Wide-Area Content-based Routing Mechanism
Arindam Mitra, Muthucumaru Maheswaran
Advanced Networking Research Laboratory
University of Manitoba
Winnipeg, MB R3T 2N2
arindam, mahes @cs.umanitoba.ca

arueda@win.trlabs.ca
Abstract
Content networking is an emerging technology, where
the requests for content are steered by content routers that
examine not only the destinations but also content descrip-
tors such as URLs and cookies. In the current deployments
of content networking, content routing is mostly confined
to selecting the most appropriate back-end server in virtu-
alized web server clusters. In this paper, we present an ar-
chitecture for wide-area content routing. The architecture is
based on tagging the requests at ingress points. The tags are
designed to incorporate several different attributes of the
content in the routing process. Simulations are carried out
to compare the performance of the proposed scheme with a
DNS-based content access scheme.
1. Introduction
As business critical applications are deployed over the
Internet, it is becoming necessary for the service or content
authors to ensure that the clients are receiving their service
or content at a satisfactory level to preserve brand equity
and customer loyalty. This issue has prompted a flurry of
activities in the area of server-initiated caching and replica-
tion.
One emerging technology in this area is content-aware
networking, where a new generation of routers specifically
designed to address the unique requirements of Web traf-
fic are used. These content-based routers have the ability to
route traffic flows based on some attribute of the content be-
ing requested such as URL or cookie values. The previous
generation of routers that routed traffic flows based on pa-
rameters such as destination IP, protocol ID, and transport
port number cannot differentiate, for example, between a
CGI script request or a streaming audio request. However,
these two requests have very different quality of service
(QoS) requirements. Content-based routers, on the other
hand, provide flexibility in defining policies for prioritizing
0-7695-1926-1/03/$17.00 (C) 2003 IEEE
Jose A. Rueda
TRLabs, Winnipeg
10-75 Scurfield Boulevard
Winnipeg, MB R3Y 1P6
traffic and for balancing load and content among servers so
that service level agreements can be met.
One way of deploying content networking technologies
is to interpose a content router between a client and the web
server or server cluster. In this approach [18], the content
router is used as a load balancer for the back-end servers
and the content router should be all knowing to make opti-
mal decisions. This approach can be generalized to arrive at
the second approach [20], where name servers may be used
to select the most appropriate server site depending on the
geographical locations of the clients, servers, and network
and server load status. The first approach has the disadvan-
tage of lack of scalability and a single point of failure and
the second approach is inefficient in handling portions of a
site.
In this paper, we present a wide-area content-based rout-
ing architecture. The routers perform content-based rout-
ing by examining the content carried by the traffic flow. In
the proposed architecture, we describe a novel content de-
rived tagging scheme where content is examined only once
and the subsequent forwarding operations are based on the
tag prepended at the ingress router. We perform simulation
studies that use various different metrics to compare the per-
formance of the proposed architecture with a DNS-based
content access scheme which is the most popular content
access scheme on the Internet [8].
The rest of the paper is organized as follows: Section 2
presents the overall architecture for our content-based rout-
ing technique and the content characterization and classifi-
cation schemes we had used to create the tags. The simula-
tion results are presented and discussed in Section 3 and the
Section 4 presents related work in the field.
2. Overall Architecture
This section describes our content-based routing archi-
tecture called the protocol independent content switching
(PICS). In PICS, client and server sites are interconnected
through an overlay network called virtual content network
(VCN). The client and server sites connect to the VCN us-
ing gateways that are themselves content-based routers. Be-
cause the routers at the outer edge of the VCN first examine
the content, we refer to these routers as the content edge
routers (CERs). The PICS architecture allows the client
or server sites to connect via multiple CERs to the VCN,
thus enabling load balancing and preventing single-point
failures.
The interior of the VCN has content-based switching
routers (CSRs). The CERs encapsulate the incoming pack-
ets using a content header. The content header contains a
content derived tag that is used to uniquely identify a con-
tent within the VCN. The CSRs at the VCN’s core steer the
content requests from the ingress CER to the egress CER
based on the tags in the content header. The CSRs are sim-
ple switches that support a single forwarding component
(i.e., algorithm for tag-based forwarding). The CERs, on
the other hand, are also responsible for characterization and
classification of content.
An example VCN setup is shown in Fig. 1 where
two client sites Client-A and Client-B are connected
to the server sites Server-1 and Server-2 through a
VCN. Client-A is connected to the VCN through two
CERs whereas Client-B is connected through a single
CER. Each of the server sites, Server-1 and Server-2
are also connected to the VCN through a single CER. The
VCN consists of several CSRs that are inter-connected by
content switched paths (CSPs). The CSPs also connect the
CERs to the CSRs. The CSPs are virtual paths that are cre-
ated when the VCN is initialized. As client and server sites
enter and depart the VCN, the corresponding CSPs may be
created and removed, respectively. The CSPs are used to
route the content requests to the appropriate server site and
the content back to the requesting client site.
Content-based Switching Router Content Switch Pathway
(CSR) (CSP)
Content Routing Plane
Client B
Client A
Physical Routing Plane
Figure 1. Virtual content network.
sites towards the client sites. The status updates are at-
tenuated appropriately to reflect the network congestion as
they cross the CSPs. The CSRs and CERs compute route
fractions based on the status updates. The routing fractions
are used by the CSRs and CERs to route the content request
towards the most appropriate content server.
2.1. Functional Components
2.2. Server Sites
The server sites hold the content that is requested by the
clients. The simplest server site can be a CER and an at-
tached server that serves the corresponding content. In the
general case, multiple CERs can act as front-ends to the
servers that are in the site. The servers can be either ori-
gin servers or surrogate servers. An origin server holds the
original copy of the content, i.e., if there are no replication
of content, this server will have the only copy. A surrogate
server is a server that “rents” its resources so that content
from origin servers can be replicated on it. The owner of
the origin server that is referred to as the content provider
is responsible for “paying” the rent for the surrogate server.
VCN does not differentiate between surrogate and origin
servers. the requested content can be delivered either by an
origin server or a surrogate server.
2.3. Client Sites
The client sites have the clients that are requesting con-
tent. Typically a client site may have a single CER connect-
ing it with the VCN. However, in some situations it may
be useful to have multiple CERs. Some advantages of such
configurations include lack of single point of failure and op-
portunity for traffic shaping within the VCN. Depending on
the local routing policies, the CERs can be allocated traffic
according to the quality of service (QoS) considerations.
2.4. Virtual Content Network
Content-based Edge Router
(CER) The VCN is a graph with the CERs and CSRs as the
nodes and the CSPs as the edges. It provides a Content
Delivery Network for the client and server sites that need
“speedy” content delivery. the clients and servers would
subscribe to the VCN by having content routing enabled
gateways (i.e., CERs as gateways) and by connecting such
Server - 1 gateways to the VCN core.
As mentioned above, if a request arrives for a content
Server - 2 that is handled by the VCN, the CER will identify the con-
tent and tag it. The tag is generated by a combination of
content and policy information. Fig. 2 illustrates the overall
request tagging process in the CERs. The IP packet is en-
capsulated by the PICS content header at the ingress CER
and is restored by the egress CER. The CERs extract the
content descriptors such as URLs or cookies from the re-
quests and use a content-based forwarding information base
(CFIB) to determine the corresponding content derived tag,
i.e., the CFIB provides the content-to-tag bindings. The
0-7695-1926-1/03/$17.00 (C) 2003 IEEE
CFIB is organized hierarchically based on the popularity the client request is resolved and mapped on to a specific
of the content to enhance the scalability. The CFIB is re- IP address and in the second phase the request is routed to
organized periodically by the content classification process this IP address using traditional Internet routing protocols
to reflect additions and deletions in content subscriptions. such as open shortest path first (OSPF) [9]. The process
The content classification process leverages the properties is illustrated in Fig. 4(a). The alternative is to unify the
of the content identified by the content characterization pro- server and path selection process. The PICS scheme pro-
cess. The content classification process uses the content at- posed here splits the name resolution process in two phases.
tributes to create content equivalence classes (CECs). A In the first phase, content is resolved onto a fixed format
CEC is identified by a tag that is derived based on the at- tag using the content attribute values. In the second phase,
tribute values of its constituent documents. The motivation the tag is further resolved using a highly distributed “flat”
for grouping documents and using content-based tags is to routing network to reach the eventual server for the content.
improve scalability of the content-based routers. The process is illustrated in Fig. 4(b). The overall routing
algorithms is shown Fig. 5.
Content Edge Router

IP Packet Tagged Packet (a) URL Name Resolver IP Route

Server selection Path selection

1. Extract the URL/Content
Information Content Header IP Packet

Content-based Forwarding 5. Attach as Content Header

Information Base

(b) URL Tag Resolver Tag Route

Content Tag Next Hop

Server selection
2. Check binding information Server set selection
3. Select content tag for request Path selection
4. Select the nexthop

Figure 2. Content tagging process in a CER. Figure 4. (a) IP resolution scheme (b) Tag res-
olution scheme.
Similar to the CERs the CSRs also maintain content-
based routing tables. Unlike CERs, the CSRs do not have
any content-to-tag binding information. The CSRs assume
Find the tag
that the appropriate tags are already created by the CERs. Client
request
CER
from FIB and
Find the Forward
examines nexthop from encapsulated
encapsulate
The CSRs use the content-based routing tables to steer the client
the request
content packet to
request routing table nexthop
requests toward the appropriate server side CER. Fig. 3 packet

shows the routing process in the CSRs.

Content-based Switching Router

nexthop is a No is nexthop
Encapsulated Packet Encapsulated Packet CSR a CER

Yes
1. Read Content Header Hosting server
deliver content to
Content Routing Table
3. Forward Packet Decapsulate
client along the Forward to nexthop is the
the request
Tag Next Hop same path hosting server egress CER
packet
traversed by the
request
2. Select Next Hop

Figure 5. Content-based forwarding process.

Figure 3. Routing process in a CSR.

A content accessing scheme essentially implements two 2.5. Content Characterization and Content Classi-
major functions: (a) server selection: selecting a server site fication for Content-based Routing
and a server within the site that can serve the requested
document and (b) path selection: selecting the path along Content characterization is process that identifies the
which the selected server delivers the content to the clients. key attributes of content which can be used to generate an
Most of the existing content accessing approaches execute accurate description of the content and its resource require-
these functions in two different phases. In the first phase, ments from the perspective of of content-based routing and

0-7695-1926-1/03/$17.00 (C) 2003 IEEE

delivery mechanism. We have identified some content at-
tributes that can be used to create a content profile a priori to
the routing process. Such a profile is used to locate the con-
tent and also allows the routing protocol to infer those char-
acterisitics that directly effects the content delivery mecha-
nism (e.g., bandwidth required for delivery, client QoS re-
quirements etc.). In our characterization scheme we group
content attributes in two distinct classes. The first class
(called structural attributes) of attribute values are known
prior to the routing process and can be used to create con-
tent description, a priori, which is then used to discover
content on a network. The second class (called semantical
attributes) of attributes are initialized only at the time a re-
quest for the content is submitted and is used for accessing
the content. A combination of these will decide the network
resources that should be allocated for a request. The Table
1 and Table 2 shows the structural and semantic attributes
used in our proposed content-based routing technique.
Table 1. Structural Attributes
Physical size, version, modification date,
ownership permissions, copyrights,
author
Name-based filename, location name, URL,
origin server name
End data type HTML, cookies, scripts, audio/video
clips, text
Popularity temporal, spatial
Table 2. Supporting Attributes
Access duration, streaming, non-streaming,
variable-bit rate, constant-bit rate,
adaptive-bit rate
Quality of service bandwidth, delay,
loss tolerance
Document Type static, dynamic
Content classification is a process of grouping the docu-
ments in to different CECs such that the expected resource
requirements for handling documents within a class are sim-
ilar. The resource requirements for handling documents in-
clude the resources used for storage and processing pur-
poses and resources used for transporting purposes. Our
motivation of using content groups is to compress the the
namespace that is used by the routing system for locating
and accessing content.
0-7695-1926-1/03/$17.00 (C) 2003 IEEE
3. Simulations
We have developed a trace driven flow-based [3] sim-
ulator to compare the performance of the proposed PICS
scheme with a DNS-based content access scheme. The sim-
ulator was written in the PARSEC [15] simulation language
and uses a hierarchical 3-tier network topology generated
using the BRITE [1] and Tiers [11] topology generators.
The input topologies closely resemble the Internet infras-
tructure comprising LAN, MAN and WAN and the nodes
are divided into different autonomous systems. The sim-
ulator is driven by traces generated by ProWGen [12], a
synthetic workload generator.
The current simulation study uses a flooding-based ap-
proach to disseminate the routing tags among the CSRs and
CERs. The overhead of using flooding is determined by an-
alytical means and accounted for in the simulation studies
reported here. The experiments performed here are con-
cerned with locating and accessing content given they are
placed at geographically distributed locations. Therefore,
we use three random placement strategies to place the con-
tent on the simulated topologies. In the DNS and PICS site-
based strategies, we copy a site on a replica server. The
number of copies and placement of them are random but
controlled by a spreading parameter. In the PICS set-based
strategy, the documents are clustered into sets based on at-
tributes such as popularity, size, and type and the sets are
replicated using a randomized strategy similar to the site-
based strategies.
Fig. 6 shows the variation of resolution time with the
number of nodes on the network for 100,000 requests and
25% replication (i.e., objects were copied onto 25% of the
replica servers available on the network). The results indi-
cate that the PICS mechanism is at most 47% faster than
the DNS mechanism in resolving the requests. Disabling
the DNS caching is one of the primary reasons for the large
DNS resolution times. This resulted in 95% of the client
requests traversing the DNS hierarchy and about 38% of re-
quests contacting the root name-server. However, in order
to accurately respond to the fast varying server and network
conditions, the client-side name-servers must avoid caching
of the DNS decisions [8].
Fig. 7 shows the variation of the average content de-
livery time with the number of nodes for 100,000 requests
and 25% replication. Similarly, Fig. 8 shows the varia-
tion of the average throughput with the number of nodes for
100,000 requests and 25% replication, where the through-
put is defined as the overall rate at which the content is de-
livered from the content servers to the clients. The PICS
set-based and site-based schemes improve the DNS-based
content access scheme by about 30% and 25%, respectively.
The PICS mechanisms achieve better performance by se-
lecting the server that is capable of delivering the content to
 1400 itize the requests based on the characteristics of the content
1200
being requested.
Resolution Time / ms

1000
30000
800

Network Throughput / Mbytes per sec

25000
600

400 20000

200 15000

0
10000
50 150 250 500
Number of Nodes
5000
DNS Site-based PICS Site-based PICS Set-based
0
50 150 250 500
Number of Nodes
Figure 6. Resolution time with 25% replication DNS Site-based PICS Site-based PICS Set-based
and 100,000 requests.

Figure 8. Overall network throughput with

the client in the least amount of time. The server selection is 25% replication and 100,000 requests.
performed by not only considering the server load but also
the load on the network path. This enables the PICS mech-
anism to reduce the network congestion. The PICS enables The average network throughput for the PICS site-based
the route selection process by computing the routing frac- scheme is lower than the DNS based access scheme for
tions based on the server and network path load status. small network sizes and the PICS site-based and set-based
schemes outperform the DNS based scheme as the network
size increases. As the network size increases, the number of
3000
alternate paths connecting two given nodes increases. Be-
2500 cause the PICS mechanisms do their own request routing
Content Delivery Time / ms

2000
1500
1000
500
0
50 150
Number of Nodes
DNS Site-based PICS Site-based
Figure 7. Content delivery time with 25% repli-
cation and 100,000 requests.
Currently, the CSRs use the routing fractions to load bal-
ance among the different candidate server replicas. One of
the side effects of this process that is observed through ex-
periments is the increases link utilization. Any large-scale
system such as the PICS that uses status information has
to handle the staleness in the available information. We are
currently investigating mechanisms for interpreting the stal-
eness in the information. Further, the basic PICS framework
we present here can be extended to: (a) load balance only
resource intensive requests (large objects or streaming me-
dia) and concentrate requests for small objects and (b) prior-
they are able to exploit the multiple paths or path segments
more effectively than a DNS based access scheme. More-
over, in PICS, the availability of multiple paths are consid-
ered in conjunction with the server placements, which fur-
ther increases the performance.
Experiments were conducted to examine the impact of
250 500
streaming content. We model streaming content by inject-
ing persistent requests. These requests are routed using the
PICS Set-based
PICS mechanisms but the route is “pinned” to the initial
configuration for the life of the connection. Fig. 9 shows the
variation of the overall network throughput with the percent
of streaming requests in the request pool. It can be observed
that the throughput increases by about 15% to 20% with the
increase in streaming content. We also observed that about
11% of the streaming content requests were rejected due to
the unavailability of capacity for routing the requests.
4. Literature Review
Content routing involves both locating and accessing
content. Locating content may include content discovery
on the network. Accessing content typically requires iden-
tifying a network path with the desired quality of service
parameters and setting up sufficient resources along the
path. Existing approaches (e.g., domain name based rout-

0-7695-1926-1/03/$17.00 (C) 2003 IEEE

 350
300
Throughput / Mbytes per sec
250
200
150
100
50
0
0 1
Request for Streaming Content (%)
50 Nodes 150 Nodes
Figure 9. Variation of the overall through-
put with percentage of streaming requests
for 25% set-based replication and 100,000 re-
quests.
ing) have devised separate schemes for locating and ac-
cessing content. In this chapter, we examine a representa-
tive set of projects from the related literature on content-
based routing approaches. These approaches can be di-
vided into two classes. The first class represents highly
distributed networked naming systems. These systems use
some performance-based metrics to resolve a high-level
content-based name to a machine name that will lead to the
location of the content. The second class represents sys-
tems that “route” actual content requests considering net-
work and server status. Currently, there are only few sys-
tems that fall into this category. These include scalable
Web server clusters that are formed by interposing content
switches between a cluster of high-performance servers and
the clients.
4.1. Class I: Name-based Routing Models
The Intentional Naming System (INS) [19] presents a re-
source discovery and routing system based on the descrip-
tion of the services. Applications create and advertise in-
tentional names for each service they provide through an
overlay network, of spanning tree topology, comprising In-
tentional Name Resolvers (INRs). The INRs locally cache
the advertisements comprising the intentional names of the
services and the IP addresses of the corresponding service
providers. Any client requiring service will probe the INR
using an intentional name of the service. The INR returns
the IP addresses that correspond to the intentional name
(called early binding) or tunnels the data received from the
client to the location with the least routing metric (e.g.,
server status), called intentional anycast, depending on the
service option required by the client. The intentional names
are described using name specifiers whose main compo-
nents consist of an attribute, to classify the service, and a
value for the attribute. A hierarchical order of attribute-
value pairs are used to identify different services and an
unique identifier is used to identify the node advertising the
service.
The Name Based Routing Protocol (NBRP) [13] imple-
ments a highly scalable Internet-wide name resolver that
takes into consideration the current “performance” level of
10 25 the target servers in performing the name resolution. This
technique is similar to the current two phased data access
250 Nodes 500 Nodes paradigm in the Internet. The NBRP aggregates content
names based on how the domains are connected to the In-
ternet, i.e., an ISP content router may represent the content
originating from the servers that are being served by the
ISP. Depending on the replication patterns of a document,
it could be part of several aggregates. An Internet name
resolution protocol (INRP) is used to translate a client re-
quest to an INRP request which is then used to route the
request to the best performing content server through a net-
work of content routers. When a request reaches a content
router that is nearest to the best performing content server,
the router returns the address of the server back to the re-
questing client using the same path as the INRP request.
4.2. Class II: Content Routing Models
The Content Addressable Network (CAN) [17] presents
a highly distributed hash table architecture that can be im-
plemented on the Internet. The main idea is to create a vir-
tual coordinate space of -dimensions and hash a key on
to this space. The virtual coordinate space is divided into
subspaces and one or more CAN nodes are responsible for
maintaining the information for their associated subspaces.
Collectively, the CAN nodes form an overlay network that
spans the coordinate space. A key is hashed to find the des-
tination coordinate position in the virtual space. Using this
position, the subspace and the corresponding CAN nodes
that hold the relevant “value” are identified. Request pack-
ets are then sent towards the destination CAN nodes. A
binning scheme described in [16] can be used to create the
coordinate space. The binning technique is a scalable and
distributed method which can also be used to perform server
selection in the Internet.
PAST [7] is a persistent P2P storage utility that uses a hy-
percube based routing scheme, called Pastry, to route con-
tents. The Pastry system [2] is an overlay network of nodes
where each node is assigned a randomly generated 128-bit
identifier to denote the node’s position in a circular nodeid
space ranging from 0 to
 . Given a message with a
 
key, a Pastry node routes the message to a node whose iden-
tification number is numerically closest to the key in O(log
N) steps, where N is the number of Pastry nodes in the net-
0-7695-1926-1/03/$17.00 (C) 2003 IEEE
work. At each routing step, a node message is forwarded to
a node whose nodeid shares with the message key a given
prefix that is at least 1 digit (or b bits, where b is a configu-
ration parameter) longer than the prefix that the key shares
with the forwarding node’s nodeid.
The Service Level Routing [14] framework provides a
general architecture for accessing and discovering services
and content on the Internet. The service level routing is per-
formed by Service Level Routers (SLPs). The framework
presented in [14] is specific to two levels, although the ar-
chitecture can be extended to more than two levels. In the
SLR framework, each client request is mapped onto a flow.
When a ingress router receives a flow request, it tunnels the
request to an egress SLR that corresponds to the best per-
forming server site. The routing metric is based on the load
condition of each server as advertised by the next level of
SLRs. In effect, the ingress SLR selects the server site that
will service the client request. the egress SLR uses possibly
current information to select the physical server.
A cooperative web caching system is proposed by [10],
where several nearby caches self-configure themselves to
form multiple overlapping multicast groups and exchange
messages to locate the nearby copies of requested data and
to find out about topology changes. It uses a content rout-
ing method Content Routing Protocol (CRP) [4] which is
based on the systematic decomposition of a URL that tra-
verses the name-space tree using the URL prefixes to iden-
tify each content uniquely. Caching servers exchange their
list of URLs periodically by sending sequences of the tree
depth and hash codes of the URL prefixes.
The Caching Neighborhood Protocol (CNP) [6] uses an
origin server and a number of proxy servers which act as
caching servers, known as cache representatives (C-Rep).
The origin server is responsible for disseminating the infor-
mation to the C-Reps in its neighborhood. A C-Rep can be
a neighbor of one or more origin servers and may be either
on the client or server-side. The origin server is responsi-
ble for inviting or dismissing the neighboring proxy servers
to act as C-Reps, performing load distribution among the
C-Reps, managing information disseminating among the C-
Reps and maintaining continuous communication with the
C-Reps. The C-Reps attempt to serve a client’s request orig-
inating from their neighborhood. When a C-Rep is unable
to serve the request, it passes the request on to the next im-
mediate C-Rep or the origin server.
4.3. Discussion
Our proposed architecture is somewhat similar to INS’s
intentional anycast but uses a combination of the URL
namespace and a fixed length content-based tag for routing
a request. In contrast to NBRP, PICS combines the name
resolution and content access phases to further reduce the
0-7695-1926-1/03/$17.00 (C) 2003 IEEE
content delivery time.
The (CAN) [17] and Pastry [2] and also Tapestry [5] rep-
resent a class of routing approaches that implement peer-
to-peer routing algorithms using highly distributed rout-
ing table architectures with low communications overhead.
These highly scalable routing schemes do not incorporate
content-based metrics but these type of overlay networks
can be modified to easily include application-level metrics
for content routing. In comparison to the CRP, SLR and the
CNP our routing method allows routing at the granularity of
URLs and also provides a solution for a more comprehen-
sive form of name-based routing.
5. Conclusion and Future Work
This paper presents a novel content-based routing archi-
tecture that can be deployed as an overlay network on the
Internet. The architecture is based on a flat routing core.
Recent advances in the design and implementation of peer-
to-peer computing substrates can be leveraged to build an
efficient system based on the PICS architecture. Because
the proposed architecture uses tags that are derived from
content attributes, the architecture can be conveniently ex-
tended to includes services and resources. Thus, the pro-
posed architecture can be deployed for content, service, or
resource discovery and access.
One of the distinguishing features of the PICS architec-
ture is that it provides a content routing infrastructure that
is flexible enough to accommodate different attributes of the
requested content. Incorporation of different attributes into
the routing process is enabled by having different fields for
the attributes in the tags.
The simulation studies based on the performance model
developed for the PICS architecture indicates PICS out-
performing a DNS-based content access scheme by a sig-
nificant margin. From the simulation studies, it can be
noted that as the content replication density increases, as ex-
pected, the average content delivery time decreases. How-
ever, the percentage improvement on the content delivery
time obtained by PICS over the time obtained by the DNS-
based access scheme increases. This indicates the suitabil-
ity of PICS for situations where the content is widely repli-
cated.
Several issues need to be further addressed before a sys-
tem based on PICS can be deployed. Some of the issues
include: (a) improved tag dissemination protocols, (b) ef-
ficient content characterization and classification schemes,
and (c) mechanisms for interpreting and exploiting the stal-
eness in status information of remote servers.
References
[1] A. Medina, A. Lakhina, I. Matta, and J. Byers. BRITE:
 Universal Topology Generation from a User’s Perspective. [18] V. Pai, M. Aron, G. Banga, M. Svendsen, P. Druschel, W.
Technical Report BUCS-TR2001 -003, Boston University, Zwaenepoel, and E. Nahum. Locality-Aware Request Dis-
Jan. 2001. tribution in Cluster-based Network Servers. In 8th ACM
[2] A. Rowstron and P. Druschel. Pastry: Scalable, distributed Conference on Architectural Support for Programming Lan-
object location and routing for large-scale peer-to-peer sys- guages and Operating Systems, pages 205–216, Oct. 1998.
tems. In 18th IFIP/ACM International Conference on Dis- [19] W. Adjie-Winoto, E. Schwartz, H. Balakrishnan, and J. Lil-
tributed Systems Platform (Middleware’01), pages 329–350, ley. The Design and Implementation of an Intentional Nam-
Nov. 2001. ing System. In 17th ACM Symposium on Operating Systems
[3] A. Yan and W. Gong. Fluid Simulation for High Speed Principles (SOSP’99), pages 186–201, Dec. 1999.
Networks. In 15th International Teletraffic Congress, June [20] W. Tang, F. Du, M. W. Mutka, L. M. Ni, and A. Esfaha-
1997. nian. Supporting Global Replicated Services by a Routing-
[4] B. S. Michel, K. Nikoloudakis, P. Reiher, and L. Zhang. Metric-Aware DNS. In 2nd International Workshop on Ad-
URL Forwarding and Compression in Adaptive Web vanced Issues of E-Commerce and Web Based Information
Caching. In IEEE INFOCOM, volume 2, pages 670–678, Systems, pages 67–74, June 2000.
Mar. 2000.
[5] B. Y. Zhao, J. Kubiatowicz, and A. D. Joseph. Tapestry:
An Infrastructure for Fault-tolerant Wide-area Location and
Routing. Technical Report UCB/CSD-01-1141, University
of California, Berkeley, Apr. 2001.
[6] C. Chiang, M. Liu, and M. Muller. Caching Neighborhood
Protocol: A Foundation for Building Dynamic Caching Hi-
erarchies with WWW Proxy Servers. In International Con-
ference on Parallel Processing (ICPP’99), pages 516–523,
Sep. 1999.
[7] P. Druschel and A. Rowstron. PAST: A Large-Scale, Per-
sistent Peer-to-Peer Storage Utility. In HotOS VIII, pages
75–80, May 2001.
[8] J. Jung, E. Sit, H. Balakrishnan, and R. Morris. DNS Per-
formance and the Effectiveness of Caching. In Proceedings
of ACM SIGCOMM Internet Measurement Workshop, Aug.
2001.
[9] J.Moy. OSPF Version 2, Apr. 1998.
[10] L. Zhang and S. Floyd and V. Jacobson. Adaptive Web
Caching: Towards a New Global Caching Architecture. In
3rd International Web Caching Conference, June 1998.
[11] M. B. Doar. A Better Model for Generating Test Networks.
In IEEE Globecom, pages 86–93, Nov. 96.
[12] M. Busari and C. Williamson. On the Sensitivity of Web
Proxy Cache Performance to Workload Characteristics. In
IEEE INFOCOM, volume 3, pages 1225–1234, July 2001.
[13] M. Gritter and D. R. Cheriton. An Architecture for Content
Routing Support in the Internet. In Proceedings of the 3rd
USENIX Symposium on Internet Technologies and Systems
(USITS), pages 37–48, Mar. 2001.
[14] N. Anerousis and G. Hjalmtysson. Service Level Routing on
the Internet. In IEEE Globecom, volume 1b, pages 553–559,
Dec. 1999.
[15] R. Bagrodia, R. Meyer, M. Takai, Y. Chen, X. Zeng, J. Mar-
tin, B. Park, and H. Song. Parsec: A Parallel Simulation
Environment for Complex Systems. In IEEE Computer, vol-
ume 31(10), pages 77–85, Oct. 98.
[16] S. Ratnasamy, M. Handley, R. Karp, and S. Shenker. Topo-
logically Aware Overlay Construction and Server Selection.
In IEEE INFOCOM, volume 3, pages 1190–1199, June
2002.
[17] S. Ratnasamy, P. Francis, M. Handley, and R. Karp. A Scal-
able Content-Addressable Network. In ACM Conference on
Applications, Technologies, Architectures and Protocols for
Computer Communications, pages 161 – 172, Aug. 2001.

0-7695-1926-1/03/$17.00 (C) 2003 IEEE