The Interdomain Connectivity of PlanetLab Nodes

The Interdomain Connectivity of PlanetLab Nodes

Suman Banerjee , Timothy G. Griffin , Marcelo Pias

Abstract. In this paper we investigate the interdomain connectivity of PlanetLab

nodes. We note that about 85 percent of the hosts are located within what we call
the Global Research and Educational Network (GREN) -- an interconnected
network of high speed research networks such as Internet2 in the USA and Dante
in Europe. Since traffic with source and destination on the GREN is very likely to
be transited solely by the GREN, this means that over 70 percent of the end-to-end
measurements between PlanetLab node pairs represent measurements of GREN
characteristics. We suggest that it may be possible to systematically choose the
placement of new nodes so that as the PlanetLab platform grows it becomes a
closer and closer approximation to the Global Internet.

1 PlanetLab

The primary goal of PlanetLab is to provide a geographically distributed platform for

overlay-based services and applications [12, 1]. Currently there are over 120 participat-

ing sites with a total of over 300 hosted nodes. The open and shared nature of PlanetLab

is enabling an exciting range of innovative experimentation in overlay techniques. With-

out its collectively supported infrastructure most participating research groups would

not have the means to set up such a rich environment.

For many of the same reasons, the PlanetLab infrastructure has attracted researchers

working with measurements of the legacy Internet. That is, PlanetLab nodes are em-

ployed to actively probe and collect various Internet metrics. We argue that this is an

opportunistic use of PlanetLab in the sense that providing an Internet measurement

infrastructure has never been among the primary goals of the project. This does not

mean that PlanetLab is not a useful Internet measurement platform, only that PlanetLab

measurements cannot automatically be taken as representative of the global Internet.

In this paper we investigate the interdomain connectivity of PlanetLab nodes. We

note that about 85 percent of the hosts are located within what we call the Global Re-

search and Educational Network (GREN) -- an interconnected network of high speed

research networks such as Internet2 in the USA and Dante in Europe. Since traffic

with source and destination on the GREN is very likely to be transited solely by the

GREN, this means that over 70 percent of the end-to-end measurements between Plan-

etLab node pairs represent measurements of GREN characteristics. Whether or not such

measurements are representative of the global Internet is something that needs more in-

vestigation.

This should in no way be misconstrued as a criticism of PlanetLab -- we are only

stating that those using PlanetLab for measurements of the global legacy Internet need

University of Wisconsin, Madison. suman@cs.wisc.edu
Intel Research, Cambridge UK. tim.griffin@intel.com
Intel Research, Cambridge UK. marcelo.pias@intel.com

to present their arguments with care. On the other hand, the GREN is in some respects

more attractive to measurement researchers than the Internet at large. Primarily this is

because it is more transparent -- that is, there is more publicly available information Matix sound of progress headphones provide digital sound

about the connectivity of the GREN and more willingness on the part of its operators

to share information with the research community. We suggest that it may be possible

to systematically choose the placement of new nodes so that as the PlanetLab platform

grows it becomes a closer and closer approximation to the Global Internet. We advo-

cate that there is still a large amount of untapped diversity within GREN which can be

explored for similar effect on the PlanetLab.

We present some preliminary results on a case study. We generate a site-to-site dis-

tance matrix for PlanetLab sites where distances between sites is taken to be minimum

round trip time. We then enumerate all possible triangles formed by three sites and in-

vestigate the violations of the triangle inequality. Low violations are important for the

feasibility of various proposals to generate synthetic coordinate systems for the Internet

based on round trip time measurements [11, 13, 17]. We find that when we classify trian-

gles as "research triangles" (all nodes on the GREN), "commercial triangles" (all nodes

off of the GREN), and "mixed triangles" (combination of commercial and GREN sites).

We find that the distribution of "bad triangles" is lowest for research triangles (about

12 percent), higher for mixed triangles (about 20 percent), and highest for commercial

triangles (about 25 percent).

2 The Global Research and Education Network (GREN)

The global Internet is comprised of a large collection of autonomously administered

networks. Some of these networks are operated by commercial enterprises, while oth-

ers are operated by nonprofit organizations. Perhaps the largest nonprofit networking

organizations exist to provide connectivity between research and academic institutions.

This includes the Geant backbone network run by the Dante organization in Europe

and the Abilene backbone network run by Internet2 in North America. Such backbones

connect many regional and national research networks into a large global network pro-

viding connectivity between diverse academic and research organizations. We refer to

this network as the Global Research and Education Network solar film
(GREN).

Figure 1 presents a simplified picture of the current GREN. The GREN is not by

any means a single administrative entity. All of the networks are independently admin-

istered, and exhibit various degrees of cooperation. There is also large diversity in the

primary goals of these networks. Some regional networks are targeted toward a specific

set of users, while others serve a larger research and education community. For exam-

ple, the CERN network exists largely to provide high bandwidth connectivity between

physics laboratories around the world. On the other hand, the WiscNet of Wiscon-

sin (http://www.wiscnet.net) exists to provide connectivity between K-12 educational

and university level institutions. Some GREN networks provide transit to commercial

network providers, while others do not. For example, the backbone of Internet2, the

Abilene network, does not provide commercial connectivity, while WiscNet does. The

ARENA project (http://arena.internet2.edu) provides a very useful online compendium

of information about the networks making up the GREN. It should also be rememberedMy favorite San Diego sedation dentist - check out the site!

ABILENE

GEANT

AARNET Austrailia

RBNnet Russia

campus

backbone

regional

CANet3 Canada

Exchange

AMPATH

telefon dinleme

RPN2 Brazil

Latin America and Caribbean

APAN Japan

TANet2 Taiwan

CERNet China

KEY

planned

Fig. 1. A highly simplified and incomplete schematic of the GREN.

that the GREN is continually changing and expanding. For example, to Latin America

from Europe, GREN traffic is routed today through the USA (Florida). A directed link

between Geant and the Latin American research networks is planned within the scope

of the ALICE project [15]. This will create a "short-cut" between these networks, which

is likely to have an effect on how experimental results should be interpreted. Similarly,

other new links are planned between Europe and USA.

The logical connectivity of the GREN networks does not immediately tell us how

traffic between them is routed. Figure 2 depicts two sites A and B that have connec-

tivity to both the GREN and the commercial Internet. One would normally expect that

both sites A and B prefer routes from the research network over routes learned from

their commercial providers. There are several reasons for this, the primary one being

that research connectivity is normally paid for out of a special source of funds, and it

may not be metered as it is normally done with commercial traffic. In addition, there

may be the perception that for research work, the research network will give better

performance. In this way, we can think of the research network as providing a "short

cut" between A and B that bypasses the commercial Internet. Of course, this may be

occasionally overridden by local routing policies. The actual connectivity of most sites

is more complex than Figure 2. For example, Figure 3 shows the connectivity of two

PlanetLab sites: planetlab2.cs.unibo.it and planetlab2.cs.wisc.edu. We have verified our

assumptions about routing policies with "AS level traceroutes". For the above example

this yields:

Commercial Internet

Research Network

Fig. 2. Simple schematic of the routing

choices as A and B

Commercial Internet

MAD-U

WISC-NET

ABILENE

GEANT

GARR

planetlab2.cs.wisc.edu

planetlab2.cs.unibo.it

GREN

Fig. 3. A slightly more complicated ex-

ample.

AS 137

: GARR

--- Italian REN

(5 routers)

AS 20965 : GEANT

--- Dante Backbone

(3 routers)

AS 11537 : ABILENE --- Internet2 backbone

(2 routers)

AS 2381

: WISCNET --- Wisconsin REN

(3 routers)

AS 59

: MAD U

--- U. of Wisconsin, Madison

(5 routes)

We compute these by first performing router-level traceroutes between PlanetLab nodes

using Scriptroute [2] and then post-processing the results to associate intermediate

routers with ASNs. With only a few exceptions, our expectations about routing were

confirmed.

3 PlanetLab and the GREN

Since those engaged in the building of PlanetLab and the deployment of overlay ser-

vices are largely academically oriented researchers, it should not be surprising that the

majority of PlanetLab sites are located in the GREN. However, as far as we know there

has never been an attempt to systematically quantify this relationship.

We must determine whether a given PlanetLab host or site is connected to the GREN

or the commercial Internet. We use the following approach. First, we obtain a reason-

able list of ASes in the GREN. This was done by extracting from Route-Views [3] BGP

routes only those routes that were announced by Abilene (this can be found in the "oix

full snapshot" source). This BGP session with Route-Views announces only research

destinations (about 8,000 routes). In a perfect world, the ASNs in the AS-PATHS of

these routes would be exactly the ASNs of the GREN. However, a small number of

routes from the commercial Internet get "leaked" into these tables (for reasons still un-

clear), and so we must eliminate some ASes that are known to be commercial providers

(Sprint, MCI, and so on). On January 15, 2004, this procedure left us with a list of 1,123

ASNs for the GREN.

Next, for each IP address associated with a PlanetLab node, we find the originat-

ing ASN associated with the longest matching prefix. In order to generate a reasonable

mapping of originating ASNs to prefixes, we merged routing information from 20 BGP

tables taken from RIPE [10] and Route-Views. This is essentially the basis of the tech-

nique described in [9, 8] for obtaining AS-level traceroutes. If the originating AS is in

our GREN list, we classify the node as a GREN node. Otherwise it is a commercial

node. If the address is associated with Multiple Originating ASNS, some being com-

mercial and others research, we classify the site as MOAS. With the 276 production

nodes of January 15, 2004, we obtain the following breakdown as shown in Table 1.

class number percent

MOAS

5.1

Commercial

9.8

Research

235

85.1

Table 1. Breakdown of hosts.

class number percent

MOAS

5.7

Commercial

9.8

Research

104

84.5

Table 2. Breakdown of sites.

We then group the hosts into sites containing multiple hosts. This produces 123 sites

with a breakdown very close to that of the hosts (Table 2).

Note that this means that over 70 percent of the end-to-end measurements between

PlanetLab node pairs represent measurements of GREN characteristics. We then clas-

sified these sites into one of four broad geographical regions: North America (NA),

Europe the Middle East and Africa (EMEA), Asia Pacific (AP), and Latin America

(LA). The breakdown of the sites are listed in Table 3.

region number percent

0.8

4.9

EMEA

20.3

74.0

Table 3. Breakdown by region.

class

region number percent

Commercial

EMEA

0.8

Research

0.8

Research

4.9

Commercial

8.9

Research

EMEA

18.7

Research

60.2

Research NA or EMEA

78.9

Table 4. Combination of regions and types.

We now look at the combinations of these two classifications (Table 4). In this ta-

ble we have ignored the MOAS sites (although they are still counted when computing

percentages) Note that this means that over 60 percent of the end-to-end measurements

between PlanetLab node pairs represent measurements of the NA and EMEA portion

of the GREN. In summary, we note that not only are PlanetLab nodes situated in the

GREN corner of the Internet, but inhabits a fairly small part of that world as well. The

majority of site-to-site data flows are carried over the Abilene and Geant networks.

Contrast this with the rich diversity illustrated in Figure 1.

3.1 Case Study -- The Triangle Inequality

A number of applications would benefit from a global topological map of the Internet,

which incorporates information such as the number of hops and the network latency

between Internet hosts. Towards this goal, research in recent years have focused on dis-

tributed techniques to construct an Internet Coordinate System, in which each host can

be assigned a virtual coordinate and the distance between the virtual coordinates of any

two hosts will approximate their actual distance for some specific network performance

metric. Network latency is an example of such a metric. There may be one such coor-

dinate system for each metric of interest. Some of the proposed techniques are Global

Network Positioning (GNP) [11], Lighthouses [13], Virtual Landmarks [17], ICS [7]

and Practical Internet Coordinates (PIC) [4]. The goal in these techniques is to signif-

icantly reduce the number of measurements required, thereby scaling to large number

of hosts. There are many applications that can leverage such an Internet Coordinate

System, e.g. topologically-aware construction of peer-to-peer systems [5], network la-

tency estimation tools [16], and distributed resource discovery mechanisms [14]. The

general problem to construct such a coordinate system can be abstracted as the problem

of mapping, or embedding, a graph into a metric space. A metric space M is defined by

the pair (X , d) where X represents the set of valid objects and d is a metric. A metric

is a function d : X × X

such that for x

, x

X, d satisfies the following

properties:

1. d(x

, x

)

0 (positiveness),

2. d(x

, x

) = d (x

, x

)

(symmetry),

3. d(x

, x

)

d(x

, x

) + d (x

, x

)

(triangle inequality)

In contrast, a vector space V is a particular metric space in which X =

and

it has distance function D that satisfies properties (1), (2) and (3). The embedding

problem consists of finding a scalable mapping : X

that transforms ob-

jects {x

, ..., x

}, of the original space, in this case the Internet graph, onto points

, ..., v

} in a target vector space V of dimensionality k.

The concept of distortion is used to measure the quality of an embedding technique.

It measures how much larger or smaller the distances in the vector space D(v

, v

)

are when compared to the corresponding distances d(x

, x

)

of the original space. The

distortion is the lowest value c

that guarantees that [6]:

. d (x

, x

)

D(v

, v

)

. d (x

, x

)

Different techniques have been recently proposed to embed the Internet graph into

a metric space, while trying to minimize the distortion at the same time [11, 13, 17, 7, 4,

5]. The distance metric used in these techniques is the minimum round trip time (RTT)

over a significantly large time interval.

3.2 Triangle Inequality on the Internet

The AS-level traffic on the Internet is forwarded based on dynamic BGP routing poli-

cies. In general, service providers are free to set their own BGP policies and make local

arrangements with peering providers and customers. Moreover, service providers are

often multi-homed, which means they have multiple connections to the Internet for var-

ious reasons such as links resilience by using backup links and traffic load balancing.

Also, the customer-provider relationship can be regarded as one reason why shortest

path routing may not be used in a given fraction of the network. There will be cases

where providers, for business-related reasons, will prefer to route traffic via another

provider as opposed to a customer, even if the shortest path is through its customer.

Thus, there is no reason to expect that the triangle inequality, an important property

of metric spaces, holds on the Internet. In this section we enumerate all possible tri-

angles formed by three PlanetLab sites. We then analyze the violations of the triangle

inequality, as well as a measure of how good these triangles are.

3.3 Data and Methodology

We used RTT data

collected between all PlanetLab hosts from November 17th to

November 23rd 2003. We then calculated the minimum RTT between each pair of hosts

available between those dates on consecutive 15-minute periods. Thus for each day in

this period there were 96 matrices of RTT measurements (each entry represented a pair-

wise RTT measurement), and the size of each matrix was 261 × 261. Over the seven

day period we therefore had 672 such matrices.

Our goal in this evaluation was to identify the various triangle inequality violations

between these pairs of hosts. Since we wanted to perform this computation over our

estimate of propagation delays on the paths only (and not the queueing and congestion

effects), we first constructed a single matrix, in which each entry represented the mini-

mum RTT between a pair of PlanetLab nodes over the entire 7-day period. This avoided

biases in the results due to high variations in RTTs, e.g. during congested periods. Our

analysis of the data indicated that by taking the minimum over a 7-day period, we can

filter out congestion related effects which have periodic weekly patterns.

Many PlanetLab sites had multiple nodes per site. For instance, the Computer Labo-

ratory (University of Cambridge) site hosted three nodes planetlab1,planetlab2
and planetlab3

under the domain cl.cam.ac.uk. The minimum RTTs be-

tween nodes within a site were very small, often of the order of 1 ms.

Examining triangles between all triples of nodes was therefore wasteful and biased

our results by the distribution of nodes in PlanetLab sites. Thus, we selected a repre-

sentative node in each site so as to build a site by site matrix D , reducing the distance

matrix to 123 × 123 RTTs.

3.4 Preliminary Results

We defined a metric r, termed the goodness of a triangle, to test violations of the triangle

inequality on the D matrix. By convention a is the longest side of a triangle, and the

other remaining sides are termed b, c. The metric r is defined as:

r =

b + c

× (1 + (a - (b + c)))

This metric is made of two terms. The first one distinguishes between `good' and

`bad' triangles (violators). Values of r less or equal to 1 represent `good' triangles. In

contrast, any triangle with r greater than 1 is a violator, i.e. the relation between its

sides lengths do not satisfy the triangle inequality property. This first term is multiplied

The all-pairs RTT measurement data was being continuously collected by Jeremy Stribling
(MIT) and it is available from http://www.pdos.lcs.mit.edu/~strib/pl_app/

by a factor which tells the `goodness' of a triangle. The higher the value of r, the

worse it is the triangle. We believe that triangles with higher r have higher impact on

the embedding distortion factors c

and c

. However, further research is required to

quantify such a distortion.

-50

-40

-30

-20

-10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

r = (a/(b+c)) * (1+(a-(b+c)))

CDF

Commercial nodes (C.C.C)

Research nodes (R.R.R)

Mixed

(a) Distribution of all triangles

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

r = (a/(b+c)) * (1+(a-(b+c)))

CDF

Commercial nodes (C.C.C)

Research nodes (R.R.R)

Mixed

(b) Distribution of violators

Fig. 4. CDF of the goodness metric r

In Figure 4, we plot the CDFs of metric (r) of triangles derived from the upper

triangular part of D . The three lines in each graph are: (1) RRR, which includes tri-

angles formed only by sites in GREN, (2) CCC, which includes only those triangles

that are formed by sites in the commercial Internet, and (3) Mixed, which includes all

the other triangles (combination of commercial and GREN sites). Figure 3.4 shows the

distributions of `good' and `bad' triangles; whereas Figure 3.4 presents only the cases

of violators, i.e. values of r greater than 1.

We can make two interesting observations from this figure. First, there are fewer

violations of the triangle inequality in the GREN than in the commercial Internet. Vio-

lators represent about 12% of the cases in the GREN and about 25% for the commercial

Internet. When mixed triangles were tested, around 20% of them were violators.

And second, the triangles in the GREN are `better' formed than the triangles in the

commercial Internet. This means that `bad' triangles in the commercial Internet tended

to be larger than the ones of the GREN and also have a very long side.

At this stage, we speculate that this difference might be the reason why the distribu-

tion of violators for mixed triangles is closer to the distribution of commercial triangles

(Figure 3.4). One would expect the opposite, as the number of GREN sites forming

mixed triangles are predominantly larger.

It seems that commercial sites when combined to GREN sites influenced the shape

of the triangles, i.e. distorting them. The nature and cause of these differences need

further detailed investigation.

4 Recommendations for improving PlanetLab diversity

The results presented in our case study leads to broad research questions -- How can

we choose a set of nodes in the Internet so that inter-node experiments can reasonably

be taken to represent behavior on the Internet as a whole? In the PlanetLab context this

might be reformulated as follows -- How can we systematically choose the placement

of new nodes so that as the PlanetLab platform grows it becomes a closer and closer

approximation to the Global Internet?

One aspect of this is the distinction between commercial and research networks.

One direction towards diversity might be to increase the number of sites on the com-

mercial networks. This may be desirable, but as a general strategy it has several short-

comings. The main problem with this approach is that many research sites are buried

deep inside of department/campus/regional networks and find it difficult if not impos-

sible to get direct access to the commercial Internet. For example, the PlanetLab hosts

at the computer science department in UW Madison are many router hops away from a

commercial router. It obtains connectivity through a departmental network, then a cam-

pus network, and finally through a state-wide research network (WiscNet). It is only

WiscNet that provides access to the commercial Internet. A solution to add diversity

in PlanetLab paths might be to work with the upstream networks and have some Plan-

etLab net blocks announced only to the commercial networks upstream from Wiscnet.

Perhaps this could be implemented with some type of BGP community sent up on the

routes when they are announced into WiscNet. It probably can be done, but requires a

lot of BGP magic.

We would advocate a different approach. The entire GREN is very large and com-

prises of a diverse set of networks. For example, CERN runs a network that is to GREN

what the GREN is to the global Internet. Out of the 1123 ASes that we found in the

GREN, PlanetLab nodes are located in 80, i.e. less than 8% of them. Perhaps the best

strategy (in terms of cost and feasibility) is to make a targeted and systematic attempt

to exploit the existing diversity of the GREN. Our exploration and research of the char-

acteristics of GREN lead us to believe that there are a lot of diverse networks within

GREN itself. For example, While some research divisions in academic environments

may have relatively low bandwidth connectivity to the Internet, other departments and

entities, e.g. some astronomy and high-energy physics communities, have very high

bandwidth connectivity and we should target this diversity of groups and communities

outside of Computer Science to add on as PlanetLab sites. We should also aggregisvely

consider adding further geographic diversity by recruiting more sites in Latin America,

Asia, Australia, Russia and so on. The connectivity characteristics of such diverse lo-

cations would certainly enhance the diversity on the PlanetLab. One possible approach

to do this might be to examine various routing and topological data that are available

from sites like Caida (http://www.caida.org), e.g. Skitter data, and identify specific sites

within GREN that will expand PlanetLab diversity.

Focusing on the GREN for diversity has additional advantages. In general, the

GREN is more transparent -- there is more publicly available information about the

connectivity of the GREN and more willingness to share information with the research

community. Hence if we can find suitable diversity just within GREN, such an enhance-

ment is certainly worth exploring further as we attempt to establish greater diversity

within PlanetLab.

4.1 Acknowledgements
We would like to thank the following people whose comments on this work were ex-

tremely helpful -- Randy Bush, Jon Crowcroft, Christophe Diot, Steven Hand, Gi-

anluca Iannaccone, Z. Morley Mao, Andrew Moore, David Moore, Jennifer Rexford,

Timothy Roscoe, Larry Peterson, James Scott, Colleen Shannon, and Richard Sharp.

Thanks to Jose Augusto Suruagy Monteiro of UNIFACS, Sidney Lucena and Ari Frazao

Jr. of RNP Brazil for helping us better understand the connectivity of the Brazilian re-

search networks. We would like to thank Jeremey Stribling for collecting the PlanetLab

round trip data and making it publicly available. In addition, this work would not be

possible without the existence of public archives of interdomain routing data including

Route Views and RIPE.

References

1. PlanetLab home page. http://www.planetlab.org.

2. Scriptroute Home Page. http://www.scriptroute.org.

3. University of Oregon Route Views Archive Project. http://www.routeviews.org.

4. M. Costa, M. Castro, A. Rowstron, and P. Key. PIC: Practical Internet Coordinates for Dis-

tance Estimation. In 24th IEEE International Conference on Distributed Computing Systems
(ICDCS' 04), Tokyo, Japan, March 2004.

5. R. Cox, F. Dabek, F. Kaashoek, J. Li, and R. Morris. Practical, Distributed Network Coordi-

nates. In ACM Workshop on Hot Topics in Networks (HotNets-II), November 2003.

6. G. Hjaltason and H. Samet. Properties of Embedding Methods for Similarity Searching in

Metric Spaces. IEEE Trans. on Pattern Analysis and Machine Intelligence, 25(5), May 2003.

7. H. Lim, J. Hou, and C. Choi. Constructing Internet Coordinate System Based on Delay Mea-

surement. In ACM SIGCOMM Internet Measurement Conference (IMC'03), Miami (FL),
USA, October 2003.

8. Z. Morley Mao, David Johnson, Jennefer Rexford, Jia Wang, and Rany Katz. Scalable and

Accurate Identification of AS-Level Forwarding Paths. In INFOCOM '04, 2004.

9. Z. Morley Mao, Jennefer Rexford, Jia Wang, and Rany Katz. Towards an Accurate AS-level

Traceroute Tool. In ACM SIGCOMM Technical Conference '03, August 2003.

10. Ripe NCC. Routing Information Service Raw Data. http://abcoude.ripe.net/ris/rawdata.

11. T.S. Eugene Ng and H. Zhang. Predicting Internet Network Distance with Coordinates-Based

Approaches. In IEEE INFOCOM' 02, New York, USA, June 2002.

12. L. Peterson, T. Anderson, and D. Culler. A Blueprint for Introducing Disruptive Technology

into the Internet. In ACM Workshop on Hot Topics in Networks (HotNets-I), October 2002.

13. M. Pias, J. Crowcroft, S. Wilbur, T. Harris, and S. Bhatti. Lighthouses for Scalable Dis-

tributed Location. In 2nd International Workshop on Peer-to-Peer Systems, February 2003.

14. D. Spence.

An implementation of a Coordinate based Location System.

Technical

Report UCAM-CL-TR-576, University of Cambridge, November 2003.

Available at

http://www.cl.cam.ac.uk/TechReports/UCAM-CL-TR-576.pdf.

15. C. Stover and M. Stanton.

Integrating Latin American and European Research

and Education Networks through the ALICE project.

Document available at

http://www.dante.net/upload/doc/AlicePaper.doc.

16. M. Szymaniak, G. Pierre, and M. van Steen. Scalable Cooperative Latency Estimation. Sub-

mitted for publication. Draft available at http://www.globule.org/publi/SCOLE draft.html,
December 2003.

17. L. Tang and M. Crovella. Virtual Landmarks for the Internet. In ACM SIGCOMM Internet

Measurement Conference (IMC'03), Miami (FL), USA, October 2003.