS architectures must a scheduler to degrade the service in a meaningful way
provide a balance between network and application- with the goal of supporting better resource utilization.
level QOS adaptation. This, he argued, would help Waldegg added that the run time structure supported
minimize QOS-related signaling which has traditionally workahead processing which had shown benefit
been integrated with routing state to support a given particularly in the case of overload.
QOS. Corson concluded by saying that two approaches
The last talk by Steven Jacobs [Jacobs,97], Columbia
seemed to warrant further investigation for ad-hoc
University, proposed that networked video systems
services; namely, minimal guaranteed QOS and
could operate sufficiently well over best effort
probabilistic QOS.
networks. Architecting support for video service in
packet networks required the addition of a number of
4.4 QOS and Video Systems valued added algorithms. Jacobs stated that there was
great demand for such services today. Furthermore, he
Chair: Alexandros Eleftheriadis, Columbia University
speculated that the demand would persist in the future
There are two schools of thought related to QOS and even when many networks supported multi-level QOS
video systems. One school argues that resource assurances. An Internet-based video delivery system
reservation is essential for video services. The other was presented which combines both image processing
argues that the network need only be engineered to and networking techniques. Jacobs presented the results
support best effort services. In this case networked from several experiments that utilize a combination of
video systems need only estimate resource availability bandwidth estimation and dynamic rate shaping of
and intelligently adapt to the observed state. This video sources.

ACM SIGCOMM 105 Computer Communication Review

Panel Discussion challenge of proving predictable QOS for multimedia
applications. In such an environment applications may
Alexandros Eleftheriadis led a lively discussion which not know their exact resource requirements in advance
addressed a number of issues raised in the session and and resource requirements and resource availability may
then opened the floor for questions. The first question be time-varying. To address these challenges,
concerned media scaling techniques. Eleftheriadis Lakshman proposed a resource management
commented that the choice of either frame dropping or architecture in which applications and the operating
quality reduction within a frame'(i.e, dropping system cooperate to dynamically adapt to variations in
transform domain coefficients) is strongly content the resource requirements and availability.
dependent, which can vary along the wide spectrum of
applications. Some applications, such as movies, have The topic of QOS management of integrated services
scenes that tolerate little content loss and require real- communication systems was addressed next by Roland
time delivery, thus they would have difficulty with best- Bless [Bless,97], University of Karlsruhe. It is likely
effort delivery. that such networks will support a wide variety of
applications that will be multiplexed into different
The panelists discussed whether hardware solutions for service classes. Bless reported on a flexible QOS
video can alleviate many of the problems experienced management scheme for such communications
using software implementations. Jacobs argued that environment. The approach provided flexibility in
hardware support for video is improving and one should tailoring the QOS management to suit specific service
address other problems instead of duplicating hardware profiles. Bless outlined the implementation of the QOS
solutions in software. Eleftheriadis pointed out that the management system and presented performance
types of hardware assists are converging as content measurements to illustrate the approach taken.
providers only use limited set of encoding standards.
However, Jha argued that even for the audio codec, Next, Stefan Fischer [Fisher,97], University of
many interoperability problems remain. In today's Montreal, presented a decentralized scheme for
Internet, lack of QOS support means that researchers cooperative QOS management. The system supports
have little choice but to use flexible software solutions. QOS management functions such as QOS negotiation
and adaptation over the network. The speaker suggested
The last question from the audience addressed the pros that the work was especially suitable for multicast
and cons of dynamic rate shaping and quantization. multimedia communications. Fisher introduced the
Jacobs pointed out that dynamic rate shaping is better notion of QOS agents, installed throughout the system
than the quantization technique when the required rate where QOS negotiations occurred. QOS agents
reduction is small. Dynamic rate shaping simply drops communicate with their neighboring agents. This
some coefficients, requires less state information to be communication is application oriented, i.e. the agents
propagated, and offers better quality under this know about QOS requirements and negotiated values on
scenario. In other words, quantization pays the penalty behalf of the users. Fischer concluded by discussing the
for SNR and speed as long as the required reduction construction of applications based on the notion of
ratio is not too large. When a large reduction is needed, cooperative QOS management.
however, quantization techniques are Superior.
Jan de Meer [de Meer,97], GMD Fokus, reported on the
specification of end-to-end QOS control. The speaker
4.5 QOS Management distinguished between the QOS control demands for
Chair: Andreas Vogel, Visigenic Corp. continuous and discrete media. Classic control theory
The next session, chaired by Andreas Vogel from seemed an appropriate vehicle to monitor and respond
Visigenic, included two long and three short papers. to time varying quality as flows progress through the
Each of these papers focussed on the issue of QOS communications system. De Meer introduced a "water-
management techniques. The presentations covered level monitoring" paradigm which consists of four
both architecture and implementation research, with components, namely, a source, container, monitor and
topics ranging from QOS support in operating systems sink, to assist in the QOS management process.
and file systems, end-to-end specification, adaptive The final talk of the session, presented by M.
QOS architecture and QOS management agents. Spasojevic [Borowsky,97], HP Labs, on using attribute-
managed storage to achieve QOS. The speaker argued
Presentations
that specification of storage systems by means of user-
The first paper, presented by Lakshman [Lakshman,97], oriented QOS attributes is the key to ease of use and
Intel, reported on an integrated approach to CPU and efficient resource utilization. Currently, most existing
network IO QOS management. Lakshman illustrated the storage management is too low level requiring the user

ACM SIGCOMM 106 Computer Communication Review

to allocate and configure arrays of disks. However, little formulation and evaluation of customized CORBA
help is provided in helping the user do this. Attribute- trader. A service distance is computed between the
managed storage systems hide details of the underlying client and the service offers taking into account QOS
storage systems through virtual storage abstractions, properties. The modified trader selects the service with
units of storage with QOS guarantees. The speaker the minimal distance to the service request. The Orbix
described a prototype matching engine called Forum trader was used to evaluate this QOS metric. Claudia
currently under development. Linnhoff-Popien concluded her talk by describing some
implementation results in comparison to an unmodified
Panel Discussion Orbix trader.
The first question from the audience addressed existing The next paper, presented by W. Almesberger
operating systems: are they sufficient to support [Almesberger,97], EPFL, surveyed QOS in
integrated QOS? Many panelists thought that work communication APIs. The speaker first contrasted the
needs to be done before a QOS-driven operating system RSVP and ATM UNI APIs and distinguished how QOS
would be available on the market. Lakshman, from his was exposed to the applications using WinSock 2,
experience with using Solaris, commented that the X/Open and Arequpa. A natural consideration is the
hardware already has a high-resolution clock, whereas mapping of these APIs to local operating system and
he considered his enhancements to the operating system network resources. Almesberger went on to summarize
targeted faster preemption and better estimation of how these APIs enable applications to control QOS for
computing time operations. their connections/flows. Each API has its own
A member of the audience asked whether the whole idiosyncrasies (e.g., native ATM APIs deal in cells and
area of QOS management research is well understood. RSVP APIs in bytes) and supports a different set of
The panel observed that there is still a lack of QOS parameters and traffic characterizations. The
experience with QOS management systems, with QOS speaker proposed the unification of QOS description
mapping across layers, and with QOS specifications. using better abstractions to resolve idiosyncrasies found
They believe that the research would shift to toward in existing APIs.
building large-scale prototypes. This would depend on In the third t.alk, John Zinky [Zinky,97], BBN, reported
the availability of a large-scale network testbed. on managing systemic meta-data for creating QOS-
Finally, Fisher commented that the IP networking view adaptive CORBA applications. Zinky made the point
of QOS management architecture is that it was far too that distributed applications must be able to adapt to
complex. It may be best to revisit fundamental control quite diverse operating conditions. The speaker
theory, focus on different time-scales and develop introduced the notion of systematic meta-data, which
simple solutions. captures how applications utilize distributed systems
technology such as CORBA, adapt their QOS
requirements, use of resources, and allocation policies.
4.6 Distributed Object Computing In this position statement, Zinky discussed some
Chair: Douglas Schmidt, Washington University at St. possible solutions. Meta-data required support from the
Louis network, distributed system and applications. The
Distributed object computing technology has been network needs to support explicit mechanism for
widely applied to resolve problems stemming from the moving and storing meta-data. Application
complexities of developing large heterogeneous programmers need APIs for QOS at the client/object
software systems. This session focuses on some of the boundary rather than at the socket level.
QOS issues that surround the use of distributed object- The fourth paper in this session was presented by D.
computing technologies. Issues addressed ranged from Reed [Reed,97], Stirling University, on supporting QOS
QOS parameterized trading services to QOS meta-data components in distributed environments. A key
management for middleware. This session included two component required in such an environment is QOS
long and three short papers. monitoring services. In his talk Reed proposed a generic
distributed monitoring service which adapts to suit
Presentations
particular applications as a means of overcoming the
The first presentation of the session, by Claudia complexity of specialized monitoring solutions. A
Linnhoff-Popien [Linnhoff-Popien,97], Aachen number of examples highlighted the flexibility of the
University of Technology, addressed the integration of monitoring service.
QOS constraints into the service selection process. The
In the final talk of the session, Andrew Grace
major challenges addressed by the work were the
[Smith,97], British Telecom, reported on a QOS

ACM SIGCOMM 107 Computer Communication Review

configuration tool for distributed applications. The video can playout at the desired rate. If the user wishes
range of distributed applications has increased to only accept the desired rate, the system makes use of
dramatically over the past several years fuelled by the advanced reservation to book resources ahead of time
growth of the Internet. Many applications tend to for the duration of the video-on-demand.
require user level knowledge of low-level technical
parameters requiring an appreciation of system
PanelDiscussion
heterogeneity issues rather than simply stating what Steve Pink and Lars Wolf joined the presenters and
service they require. In this presentation Grace Hideyuki Tokuda for the panel discussion on advanced
described a working system for QOS configuring for reservation. The session chair raised a number of
Mbone conferencing applications. QOS profiles were interesting questions and directed them at the panel.
adopted as a means of specifying resources and Tokuda asked if we really need advanced reservations.
requirements in the end-system and networks. If so, what type of quality demanding services shall we
use advanced reservation for? How can we maintain
reservation state and how would failure be handled?
4.7 A d v a n c e d Reservation
Steve Pink was unsure of the demand for advanced
Chair: Hideyuki Tokuda, Keio University
reservation services. State management of reservations
Is there a need to make reservations in advance? In this looks troublesome since the models introduce a lot of
session, a number of speakers argued that there are sets state and require switches/routers with plenty of
of applications that require a high degree of resources memory/storage to maintain it. Pink suggested that
availability in advance. The first paper takes an existing reservation architectures might be too complex
empirical look at advance reservation from the network for the network at the moment. Therefore, he proposed
viewpoint. The second and final paper in this session to separate control functions from packet forwarding.
reports on a scalable video-on-demand system which Pink also argued that RSVP is not the most appropriate
utilizes advanced reservation techniques. choice for advanced reservations signaling. For RSVP
Presentations to function properly, the senders should be present in
advance. This cannot always be guaranteed for
Olov Schelen [Schelen,97], Lulea University, began advanced reservation applications. Therefore, advanced
this session with a presentation on sharing resources reservations are a function of the management layer and
through advanced reservation agents. The speaker not the packet-forwarding network.
proposed an architecture where clients make advance
reservations through agents responsible for advance Lars Wolf summarized what he considered to be the
admission control. The agents allocate resources in the open issues on the topic. These included the duration of
routers just before they are needed for packet reservation, stacks, failures, distribution of
forwarding. Schelen illustrated that network resources announcement information, management of resource
can be shared between immediate and advance and required protocol changes. The most challenging
reservations applications without pre-partitioning. problems are state maintenance and failure handling.
Admission control decisions for immediate reservations All systems performing advance reservation must keep
use information about resources to be allocated for the associated state information for potentially long
advance reservations in the near future. The speaker periods of time. This must be stored in non-volatile
introduced a new parameter in the admission control memory to survive system shutdown. Wolf described
algorithm for the lookahead. this as the hard-state approach. Alternatively, similar to
the approach followed by RSVP, the reservation may be
Next, Abdelhakim Hafid [Hafid,97], Computer refreshed from time-to-time - he called this the soft-
Research Institute of Montreal, proposed a scalable state approach. W01f believes advanced reservations
video-on-demand system that uses advanced reservation could be useful for several application classes.
techniques to support services. Typically, video-on- However, advance reservation raises difficult questions
demand systems check whether there are enough that need to be resolved.
available resources to deliver the requested movie to the
user's host. Given sufficient resources, the movie
presentation will commence, otherwise, a rejection is 4.8 QOS-based Transport Protocols
sent back to the user. Hafid described an advanced Chair: Steven Pink, Swedish Institute of Computer
reservation signaling system called NAFUR. If a QOS Science
request for a video stream cannot be immediately
Historically, transport protocols have been a hotly
supported at the desired rate NAFUR determines at
researched topic in computer networking. With the
which point in advance of the current request time the

ACM SIGCOMM 108 Computer Communication Review

advent of multimedia, there has been a move away from In the final talk of this session, Glenford Mapp
designing reliable, high performance data transports to [Mapp,97] Olivetti & Oracle Research Lab, reported on
transports that Support end-to-end QOS guarantees. development of a QOS-based transport protocol called
This session consisted of two long and two short paper A1. The transport was designed to provide QOS trade
presentations on issues surrounding the development of offs rather than strong guarantees. Mapp discussed the
QOS-based transport systems. trade offs between qualitative QOS such as order and
reliability and quantitative QOS such as delay and
Presentations throughput. The transport service supported the notion
The first paper presented, by K. Fukuda [Fukuda,97], of a QOS vector to specify all transport requirements at
investigated QOS mapping issues between a user and a the API. Preliminary performance results for A1
video transport system. Fukuda described a QOS running over ATM were presented and compared with
mapping method between user preference for video an efficient kernel implementation of TCP/IP.
quality a n d t h e required bandwidth to transport the
Panel Discussion
resulting video flows. This work assumed that the
underlying network is capable of supporting a In the panel discussion that followed, the Chair, Steve
bandwidth allocation mechanism such as deterministic Pink initiated a discussion by making the observation
bit rate service class in ATM, RSVP, IPv6, etc. Based that two opposing trends seem to be emerging within
on spatial, SNR and time resolutions QOS parameters, transport design. On one hand, as applications become
the QOS mapping function derives the required increasingly sophisticated in their requirements, newer
bandwidth to support MPEG-2. The mapping between transport protocols should be developed to support t h e
QOS parameters and user perceived video quality is required functionality. On the other hand, it is argued
then calculated using classic mean opinion score that since only applications can truly understand their
evaluation testing. data semantics, traditional transport functions should be
removed from the transport to the applications layer.
Jean-Franqois Huard [Huard,97], Columbia University,
Pink then solicited questions from the audience.
reported on end-to-end QOS mapping. A simple
mathematical formulation for mapping QOS parameters A member of the audience asked that if trends were
between application and transport was derived. A toward thinner and thinner transports then why have a
.platform was developed for evaluating end-to-end QOS transport layer in the first place? The panel's response to
by performing concurrent network, transport and this question was mixed. Glenford Mapp agreed with
application level measurements. The loss bound this comment and was of the opinion that the transport
empirically obtained under the assumption of uniformly in its traditional form was on its way out! Jean-Francois
distributed cell losses within a video frame is too Huard, however, took a different line. In his opinion,
conservative. Early results suggest that the existing the classical algorithms for providing reliable, flow-
literature on loss mapping is typically too conservative control and sequenced delivery are too complex to be
by a factor of three. Huard concluded that in order to left to the average application programmer. Pink agreed
obtain better empirical QOS mapping rules between the with this and noted that the recent move toward
application, transport and network, more data needs to multicast communications may redefine the role of
be collected and analyzed. current transport protocols. Here, the difficulty lies in
providing scalable reliable multicast flows. Many of the
Next, P. Conrad [Conrad,97], University of Delaware
well-understood techniques including positive
made the following position statement: "Transport QOS
acknowledgements do not work in this environment.
over Unreliable Networks: No Guarantees, No Free
Lunch!". The talk presented an approach to transport in Pink followed by leading a discussion on user versus
unreliable networks, investigating trade offs between kernel level transports. He noted that there are primarily
qualitative QOS parameters (e.g., order and reliability), two types of processing activity in any transport
and quantitative parameters (e.g., delay and protocol: the more expensive per-byte processing for
throughput). Conrad focused on partially ordered and computing packet checksums and the less
partially reliable transport services. The key results are expensive per-packet processing for flow-control,
that both sender-based and received-based reliability acknowledgment, etc. The classic argument for kernel-
schemes for providing partial reliability achieve almost level transport protocols has been to perform efficient
identical reliability and delay. On the other hand, a per-byte processing. However, as network speeds
sender-based approach provides better throughput than increase to gigabits, it seems sensible to delegate much
a receiver-based approach at higher loss rates. of this computer intensive functionality to specialized
hardware. This leaves the task of per-packet level

ACM SIGCOMM 109 Computer Communication Review

processing to software. The consensus among the quality representation and the underlying operating
panelists was that user-level transport implementations system resource, namely the CPU. A communications
could be as efficient as kernel level implementations. model was analyzed for different MPEG grouping
This, plus the added flexibility of being able to schemes and communications paradigms. Nahrstedt
decouple the transport from the operating system makes commented that a middleware level seemed like a
user-level transport protocols ideal for the increasingly natural point where the user can specify QOS
specialized needs of today's sophisticated applications. requirements using application language and the
operating system/network derives QOS parameters
expressed in its own language
4.9 QOS Mapping
In the final paper of the session, Valerie Issarny
Chair: Jean-Pierre Hubaux, EPFL
[Issarny,97], INRIA, discussed the translation of QOS
QOS mapping performs the function of automatic specifications in a QOS architecture. Issarny discussed
translation between representations of QOS at different the development of customized software architectures.
system levels (i.e., application, operating system, These include the specification of execution properties
transport and network, etc.) and thus relieves the user of such as interaction properties, that capture the
the necessity of thinking in terms of lower level communication patterns, and QOS properties, which
specifications. This session included two long and two represent resource management policies implemented
short papers which addressed the translation between by middleware. The main challenge of QOS translation
application QOS specification, and the operating system is to correctly specify the interaction and QOS
and network QOS. properties so that a proper customization of the
Presentations resulting QOS architecture may be provided.

In the first talk, N. Nishio [Nishio,97], Keio University, Panel Discussion

presented a simplified method for session coordination Jean-Pierre Hubaux highlighted the layers between the
using a three-level QOS specification and translation application and operating/network for which QOS
scheme. It may be unrealistic to expect the application mapping was needed. He then posed a general question
to specify its QOS requirements using operating to the panel: given the diversity of application and
system/network specific language, e.g., memory size in underlying networks and operating systems, is it
Kbytes or bandwidth in Mbps, etc. However, such feasible to try to formulate a generic framework for
information needs to be distilled from the application QOS mapping and, if so, how far are we from it? The
specification for admission control and resource response from the panel was unanimous in that they
reservation. To address this need, Nishio introduced a believe that QOS mapping is an area that is just
QOS architecture which presented the user-level beginning to be understood.
specification in .terms of application program QOS,
middleware QOS, and system-level QOS. Nishio However, the reasons given by each member of the
described a conductor/ performer paradigm used to panel varied widely. Klara Narhstedt reasoned that any
handle the translation function between these system form of generic framework would have to be qualified
components. given the diversity. The typical technique used today is
through some form of application profile
In the next presentation, H. Knoche [Knoche,97], characterization. She believes that the whole field of
University of Hamburg, reported on a quantitative QOS QOS mapping is still in its infancy given that the scope
mapping approach. Knoche identified QOS mapping as of mapping considered in most schemes is still fairly
the process of translating QOS parameter bounds from simplistic and static in nature: focusing on a certain
layer to layer and finally to specific resources. General application, operating system and network.
mapping between video frame service data units and
network quality requirements such as delay jitter, Another point raised by the floor was that most QOS
throughput and reliability were presented. QOS mapping schemes address only continuous media
mapping took into account the cases of common service services avoiding other communications services, e.g.,
functions such as segmentation/reassembly, blocking, transactions. Hendrick Knoche felt that the primary
playout buffer, interleaving, coding, or peak smoothing. difficulty in formulating good QOS mapping stems from
the fact that we do not yet understand the perceptual
Klara Nahrstedt [Kim,97], University of Illinois, effect of QOS. Nobuhiko Nishio noted that the
Urbana-Champaign, presented an integrated view of interaction problem among the many layers in a system
QOS translation and admission control. Nahrstedt complicates the issue of QOS mapping since the
discussed a translation between the MPEG-2 video

ACM SIGCOMM 110 Computer Communication Review

dynamics are usually non-linear and difficult to issue of fairness nor interaction between adaptive
characterize. traffic, e.g., TCP. Sisalem concluded by indicating that
adaptive schemes can suffer from fairness problems
Jim Van Lot, Sun Microsystems, asked the panel if
potentially causing starvation of reactive flows such as
there was a good abstraction or metaphor for expressing
TCP.
QOS mappings? This question cut to the heart of the
problem. Nobuhiko reflected that whatever the The third paper, by Max Ott [Ott,97], NEC, reported on
abstraction might be, it would likely be based on adaptive QOS in multimedia systems. Ott made the
economic concepts and be closely indicative of the cost point that most QOS architectures present QOS-aware
involved of rendering the service to be mapped. APIs to the applications. This is either achieved by
adding QOS parameters to standard system calls, or by
raising the system abstraction to a higher level, filling
4.10 Q O S A d a p t a t i o n the gap with what is often referred to as middleware.
Chair: Klara Nahrstedt, Unwersity of Illinois at The speaker reflected that in either case there seems to
Urbana-Champaign be a serious desire to draw a strict line between "us",
Many distributed multimedia applications are adaptive the QOS provider, and "them" the elusive application.
in nature and exhibit flexibility in dealing with Ott argued that it seemed natural to define an
fluctuations in network conditions. QOS adaptation architecture that allows the introduction of QOS at any
algorithms can, for example, trade temporal and spatial level: from the CPU and network resources to the user's
quality to available bandwidth or manipulate the "satisfaction". He reported on an architecture that has
playout time of continuous media in response to evolved over the past few years at NEC that has
variations in delay. This session comprised two long introduced the concept of QOS adaptation at each level
and three short papers on the topic. of the architecture. QOS is specified by contracts which
are established between clients and service providers
Presentations using a single generic programmable API.
In the first talk of the session, Pratyush Moghe Oh [Oh,97], Osaka University, presented the next paper
[Moghe,97] Bell Labs, addressed what he described as on a dynamic QOS adaptation mechanism for a
"terminal QOS" for adaptive applications. Next networked virtual reality system. The motivation behind
generation terminals are expected to support the work is to maintain acceptable user presentation
sophisticated adaptive applications. Since some quality when resources fluctuate in a networked virtual
terminals have limited power (e.g., personal digital reality system. Oh introduced the notion of "importance
assistants and network computers), the application of presence" which when applied to objects in a virtual
processing delay can be a significant component of end- reality system can be used to trade off available
to-end delay. Moghe called this the terminal QOS resources and objects. Importance of presence is based
measure. Currently, each adaptive application has its on the maximum visible distance and angle of incidence
native adaptation algorithm that operates independently between the users and an object in the user's virtual
of other applications, their adaptation algorithms, or space. The speaker discussed an adaptive algorithm that
scheduler. Moghe presented an analytical relationship reduced the QOS of an object based on its importance
between network feedback and the level of adaptation. of presence indication. At the end of his presentation
This theoretical framework is useful in understanding the speaker demonstrated the technique using a
the relationship and interaction between the adaptive videotape of the networked virtual reality system using
application, end-system and network. The speaker adaptation based on the importance of presence.
concluded by saying that the notion of terminal QOS
may be used to tune adaptation algorithms. In the final talk of this session Dan Revel [Revel,97],
Oregon Graduate Institute of Science and Technology,
Next, Dorgham Sisalem [Sisalem,97], GMD Focus, discussed predictable file access latency for multimedia.
presented an approach for dynamically adjusting the The speaker asserted that multimedia applications are
sending rate of applications to the congestion level sensitive to IO latency and jitter when accessing data
observed in the network. The speaker discussed how from secondary storage. To address this challenge,
senders could increase their sending rate during lightly Revel introduced the concept of transparent adaptive
loaded situations and reduce it during overload periods. prefetching which uses software feedback to provide
Sisalem presented results which illustrated the multimedia applications with file system QOS
efficiency of a direct adjustment algorithm responding guarantees: predictable low-latency access to data on
to fluctuations in available bandwidth while maintaining secondary storage. The research at Oregon Graduate
low loss rates. Currently, the work does not address the Institute is strongly focused on adaptive software

ACM SIGCOMM 111 Computer Communication Review

feedback algorithms for Internet video, mobile systems Currently, it is a fiction since no such product or public
and adaptive QOS access to storage. Currently, the domain platform exists. It will become a fact when
QOS interface to transparent adaptive prefetching researchers and developer are using QOS-based
allows applications to express adaptive needs in a middleware as they do C++ or JAVA. Clearly, we need
vocabulary that is meaningful to them. to ask ourselves what does such a rniddleware bring?
Yet another layer?
Panel Discussion
The panel argued that distributed object computing was
Klara Nahrstedt commented that QOS adaptation essential for architecturing complex software systems,
techniques, while complex in nature help to efficiently
ease of programmability and interworking across
share resources in the end-systems and network for a heterogeneous systems. A number of research groups
class of applications that can accommodate varying are building real-time ORBs and exposing QOS at the
resource availability. Nahrstedt followed up by asking
API for user programmability. Freeware is becoming
the panel whether they now considered QOS adaptation
available (e.g., Glenford Mapp, Olivetti & Oracle
as a replacement for reservation? A wide variety of
Research Lab, announced that ORL had put their ORB
opinions were articulated. Dorgham Sisalem used TCP
2.0 on the web). Andreas Vogel stressed that real-time
as an example to argue that TCP's congestion CORBA products will become a reality when there is
adaptation mechanisms did very well without any
demand from the customer. When the demand
explicit call setup and resource reservation; therefore transpires the industry would react positively.
why is reservation needed?
A member of the audience asked why CORBA has
Max Ott argued that if strong QOS guarantees were
become the common language, ORB of choice and
required (e.g., timing guarantees) then reservation was
subject of this panel? Guru Parulkar pointed out that
unavoidable. Nahrstedt said that QOS adaptation filled
CORBA is an existing standard and has been widely
the middle ground providing a poor person's QOS
implemented by multiple vendors. It has a number of
guarantee. Pratyush Moghe thought that many
deficiencies, he asserted, many ORB implementations
continuous media applications could operate at a
are not highly inefficient and support of QOS is
minimum level guarantee and use QOS adaptation
missing. However, Guru Parulkar felt optimistic that
techniques to achieve better quality when additional
these issues could be amended and were not "show
resources became available. Such a hybrid approach
stoppers".
grew out of the consensus that many applications can
not degrade below a certain level. Each of the panelists provided a list of open issues
which they believed must be addressed before QOS-
based middleware became a reality:
5. Workshop Panels
• Schmidt maintained that the ORB needed a real-
The workshop included two panels. The first panel time capability and middleware, such as
[Schmidt,97] looked at the emerging field of QOS in CORBA, is being modified to incorporate such a
distributed object computing environments. The second feature;
panel [Schulzrinne,97] highlighted some of the
concerns researchers have about a reservation driven • Ott argued that any emerging platform needs to
Internet. take seriously the user perspective to middleware
to ease programmability and simplify QOS
abstractions;
5.1 Q O S for Distributed Object C o m p u t i n g
• Parulkar stressed that we need to fix CORBA to
M i d d l e w a r e - Fact or Fiction?
support QOS in the end-system and at the
Chair and Organizer: Douglas C. Schmidt, Washington network access points in order to provide service
University guarantees at the I/O level, appropriate packet
Panelists: Max Ott, NEC, Guru Parulkar, Washington and process scheduling was needed;
U., Rolf Stadler, Columbia U., Andreas Vogel,
• Stadler remained upbeat about the use of the
Visigenic
technology in the network arguing for distributed
The panel discussion started with a statement by the interactive resource controllers with global
Chair that nobody uses a distributed middleware with control to deliver services with a minimal set of
QOS support, therefore, do we need such a layer? If guarantees and, in case of higher quality, they
yes, what does QOS-based middleware look like? Is should support adaptive behaviors;
QOS-based distributed middleware a fact or a fiction?

ACM SIGCOMM 112 Computer Communication Review

 • Vogel reflected that ORB vendors (e.g., Next, Roch Guerin, IBM, asked what the driving force
Visibroker) will eventually incorporate new behind reservations was. He mentioned that there is
CORBA services and QOS bindings for currently no application that is so critical that it can
multimedia streams and control when the market only function by making reservations. Alternatively,
demands it. He predicted that it will happen there are so many of these applications that it is
soon. impossible to define a generic reservation framework to
satisfy them all. Guerin argued that the main drivers for
Toward the end of the session questions of whether
reservations are the economic and contractual factors.
ORBs should explicitly support multimedia streaming
As the Internet is moving towards a commercial
were raised. There was clear disagreement about this
network, people would like to know what they are
issue. The question is whether the ORB is used to set up paying for - and ISP's also need a pricing model. This
streams and then moves out of the way or, conversely, it
requires enforceable and observable service contracts.
can explicitly support isochronous communications via
These contracts should be simple and deterministic to
RPC. Some panelists stated that this should not be part
avoid adding additional complexity to the infrastructure.
of CORBA. Others argued that for continuity the ORB
Guerin also pointed out that there are many conflicting
should support streaming. Andreas Vogel mentioned
forces to economic factors such as the cost of resources
that OMG would not support streaming through the
and their exploitation. In particular, cheap resources
ORB.
lead to simple signaling, where value is added at end-
systems. In contrast, if resources remain expensive, a
5.1 R e s e r v a t i o n s a b o u t R e s e r v a t i o n s network will need complex signaling, thus leading to
providers and equipment vendors adding most of the
Chair and Organizer: Henning Schulzrinne, Columbia value.
University
Panelists: Fred Baker, CISCO, Andrew T. Campbell, Dilip Kandlur, IBM, expressed some reservations about
Columbia U., Jon Crowcroft, UCL, Roch Guerin, IBM RSVP scalability. He cautioned that RSVP is able to
and Dilip Kandlur, IBM scale to a large number of receivers in a session, but not
a large number of unicast sessions. He provided
The panel discussed the current state and future
possible solutions to achieving the latter, which
developments of QOS support in an Internet and the
included aggregation of sessions, flow state
impact of a reservation driven network. Each member
management and path management.
of the panel started off by presenting a short position
statement of their vision of the rollout of QOS in the Henning Schulzrinne, Columbia University, discussed
Internet. This was followed by heated debate on several points: the need for a basic service with call
concerns about reservations. admission control; at what point is reservation the best
option; flow aggregation and RSVP issues. The existing
Fred Baker, Cisco, stated that QOS routing, line
consumer ISP model is based on a multiplexing model
protocols and queueing management can improve QOS. that supports 200-300 concurrent users with 10-15
He emphasized that for queuing, congestion
customers sharing an ISP line. With Internet telephony,
management algorithms, such as weighted fair queuing,
radio-like services and content pushing such a model is
serve as good tools for low speed links when there is a not sustainable. This implies that volume-based
limited number of flows. However, on the average OC-
charging is required rather than reservations.
3 link where there are hundreds and thousands of Schulzrinne mentioned that reservations are still
simultaneous flows (e.g., in a backbone network),
necessary for guaranteeing special purpose QOS, for
statistical approaches like random early detection prove services that cannot tolerate disruption. RSVP as a
to be very useful. In this case, the binding is based on IP
reservation protocol for the Internet introduces
precedence. This precedence level can be set either by a
unnecessary complications such as flow merging, delay
traffic originator or via administrative controls in the
guarantees and receiver diversity.
touters. However, like random early detection, being a
FIFO queueing algorithm, is not predictable and is not The fourth presentation in this panel session focused on
deterministic since it depends on host behavior. the use of a simple approach to QOS provisioning. Jon
Moreover, it is also dependent on bulk of traffic being Crowcroft had arranged to come in live over the
TCP. Besides these approaches, Baker suggested that Internet. However, moments before the panel was to
RSVP is a reasonable answer to QOS-enhanced traffic, begin, UCL experienced a complete brown out and Jon
particularly for edge networks and large flows. Crowcroft was source squenched! Andrew Campbell
stepped in to present the main points of Crowcrofi's

ACM SIGCOMM 113 Computer Communication Review

position. Crowcroft §uggested that were three simple real-time distributed multimedia applications, which
choices - each with their pros and cons: very often are based on networked computer systems.
These applications require time-dependent data
• over-engineering selected paths, where a processing and place huge processing demands on
particular route could be over provisioned. That distributed computer systems.
route might be shared by multiple networks as
well as users, hosts or applications, The goal of QOS architecture is to ensure the overall
presentation of multimedia data with respect to the end-
• subscription for selected terminals or addresses, to-end QOS requirements of applications. Because QOS
where a virtual private Internet offering is end-to-end, he argued, end-systems, servers,
improved QOS can be created by assigning networks, system software and applications must handle
resources based on address prefixes or IP the data accordingly. QOS model generally have a
network numbers; and number of components/viewpoints (e.g., user,
• on demand service, where a virtual circuit is set application, system and network QOS) which address
up using a signaling protocol, call admission various approaches to QOS provisioning, both
control, traffic policing, traffic accounting, and pessimistically (e.g., worst-case assumptions based on
service discrimination through scheduling. peak rate resource allocation) and/or optimistically
(e.g., resource reservation and adaptive mechanisms).
Crowcroft preferred the first choice of over-engineering Among the optimistic approaches, resource reservation
selected paths but cautioned that it might be difficult to is preferred since adaptive methods like media scaling
support new services due to such an over-simplistic and filtering cannot offer hard QOS.
view of resources. However, his contention was that the
cost of providing new services outweighs the benefit. Wolf proceeded to describe the fundamental steps in
The balance, he suggested, should be redressed toward provisioning QOS in the end-system and network based
simpler end of the QOS provisioning spectrum: that is, on signaling, resource reservation and scheduling
over-engineered selected paths. mechanisms. These steps can be divided into the QOS
negotiation phase (including QOS specification,
Andrew Campbell presented his own reservations about capacity testing and QOS calculation, resource
reservations, pointing out that both ATM and RSVP reservation) and the data transmission phase, where the
signaling are too heavy weight and complex. He also negotiated QOS is enforced by appropriate resource
revealed his concerns about scalability and stability of scheduling mechanisms. Several resource management
RSVP as a widely deployed signaling system for the components need to interact to provide end-to-end
Internet. Out-of-band signaling systems, he asserted, assurances : applications, QOS translators, admission
need to be sized, engineered and their stability controllers, resource schedulers and resource monitors.
thoroughly, investigated. Resource reservation protocols serve as a means to
The presentations sparked off an interesting debate on transfer information about reservations and to
reservations about reservations. The panelists seem to participate in the negotiation of QOS values.
agree that reservations are at times inappropriate.
Guerin, who is less comfortable with the need for
reservations, felt that reservations would only add 7. Closing R e m a r k s
complexity in the end. Andrew Campbell provided the closing remarks. He
announced that the next workshop would be held in
Napa Valley, California in May 1998 and would be
6. Q O S : W h e r e are we ? hosted by Ed Knightly from Rice University and Rich
IWQOS'97 Invited Workshop Paper: by Ralf Steinmetz Friedrich from HP Labs.
and Lars Wolf Darmstadt University The theme of IWQOS'97 building QOS into
Lars Wolf [Steinmetz,97] presented the invited distributed systems - begs the question: in today's world
workshop paper on topic of "QOS" Where are we?". He is QOS a fact or a fiction7 For those that attended the
began with an overview of terminology, issues and workshop, with its excellent technical sessions and
trends in the provisioning of QOS concentrating on panels, it is indeed, a fact. During the three day
QOS principles and the architectural issues addressed meeting, participants had heard reports about actual
by a number of research groups and standards bodies implementation results and more speculative work.
(e.g., IETF's int-serv work). Wolf reflected that over While this looks very promising, some practitioners
the past several years, QOS has evolved as a major have reservations about QOS research. Perhaps, this
field of research to support new applications such as was best typified during the panel on reservations about

ACM SIGCOMM 114 Computer Communication Review

reservations. The catchy title of the panel captures that For details on IWQOS'98, Napa Valley, see:
feeling of concern. For example, the entire reservations
• http://www-ece.rice.edu/conf/iwqos98/
panel agreed that it was very unlikely that a single
approach (e.g., RSVP) would fit all applications, with
different trade offs between complexity, level of
guarantees and scaling issues.
8. Acknowledgements
We would like to thank the Program Committee for all
Returning to the theme of whether QOS was a fact or their efforts to make this a successful workshop, the
fiction, Campbell asserted that QOS is fiction in the Center for Telecommunications Research at Columbia
sense that it is not a concrete fixture in our lives, e.g., University for hosting this year's workshop and to the
no network exists today that lets the user configure QOS local Organization Committee (particularly, Koon-Seng
on-demand. So how long do we have to wait until Lim and Oguz Angin) for all their hard work. Finally,
fiction becomes fact? Campbell concluded by telling the any errors or omissions are ours and for these, our
audience to "stay tuned" for Napa '98. apologies in advance.

ORGANIZED BY
Center for Telecommunications Research, Columbia University

IN A S S O C I A T I O N W I T H
ACM SIGCOMM
IEEE Communications Society
IEEE Signal Processing Society
IFIP Joint International Conference on Open
Distributed Processing and Distributed Platforms
IWQOS'97 WORKSHOP PATRON
Hewlett-Packard Company

PROGRAM CO-CHAIRS
Andrew T. Campbell, Columbia University
Klara Nahrstedt, University of Illinois at Urbana Champaign

PROGRAM COMMITTE
Gordon Blair, University of Lancaster Mahmoud Nagshineh, IBM
Simon Crosby, Cambridge University Elie Najm, ENST
Jon Crowcroft, UCL Max Ott, C&C Research, NEC
Hermann de Meer, University of Hamburg Guru Parulkar, Washington University
Jan de Meer, GMD Fokus Jerry Rolia, Carleton University
Alexandros Eleftheriadis, Columbia University Cormac Sreenan, Lucent Technologies
Richard Friedrich, Hewlett-Packard Company Chris Sluman, OpenlT Ltd
Michael Fry, University of Technology Hideyuki Tokuda, Keio University
Andrew Herbert, APM Ltd., UK James VanLoo, Sun Microsystems
Jean-Peirre Hubaux, EPFL Andreas Vogel, Visigenics Corp
Kevin Jeffay, UNC at Chapel Hill Andrew Watson, OMG
Brigitte Kerherv6, U. du Q. h Montr6al, Hartmut Wittig, Multimedia Software
S. Keshav, Cornell University Lixia Zhang, UCLA
Glenford Mapp, Olivetti Research Ltd Hui Zhang, Carnegie Mellon University
Steve McCanne, UCB Martina Zitterbart, TU Braunschweig

ACM SIGCOMM 115 Computer Communication Review

LOCAL ORGANISATION COMMITTEE
Oguz Angin, Laitee Cheok and.Koonseng Lim
York, USA, Pages 347-358.
9. References [Fischer, 97] Fischer S., Salem M.-V.O.M. and Bochmann
G.v., "Application Design for Cooperative QoS
[Almasberger, 97] Almasberger A., Sasyan S. and Wright S.,
Management", Proc. 5th International Workshop on Quality of
"Quality of Service (QoS) in Communication APIs", Proc. 5th
Service (IWQOS'97), Columbia University, New York, USA,
International Workshop on Quality of Service (IWQOS'97),
Pages 191-194.
Columbia University, New York, USA, Pages 237-248.
[Fukuda, 97] Fukuda K., Wakamiya N., Murata M., and
[Badrinath, 97] Badrinath B.R. and Talukdar A.K., "IPv6 +
Miyahara H., "QoS Mapping between User's Preference and
Mobile-IP + MRSVP = Internet Cellular Phone ?", Proc. 5th
Bandwidth Control for Video Transport", Proc. 5th
International Workshop on Quality of Service (IWQOS'97),
International Workshop on Quality of Service (IWQOS'97),
Columbia University, New York, USA, Pages 49-52.
Columbia University, New York, USA, Pages 291-302.
[Beeker, 97] Becker D., "Comprehensive Queueing Analysis
for a Partial Buffer System with Discrete Markovian Arrival [Hafid, 97] Hafid A., "Providing a Scalable Video-on-
Processes", Proc. 5th International Workshop on Quality of Demand System using Future Reservation of Resources and
Multicast Communications", Proc. 5th International
Service (IWQOS'97), Columbia University, New York, USA,
Workshop on Quality of Service (IWQOS'97), Columbia
Pages 89-100.
University, New York, USA, Pages 277-288.
[Blair, 97] Blair. G.S., Davies N., Friday A. and Wade S.P.,
[Huard, 97] Huard J.F. and Lazar A.A., "On End-to-End
"Quality of Service Support in a Mobile Environment: An
QoS Mapping", Proc. 5th
Approach Based on Tuple Spaces", Proc. 5th International
Workshop on Quality of Service (1WQOS'97), Columbia International Workshop on Quality of Service (1WQOS'97),
University, New York, USA, Pages 37-48. Columbia University, New York, USA, Pages 303-314.
[Bless, 97] Bless R., Jacob M. and Schmidt C., "Service- [Issarny, 97] Issarny V., Bidan C., Leleu F. and Saridakis T.,
Tailored QoS Management in High Performance Networks", "Towards Specifying QoS-Enabling Software Architectures",
Proc. 5th International Workshop on Quality of Service Proc. 5th International Workshop on Quality of Service
(IWQOS'97), Columbia University, New York, USA, Pages (IWQOS'97), Columbia University, New York, USA, Pages
179-190. 363-366.
[BorowskT, 97] Borowsky E., Golding R., Merchant A., [Jaeobs, 97] Jacobs S., Eleftheriadis A., "Adaptive Video
Scheier L., Shriver E., Spasojevic M., Wilkes J., "Using Applications for Non-QoS Networks", Proc. 5th International
attribute-managed storage to achieve QoS", Proc. 5th Workshop on Quality of Service (IWQOS'97), Columbia
International Workshop on Quality of Service (IWQOS'97), University, New York, USA, Pages 161-165.
Columbia University, New York, USA, Pages 199-202.
[Jha, 97] Jha S.K., Wright P.A. and Fry M., "Playout
[Campbell, 97] Campbell A.T., "QoS Challenges for Next Management of Interactive Video - an Adaptive Approach",
Generation Mobile Middleware", Proc. 5th International Proc. 5th International Workshop on Quality of Service
Workshop on Quality of Service (IWQOS'97), Columbia (IWQOS'97), Columbia University, New York, USA, Pages
University, New York, USA, Pages 53-56. 145-156.
[Castellanos, 97] Castellanos, J.G. and Naghshineh M., "Link [Kim, 97] Kim K. and Nahrstedt K.," QoS Translation and
Error Impact on MPEG Video over Wireless Broadband Admission Control for MPEG Video", Proc. 5th International
Networks", Proc. 5th International Workshop on Quality of Workshop on Quality of Service (IWQOS'97), Columbia
Service (IWQOS'97), Columbia University, New York, USA, University, New York, USA, Pages 359-362.
Pages 57-60.
[Knoehe, 97] Knoche H. and de Meer H., "Quantitative QoS-
[Conrad, 97] Conrad P., Amer P., Golden E., Iren S., Marasli Mapping: A Unifying Approach", Proc. 5th International
R. and Caro A., "Transport QoS over Unreliable Networks: Workshop on Quality of Service (IWQOS'97), Columbia
No Guarantees, No Free Lunch", Proc. 5th International University, New York, USA, Pages 347-358.
Workshop on Quality of Service (IWQOS'97), Columbia
University, New York, USA, Pages 315-318. [Koodli, 97] Koodli R. and Krishna C.M., "Supporting
Multiple-tier QoS in a Video Bridging Application", Proc. 5th
[Corson, 97] Corson M.S., "Issues in Supporting Quality of International Workshop on Quality of Service (IWQOS'97),
Service in Mobile Ad Hoc Networks", Proc. 5 th International Columbia University, New York, USA, Pages 133-144.
Workshop on Quality of Service (IWQOS'97), Columbia
University, New York, USA. [Lakshman, 97] Lakshman K., Yavaykar R. and Finkel R.,
"Integrated CPU and Network I/O QoS Management in an
[de Meer, 97] de Meer J., "On the Specification of End-to- Endsystem",Proc. 5th International Workshop on Quality of
End QoS Control", Proc. 5th International Workshop on Service (IWQOS'97), Columbia University, New York, USA,
Quality of Service (IWQOS'97), Columbia University, New

ACM SIGCOMM 116 Computer Communication Review

Pages 167-178. Columbia University, New York, USA, Pages 393-396.
[Lazar, 97] Lazar A.A., "Programming Telecommunication [Pope, 97] Pope S. and Webster P., "QoS Support for Mobile
Networks", Proc. 5 th International Workshop on Quality of Computing", Proc. 5th Intemational Workshop on Quality of
Service (IWQOS'97), Columbia University, New York, USA, Service (IWQOS'97), Columbia University, New York, USA,
Pages 3-24. Pages 61-64.
[Lee, 97] Lee H., "Guaranteeing Multiple Cell Loss Classes [Reed, 97] Reed D.A. and Turner K.J., "Support Components
in Shared ATM Output Buffer", Proc. 5th, International for Quality of Service in Distributed Environments:
Workshop on Quality of Service (IWQOS'97), Columbia Monitoring Service", Proc. 5th International Workshop on
University, New York, USA, Pages 77-88. Quality of Service (IWQOS'97), Columbia University, New
York, USA, Pages 255-258.
[Linnhoff-Popien, 97] Linnhoff-Popien C. and Thissen D.,
"Integrating QoS Restrictions into the Process of Service [Revel, 97] Revel D., Cowan C., McNamee D., Pu C. and
Selection", Proc. 5th International Workshop on Quality of Walpole J., "Predictable File Access Latency for
Service (IWQOS'97), Columbia University, New York, USA, Multimedia", Proc. 5th International Workshop on Quality of
Pages 225-236. Service (IWQOS'97), Columbia University, New York, USA,
Pages 401-404.
[Lu, 97] Lu S., Lee K.-W. and Bhargavan V., "Adaptive
Service in Mobile Computing Environments", Proc. 5th [Schelen, 97] Schelen O. and Pink S., "Sharing Resources
International Workshop on Quality of Service (IWQOS'97), through Advance Reservation Agents", Proc. 5th International
Columbia University, New York, USA, Pages 25-36. Workshop on Quality of Service (IWQOS'97), Columbia
University, New York, USA, Pages 265-276.
[Ma, 97] Ma Q. and Steenkiste P., "Quality of Service
Routing for Traffic with Performance Guarantees", Proc. 5th [Schmidt, 97] Schmidt D., Ott M., Parulkar G., Stadler R.
International Workshop on Quality of Service (1WQOS'97), and Vogel A., "QOS .for Distributed Object Computing
Columbia University, New York, USA, Pages 115-126. Middleware", Proc. 5th International Workshop on Quality of
Service (IWQOS'97), Columbia University, New York, USA,
[Mapp, 97] Mapp G. and Hodges S., "QoS-Based Transport",
Pages 205-207.
Proc. 5th International Workshop on Quality of Service
(IWQOS'97), Columbia University, New York, USA, Pages [Sehulzrinne, 97] Schulzrinne H., Baker F., Campbell A.T.
319-322. and Kandur G., "Reservations about Reservations", Proc. 5th
International Workshop on Quality of Service (IWQOS'97),
[Maryni, 97] Maryni P. and Pacifici G., "Real-Time
Columbia University, New York, USA, Pages 325-331.
Estimation of the Link Capacity in Multimedia Networks",
Proc. 5th International Workshop on Quality of Service [Smith, 97] Smith S. and Grace A., "A QoS Configuration
(IWQOS'97), Columbia University, New York, USA, Pages System for Distributed Applications", Proc. 5th International
101-112. Workshop on Quality of Service (IWQOS'97), Columbia
University, New York, USA, Pages 259-262.
[Moghe, 97] Moghe P. and Kalavade A., '~rerrninal QoS of
Adaptive Applications and its Analytical Computation", Proc. [Steinmetz, 97] Steinmetz R. and Wolf L., "Quality of
5th International Workshop on Quality of Service Service: Where are we?", Proc. 5th International Workshop on
(IWQOS'97), Columbia University, New York, USA, Pages Quality of Service (IWQOS'97), Columbia University, New
369-380. York, USA, Pages 211-222.
[Nishio, 97] Nishio N. and Tokuda H., "Simplified Method [Verma, 97] Verma S., Pankaj R.K. and Leon-Garcia A.,
for Session Coordination Using Multi-level QOS "QoS Based Multicast Routing for Multimedia
Specification and Translation", Proc. 5th International Communications", Proc. 5th International Workshop on
Workshop on Quality of Service (IWQOS'97), Columbia Quality of Service (IWQ.OS'97), Columbia University, New
University, New York, USA, Pages 335-346. York, USA, Pages 127-130.
[Oeehslin, 97] Oechslin O., "Worst Case Arrivals of Leaky [Waldegg, 97] Waldegg D.B., "A Temporal QoS based CPU
Bucket Constrained Sources: The Myth of the On-Off Scheduling Model for Multimedia Applications in General
Source", Proc. 5th International Workshop on Quality of Purpose Operating Systems', Proc. 5th International
Service (IWQOS'97), Columbia University, New York, USA, Workshop on Quality of Service (IWQOS'97), Columbia
Pages 67-76. University, New York, USA, Pages 157-160.
[Oh, o,/] Oh. S., Sugano H., Shimojo S., Miyahara H., [Zinky, 97] Zinky J.A. and Bakken D.E., "Managing
Fujikawa K., Matsuura T., and Arikawa M., "A Dynamic QoS Systemic Meta-Data for Creating QoS-Adaptive CORBA
Adaptation for Networked Virtual Reality', Proc. 5th Applications", Proc. 5th International Workshop on Quality of
International Workshop on Quality of Service (IWQOS'97), Service (1WQOS'97), Columbia University, New York, USA,
Columbia University, New York, USA, Pages 397-400. Pages 251-254.
[Ott, 97] Ott M., M~chelitsch G., Reininger D. and Welling
G., "Adaptive QoS and ifi Multimedia Systems", Proc. 5a~
International Workshop on Quality of Service (IWQOS'97),

ACM SIGCOMM 117 Computer Communication Review