%%%%%%%% Article style document. %%%%%%%% %\documentstyle[fullpage,times,psfig,11pt]{cmu-art} \documentstyle[fullpage,times,psfig,11pt]{article} %\documentstyle[titlepage,times,11pt,picture]{article} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Get the support for EPSF in subdirectory % Will typeset incorporated pictures on any unix system % Will typeset and preview incorporated picture on macintoshes (Textures) % if custom EPSF.sty is installed. Available from tomstr@cs.cmu.edu %\input epsf.sty %\def\mac{mac} %\ifx\sys\mac\def\figpath{:fig:} % \else\def\figpath{./fig/} %\fi %\def\figpath{../graph/} \def\figpath{./} %%%%%%%% Margin declarations. %%%%%%%% \oddsidemargin 0in \evensidemargin 0in \topmargin 0in \textwidth 6.5in \textheight 8.75in %%%%%%%% Macro definitions. %%%%%%%% % allows multiple line comments enclosed by brackets. \newcommand{\comment}[1]{} % define un-numbered footnote for all kinds of disclaimers and support messages \makeatletter \long\def\unmarkedfootnote#1{{\long\def\@makefntext##1{##1}\footnotetext{#1}}} \makeatother % allows multiple line comments enclosed by brackets. \newcommand{\ignore}[1]{} % allows multiple line comments enclosed by brackets (more visible). \newcommand{\COMMENT}[1]{} % allows multiple line comments enclosed by brackets. These are to be % omitted from the current version of the document but should be included % in the final version. \newcommand{\OMIT}[1]{} % changes output to double spaced. \newcommand{\doublespace}{\baselineskip 22pt} % changes output to single spaced. \newcommand{\singlespace}{\baselineskip 16pt} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % DOCUMENT STARTS HERE % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \begin{document} % set line spacing to \singlespace or \doublespace. \singlespace % choose the output format for the bibliography. % @use(bibliography = /afs/cs/usr/prs/bib/nectar.bib) % @use(bibliography = /afs/cs/usr/prs/bib/network.bib) %\bibliographystyle{cmu-bib} %%%% to use\bibliographystyle{plain} %\bibliographystyle{alpha} %%%%%%%% TITLE PAGE %%%%%%%% \title{Routing High-bandwidth Traffic in Max-min Fair Share Networks} \author{ Qingming Ma\ \ \ \ \ \ \ \ \ Peter Steenkiste\ \ \ \ \ \ \ \ \ Hui Zhang\\ School of Computer Science\\ Carnegie Mellon University\\ Pittsburgh, PA 15213\\ \{qma, prs, hzhang\}@cs.cmu.edu\\ %(412) 268-8079 } \date{} \abstract{ Add abstract } \maketitle \unmarkedfootnote{ Need Credit Net research credit. } \newpage \section{Introduction} A growing number of applications that communciate over networks is being developed. Examples include distributed scientific computations, World-Wide Web, and video conferencing. In many cases, low message latency is required and since many applications use large messages, this often translates in a requirement for high bandwidth. Examples include scientific computation, where data sets as large as a few gigabytes need to be processed in a few seconds, and Web browsers, which are retrieving increasingly larger images and animation files from a few kilobytes to a few megabytes. Fortunately, advances in broadband networking and switch-based networking technology, including ATM and switched-LANs (e.g., switched-Fibrechannel, switched-Ethernet, and switched-FDDI), has greatly increased network bandwidth available to applications. While most networks support only "best-effort" communication, the diversity in applications has started the development of a variety of communication services that can better meet the communication demands of applications. Many people have looked at services classes that provide quality of service guarantees to applications between real time requirements. Even within the class of best-effort communication, it might be necessary toe define multiple service classes, because applications often want to see different traffic parameters optimized. For example, most data traffic falls into either of the following two classes: \begin{itemize} \item {\bf Low-latency traffic.} Applications operate on relative small amounts of data and are sensitive to end-to-end message delay. Since a message typically consists of one or a small number of packets, the latency of individual packets must be minimized. A typical example is RPC. \item {\bf High-bandwidth traffic.} Applications are sending a large amount of data and the elapsed time for transmission depends on the average throughput, rather than the delay of individual packets. Applications include ftp, Web browser, and I/O intensive scientific computation. This class of applications is growing rapidly. \end{itemize} For best effort traffic, network resources are shared dynamically by many applications. Resource allocation is typically done at two levels. At the higher level, a {\em routing} entity directs traffic along less congested paths to balance the network load and achieve high performance. At a lower level, {\em congestion control} mechanisms limit how much traffic enters the network, to minimize buffer overflow and avoid inefficient network use. The following mechanism are typical of what is used in today's internet: \begin{itemize} \item {\bf Shortest-path routing.} Average packet delay is used as link metric. The hypothesis is that short delay is a good indication of low link utilization, and thus of high available bandwidth. \item {\bf Window-based congestion control.} Rate adjustment is performed by the sender based on end-end feedback. For example, the sender will reduce its rate based on an implicit indication of congestion, such as packet loss or increased in end-end packet delay. \end{itemize} Recent developments in both the areas of routing and congestion control suggest that defining separate services for low-latency and high-bandwidth might be beneficial. First, it has been observed recently \cite{??} that, due to the TCP dynamic adjustment of sending rate, the average packet delay is not affected much by network load. As such, the shortest path routing policy is likely to be more effective in minimizing packet latency than in maximizing sustained throughput. Second, recent changes in congestion control and scheduling mechanisms have made it feasible to provide routing algorithms with information on per-connection bandwidth instead of per-packet latency. An example is the rate-based per-connection flow control adopted by the ATM forum: the EPCRA algorithm gives every source an explicit rate at which it is allowed to transmit. Fair sharing scheduling policies for packet networks can provide the same type of information. These two developments suggest that high bandwidth applications could benefit from a high-bandwidth service that uses bandwidth information to optimize sustained throughput. Note that a good initial route selection is important in connection-based networks such as ATM since all packets use the route, and rerouting can be relatively expensive. Routing data traffic in a max-min fair share network is challenging due to the dynamic sharing nature of the max-min fairness. It differs from routing in the traditional ``residual bandwidth'' or ``link capacity'' network model in two ways. First, the available bandwidth of a connection no longer held for the life of the connection, it can obtain more or less bandwidth at any time when a connection leaves or enters the network. Second, the performance goal used in ``residual bandwidth'' model, such as ``bandwidth routed'' and ``blocking rate'' is no longer applicable. Instead, it is the average throughput of all connections the main concern. In this paper, we compare the effectiveness of a number of routing algorithms in routing high-bandwidth traffic, assuming a network model based on max-min fair sharing. Our evaluation is based on simulation of a number of traffic loads on several network topologies. We look both at single-path and multi-path algorithms. For multi-path routing, we introduce a prioritized multi-level max-min fairness model, in which multiple paths are assigned different priority. This approach prevents a multi-path connection from grabbing an unlimited amount of bandwidth by using a large number of paths, i.e. additional paths use only unused bandwidth and do not affect the bandwidth available to primary paths. The remainder of this paper is organized as follows. In Section \ref{section:related} we discuss related work, and in Section \ref{section:model} we present the network model used by the routing algorithms. We describe the routing algorithms we evaluated in Section \ref{section:algorithms} and our simulator in Section \ref{section:simulator}. Section \ref{section:results} presents our results. We conclude in Section \ref{section:conclusion}. \section{Related work} \label{section:related} \subsection{Service models} Work on introducing new no-real-time data service models David Clark: extending the Interent by adding features that permit allocating different service levels to differnet users. Pay more to get more. - hard to apply to public networks Scott Shenker: ASAP vs AMAP - no detail work ATM Forum: ABR and UBR \subsection{Routing} Early routing work for APARNET Dynamic and adaptive re-routing: Dynamic or adaptive re-routing is a way to reblance network load and make better use of network resources and get increased network throughput. However, it causes problems in practice. The earlier APARNET experience suggests that doing adaptive routing frequently can cause the well-know oscillation problem because each swicth makes routing dicision indepentently. Also, packets for the same source and sink pair can be delivered out of order, which may greatly affect TCP window size. In a network using virtual-circuit, unless routing decision is decided by a centralized routing server, the simiular oscillation problem can still occur. Moreover, signalling overhead of tearing down old and setting up connections can not be neglected, although data set is large in high-bandwidth traffic. In the studies by Rudin and Mueller and by Grandjean, it was shown that for circuit-switched networks, dynamic routing can improve network throughput or delay, but only over a limited range of network utilizations, and can worsen performance at very high utilizations. alternative path when the network is congested multi-path routing Other misc: burst reservation Routing for different network models (different goals) metric used for routing: hops, residual bandwidth, ... Routing algorithms Two main classes of routing algorithms have been studied in the literature. The first class, generally called the shortest-path algorithms. provides a least-cost path between a given source-destination pair, where the ``cost'' of path is typically defined as the linear sum of the costs of each hop in a given path. Routing algorithms used in most existing networks fall into this category, althgough cost of a hop is defiend from network to network. The second class of routing algorithms are optimal to minimize the network average network-wide delay. The use of this performance leads to multi-path routing. Routing data traffic for the network with underlying congestion control based on max-min fairness is new. \section{Network model} \label{section:model} The model of the network used by the routing algorithm can have a significant impact on the results. In this section we describe our network model. We describe how competing connections share bandwidth, the traffic classes, and costs associated with routing related operations. \subsection{Max-min fair sharing} We assume that bandwidth sharing is based on max-min fairness. Intuitively, with max-min fair sharing, all existing connections in the network are treated fairly in the sense that the available link bandwidth is evenly spilt among all connections who share the link. If some connections can not make full use their fair share because they are bottlenecked by some other links, the unused portion is split fairly among those who can use more bandwidth. Given a network $\langle V, E\rangle$, a set $S$ of connections, and a bandwidth function $B: E\rightarrow R$. Let $S_{e}$ be the set of connections in $S$ that use link $e$ and $B(e)$ the total bandwidth available to all connections in $S_{e}$. The maximal rate $r_{s}$ that a connection $s$ can achieve within max-min fair sharing is called its max-min rate; once a connection gets its max-min rate, it is said to be saturated. A link $e$ is said to become bottleneck if all connections use the link are saturated. A simple centralized algorithm to calculate the max-min rate is based on gradually increasig the rate of unsaturated connections until all connections are saturated. The iterative algorithms can be summarizes as follows: \begin{enumerate} \item For every connection $s$, $r_{s}$ is initially set to $0$ and made unsaturated. Links that are not used by any connection in $S$ are marked bottlenecked immediately and are removed from the pool. \item Find a link $e$ with a minimal incremental rate, defined by \[ {\sf inc}_{e} = \frac{B(e) - \sum_{e\in S_{e}} r_{e}}{\rm number\ of\ unsaturated\ connections\ using\ link\ e} \] \item Add this incremental rate to all unsaturated connections. The link $e$, and possible other links will become new bottlenecks. Remove all new bottlecks and mark all connections use these bottlenecks saturated. \end{enumerate} The above algorithm results in ``basic'' fair sharing, but many other versions have been defined. A first variation is ``weighted'' fair sharing, which allows which allows different connections to get a different fair share, e.g. one connection gets twice as much bandwidth as another. A second variation is to break up the available bandwidth in different classes, and to apply the fair sharing algorithms within each class. This approach provides isolation between traffic streams in different classes but sharing inside a class. We will use the basic max-min fair sharing, both because it is simpler, and because it is likely to be the first fair sharing algorithm to be supported by commercial systems. Our notion of fair sharing matches what is currently being standardized for ABR traffic in ATM by the ATM Forum, although the ATM Forum does not mandate a specific definition of fairness. Similarly, many research groups are looking at approximating max-min fair sharing behavior inside packet networks. \subsection{Traffic classes} Our network model supports three classes of traffic: \begin{itemize} \item {\em real time traffic} reserves a certain amount of bandwidth for a certain time interval. The bandwidth used are a relatively small fraction of the link bandwidht, and real time traffic is only responsible for a small portion of the total link bandwidth. {\bf Is there really real-time traffic in teh simulations?} \item {\em low latency traffic} consists connections that carry an amount of data that is lower than some threshold {\bf LBvsHB}. A low-latecny connection typically sends a few kilabytes to less than one megabyte, and requires low sending rate due to the end system limitation (e.g., human interaction and low disk bandwidth). Thus, we assume that a low-latency connection is upbounded by a peak sending rate. \item {\em high bandwidth traffic} consists connections that carry an amount of data that is larger than the threshold {\bf LBvsHB}. A high-bandwidth connection can send a few megabytes to one gigabyte of data. A traffic source can typically make full use of available bandwidth. \end{itemize} The partition between high-bandwidth and low-latency can be vary and depends on individual applications. In this papaer, we assume that any request below {\bf LbvsHB} is sent through low-latency traffic, others through high-bandwidth traffic. For the non-real time traffic connections, the performance metric is the total elapsed time to send the amount of data for that connection, which proportional to the average bandwidth for the duration of the connection. \section{Routing algorithms} \label{section:algorithms} Our performance goal for high-bandwidth routing is to maximize the average throughput of all connections. In section we discuss the routing algorithms we consider and we also briefly discuss routing protocol issues: who make the routing decision, what routing information is needed, and how the information is distributed. \subsection{Single path routing algorithms} One can view the transmission of a data stream over a connection as a pipeline. Each hop that is part of the connection represents a pipeline stage, and multiple packets can be traveling through the pipeline at the same time. The elapsed time needed to send a certain set of packets through the network will be determined by the delay introduced by the pipeline, and by the rate at which the data flows through the pipeline. The delay is determined by the number of pipeline stages (hops) and by the delay introduced by each stage (propagation and queueing delay). For small data transfers, the pipeline delay experienced by individual packets dominates the elapsed time, and our routing measure will be minimizing the number hops, which widely used as an approximation for delay. For large data transfers, the average rate will typically be more critical, and the routing algorithm we use is based on the bandwidth that a new connection can expect on each link. This expected bandwidth is calculated using the max-min fair sharing algorithm described in the previous section. However, there are some intuitive arguments why picking the ``widest'' path is bad idea. The reason is that the widest path might have a large number of hops, which means that the path might be very resource intensive, and also increases the chance that traffic that enters the network later will lower the available bandwidth. As a result, we will consider a number of algorithms that try to strike a balance between finding the ``widest'' and the ``shortest'' path. Specifically, we evaluate the following set of algorithms: \begin{itemize} \item {\bf Widest-shortest path}: the path with the smallest number of hops. If there are multiple ``shortest'' paths, the one with the maximum max-min rate is selected. This algorithm is used for low bandwidth traffic, and as the base case for high bandwidth traffic. \item {\bf Shortest-widest path}: the path with maximum max-min rate among all possible paths. If there are multiple ``widest'' paths, the one with the one with the fewest number of hops is selected. \item One can define a family of algorithms that covers the spectrum between widest-shortest and shortest-widest path. Let $r_{1}, \cdots, r_{k}$ be the max-min rates of a path $P$ with $k$ hops. The polynomial distance of $P$ with power $n$ is defined as \[ {\rm dist}(P, n) = \sum_{i = 1}^{k} \frac{1}{r_{i}^{n}} \] A shortest path algorithm based on the distance ${\rm dist}(P, n)$, approachest the shortest-widest path as $n$ becomes very large and is equal to the widest-shortest path algorithm for $n = 0$. We will use look at the performance of intermediate algorithms corresponding to $n = 0.5$, $n = 1$, and $n = 2$. Note that ${\rm dist}(P, 1)$ can be interpreted as the bit transmission delay from the traffic source to destination if the connection should get the rate $r_{i}$ at hop $i$. This delay is different from the measured delay used in some shortest path implementations in two ways. First, the focus is on bit transmission delay (i.e. bandwidth) instead of total packet delay. Second, the measure is for data belonging to a specific connection instead of a link average. \end{itemize} \subsection{Multipath routing} {\bf How to use multi-path without destroy max-min fairness:Figure \ref{fig:multi:path}. a prioritized max-min scheme is introduced, need no scheduling support, only support from source to set priority } \begin{figure}[htb] \centerline{\psfig{figure=fig/multipath.eps,height=1.8in}} \caption{Prioritized multipath routing} \label{fig:multi:path} \end{figure} \subsection{Routing protocol} Since the high bandwidth routing algorithms depend on global network information, we expect that they will be implemented as a link state protocol (e.g., IDPR \cite{RFC1478}, OSPF \cite{ospfv2}, and IS-IS \cite{isis,perlmanBook}), as opposed to distance vector (e.g., BGP \cite{BGP}, IDRP \cite{IDRP}, and RIP \cite{hedrick}). We also expect that source routing is used, as opposed to hop-hop routing. A critical question is how much information has be propagated from the switches in the network to the routing entity. In theory, to calculate the rate for a new connection, we need to know the network topology and paths of all existing connections so as to mimic max-min fairness algorithm. In reality, this may not be practical and unnecessary since the amount of routing information can be very large when if thousands of connections exist in the network. We propose an approximate way to calculate the rate of a new connection, under which each switch only needs to send a vector of counters for the number of connections in discrete bandwidth intervals. For example, assume that the link bandwidth is $155 Mb/s$. For each output port of a switch, we have a vector, and the entries of the vector is number of connections with a certain amount of bandwidth assigned. This significantly reduces the amountof information that has to be propagated in the network. {\bf Is this what is used in the simulations? Add example??} \section{Simulation approach} \label{section:simulator} In this section, we first briefly describe our simulator and we then parameters and simulator inputs used for the simulations. \subsection{Simulator Design} We designed and implemented a connection-level event-driven simulator. The simulator allows us to specify a topology and code modules representing the various traffic sources, the algorithm used by the switches for distributing bandwidth between connections sharing links, and the routing algorithm. For the results we report, the switches use max-min fair sharing algorithm described earlier. The simulator simulates connection setup and tear-town as follows. An incoming request presents the number of bytes to be sent. Paths are found using the routing algorithm and connections are set up. After completing sending all data, connections are torn town. The rates of all connections are dynamically adjusted as connections are added and removed. The main performance measure reported by the simulator is the average bandwidth achieved by each connection, plus the the variance of actual bandwidth obtained by different connections. For multi-path routing, we compare the average bandwidth gained by those connections that use more than one path. \subsection{Simulation parameters and inputs} {\bf Topology.} The network topology being picked has big impact on simulation results. For this resaon, we pick three topologies as shown in figure ... Two of them with 8 switch nodes and one with 16 swicth nodes. The difference between the two 8-node topologies is that one is the symmetricity and connectivity. The 16-node topology will let us to find out how performnce mesaure changes when the network grows large. In particular, we assume that a fixed number of host nodes are attached to each switch and traffic arrives from host nodes. This allows us to explore the network performance by partitioning host nodes into clients and servers so as generate more realistic and unbalanced traffic load. Also, it allows us to take into consideration the cases when the end sysems are the bottleneck. {\bf Traffic load.} In our simulation, we use two types of traffic, namely high-bandwidth and low-latency traffic. \begin{itemize} \item The exact partition between high-bandwidth and low-latency depends on application. In our simulation, we assume that any request below one megabytes is sent through low-latency traffic, others through high-bandwidth traffic. We also show the sensitive of this this cut-off between high-bandwidth and low-latency tarffic to the simulated network performance. We also assume that, for high-bandwidth traffic, the traffic source can make full use of bandwidth assigned by the network. For low-latency traffic, traffic source suggests a peck rate (PCR) and can make full use of assigned bandwidth that is no higher than the peak rate. Data are assumed to flow out from traffic source smoothly. \item Traffic arrival: poisson? \item Ratio of bytes in high-bandwidth and low-latency traffic. \end{itemize} {\bf Routing and connection establishment costs.} Since all communication is connection based, we charge a per-connection cost that is proportional to the number of hops included in the connection. The cost is {\bf HopCost} for each hop. Since we expect that the routing for high-bandwidth connections might be fairly expensive, we will charge a cost for routing for high-bandwidth connections. The cost is represented by {\bf RoutingCost}. Both the routing and the connection cost are included in the elapsed time of the connection since they are in the critical path of getting the data to the receiver. These costs will, in effect, reduce the average bandwidth observed by applications. {\bf what is PP Delay??} \begin{tabular}{|c|c|c|c|c|} \hline {\bf HopCost} & {\bf RoutingCost} & {\bf LBvsHB} & {\bf LinkBth} & {\bf PP Delay} \\ \hline 3ms & 10ms & 1000 KB & 155 mb/s & 20ns \\ \hline \end{tabular} = routing algorithms: shortes-widest, widest-shortes path, hybrid = routing information distribution interval = client-server vs uniformal host nodes = message size distribution: uniform vs long-tail = number of paths = ratio of high-bandwidth over low latency traffic \section{Simulation results} \subsection{Single-path algorithms} In Figure \ref{}, it is assumed that the network load is uniformally distributed and there are $90\%$ of bytes of data are in high-bandwidth traffic. The network topology is relatively symmetric in the sense that there is no obvious bottleneck link. \begin{itemize} \item {\bf Shortest}-${\rm dist}(P, 1)$ performs well in both cases of light load and heavy load. \item {\bf Shortest}-${\rm dist}(P, 1/2)$ is close to {\bf Shortest}-${\rm dist}(P, 1)$ when the load is heavy and slighltly worse than {\bf Shortest}-${\rm dist}(P, 1)$ when the load is light. \item Widest-shortest path lags behind {\bf Shortest}-${\rm dist}(P, 1/2)$ in both light load and heavy load. \item {\bf Shortest}-${\rm dist}(P, 2)$ performs as well as {\bf Shortest}-${\rm dist}(P, 1)$ when the network load is light and slightly worse than widest-shortest-path when the network load is heavy. \item Shortest-widest path performs very well when the network load is light and much worse than all others when the load is heavy. \end{itemize} In general, the average throughput decreases linearly while the network load increases, but with different gradient. Shortest-widest-path decreases most quickly. In Figure \ref{}, the percentage of bytes in high-bandwidth traffic is changed from $90\%$ to $50\%$. We can draw similar conclusion as for Figure \ref{}, except that the scale of the performance difference is down a little bit. It is reasonable, since more bytes of data loads are shipped through low-latency traffic, therefore, less interference to high-bandwidth connections using more hops. It is interesting to observe that, in Figure \ref{}, when the network topology is less symmetric, where there is a bottleneck link that sits in many shortest paths. Widest-shortest-path becomes a much worse choice than all others. Hence, the performance of shortest-widest-path and widest-shortest-path can perform dramatically good and bad, depending on particular topology. Shortest-${\rm dist}(P, 1)$-path is much more stable. It is also interesting to notice that, when the network load is extreamely heavy, the performance difference between these different algorithms disappears. The resaon is that, in such extreamely heavily loaded cases, the number of connections over any link is very large, the obtainable rate is almost the same from any link. Path selection becomes much less sensitive to the obtainable rate and much more sensitive to the number of hops in a path. Eventually all connections tend to pick widest-shortest paths. In Figure \ref{}, the traffic load is no longer uniformally distributed. Host nodes are partitioned into clients and servers. {\bf Shortest}-${\rm dist}(P, 1)$ continues to outperform all others. To make it even more realistic, we assume that servers use $622 mb/s$ access link in Figure \ref{}. The performance results are consistent with $155mb/s$ access link, but larger scale. By looking into all these Figures carefully, we can draw an important conclusion: when the amount of high-bandwidth traffic increses, then performance impact of different routing algorithms become more significant. widest-shortes path for low-latency traffic, widest-shortes, shortest-widest, and hybrid for high-bandwidth change routing information distribution interval, change topology (both size and connectivity), change traffic load and load ratio of high-bandwidth over low-latency traffic ratio change all client mode to client server mode \subsection{Multi-path routing} high connectivity a fixed algorithm from the previous subsection change network load and load ratio, change topology (increasing connectivity), change all client mode to client-server mode change number of paths to use \subsection{Sensitivity analysis} We have to look at how sensitive the results are to some of the key parameters. The list of key parameters includes: hopcost, routingcost, LBvsHB, PCR for LB traffic. \section{Conclusions} \label{section:Conclusions} \subsection{Dynamic rerouting} Pro: better resource allocation When should source reroute? Cons: oscillation (if multiple flows reroute at the same time. overhead in practice \subsection{Dynamic rerouting} fixed algorithm, single path change the condition when rerouting should happen change topology (connectivity) Separate high bandwidth traffic from less critical traffic + motivation Assuming fair sharing + motivation Also evaluating multipath + motivation widest-shortest versus shortest-widest influence of topology+traffic on results cut off between the two traffic classes effect of using fair sharing multi-path influence of topology+traffic on results effect of "fairness" algorithm effect of using using fair sharing - Algorithms shortest-widest versus widest-shortest sources of overhead hybrid multi-path sources of overhead adopted by PNNI, and be used to find a pat on per-connection bese. \begin{itemize} \item {\bf Variety} Applications requires quality-of-service guarantee, e.g., digital audio and video, interactive learning, and video conference \item {\bf Size} The amount of data being transferred is becoming increasingly larger due to the growing application size and complexity. For example, in scientific computation (e.g., ...), data sets being processed can be as large as a few gigabytes in a few seconds. Also, The size of files (including images, pictures, and animation files) retrieved through FTP and Web browsers is also growing, from a few hundreds of kilabytes to megabytes. \end{itemize} - without adding much complexity to the existing service models - switch-based - virtual circuits - high link bandwidth \end{document}