Peer to peer networking:

Main aspects and conclusions from the view of Internet service providers
Gerhard Haßlinger, Department of Computer Science, Darmstadt University of Technology, Germany

Abstract: - Peer to peer applications presently con- the network or they work completely distributed
tribute the main part to the traffic volume on broad- without central servers as e.g. the Gnutella network.
band access platforms to the Internet. The underly- The download phase following a successful search
ing distributed file sharing systems and data ex- is always organized between peers.
change protocols supports new service types to be Initial versions of peer to peer protocols are often
integrated in IP networks. subjected to a large messaging overhead, which
This paper investigates the traffic profiles and im- amounts to e.g. 6 kbit/s per connection according to
plications of peer to peer applications for resource measurement in the Gnutella network [9]. Protocols
efficiency, monitoring, control and quality of ser- for large peer to peer networks with several million
vice on the network layer of service provider plat- active nodes obviously have reduced the overhead
forms. and improved scalability. Nevertheless, a significant
messaging overhead often remains.
Key-Words: - Peer to peer networking, effect on
The data search and selection of sources for down-
traffic profiles and quality of service, resource effi-
loading establishes routing mechanisms on the ap-
ciency, monitoring, routing
plication layer, which are usually independent of the
routing on the IP network layer. The distribution of
1 Introduction data over the peer to peer network and the match of
routing strategies on different layers is decisive for
In addition to the client server architecture, peer to an efficient bandwidth utilization in IP platforms.
peer networks have established an alternative con- This is a precondition for a good data throughput
cept for the distribution of information and data, and thus for the success of the peer to peer protocol
which has become popular within a few month after and in addition has to be optimized by the service
Napster made a data exchange protocol available on providers. On the other hand, the diversity and vari-
the Internet for music files. Meanwhile a variety of ability of peer to peer protocols makes it difficult to
peer to peer protocols contribute the major portion monitor or control their traffic [8].
of the Internet traffic by exchange of video and other In section 2 we start with a closer look at peer to
data and are a main driver of a recent steep increase peer traffic profiles, followed by implications of the
in traffic volumes over broadband access platforms. up- and download symmetry and the application
Figure 1 shows a representative measurement of layer routing in sections 3 and 4. Monitoring and
components in the present Internet traffic mix, possible caching of peer to peer traffic is discussed
which depends on the portion of broadband access in sections 5 and finally the consequences on the
users. Those figures are also varying in a daily traf- quality of service and the prediction of future traffic
fic profile, where e.g. web browsing (HTTP) is development are addressed in sections 6 and 7.
more relevant in heavy traffic hours than throughout
the night time [7][8][10][11]. Besides file sharing,
the peer to peer principle supports distributed com- 2 Peer to peer traffic characteristics
putation with a balanced utilization of vacant re- Peer to peer applications generate a large traffic
sources including bandwidth in access networks. volume with downloads of video files of several
Communication in peer to peer networks restrict Gigabyte size as the main driver of long lasting
server tasks to the support of the search for data in background file transfers.

50 - 75% 10 - 20% 15 - 25% < 5%

Peer-to-Peer networking & data exchange HTTP other TCP UDP
Figure 1: Typical composition of the main components of present Internet traffic
Despite the increasing volume, peer to peer applica- 3 Relation of uploads and downloads
tions have an overall smoothing effect on traffic
profiles as compared to client server architectures: Peer to peer traffic is symmetrical in up- and down-
stream direction when comparable nodes are in-
The daily traffic profiles in broadband access volved as source and destination of a transfer. Since
platforms typically show busy hours of high ac- most users are interested in downloading files at
tivity in the day or in the evening time. For peer first, the peer to peer protocols have to take care of a
to peer traffic, the ratio of the peak to the mean balanced transfer in both directions. The protocols
rate is smaller than 1.5 owing to background with widespread usage start to upload data from a
transfers which often last during the night time, node immediately after it has been downloaded as
whereas web browsing and many other applica- soon as data chunks are available on the application
tions have ratios larger than 2 from the peak to layer. Moreover, they enforce a give and take policy
the mean. Thus the long lasting downloads have for each participant such that users with a higher
a smoothing effect on the periodical changes of upload volume are preferred when they are request-
data transfer activity. ing downloads.
There are less hot spots in peer to peer applica- On the other hand, broadband access lines are usu-
tions than in client server based architectures. ally asymmetrical with an essentially smaller up-
For the search phase, this depends on the level stream rate e.g. 128kbit/s ÷ 768 kbit/s for ADSL
of distributed versus server based computing of access as offered by Deutsche Telekom. Therefore
the peer to peer protocol, whereas the download the upstream capacity in broadband access platforms
phases run completely distributed among the becomes a bottleneck for peer to peer applications.
nodes. While spontaneous hot spot accesses In measurement of upstream traffic the peer to peer
from many clients to a server can lead to bottle- portion is even more prevalent with up to 90% than
necks, the hot spot data is soon replicated and on backbone links as in figure 1 or downstream.
afterwards downloaded from many nodes in a
peer to peer network. Symmetrical access lines would be more appropriate
for peer to peer data exchange, but if service provid-
Most of the peer to peer source traffic ers would replace ADSL lines with symmetrical
originates from subscribers with a limited DSL at comparable speed, i.e. 448kbit/s ÷ 448 kbit/s
access rate especially for uploads. The up- instead of 128kbit/s ÷ 768 kbit/s, then peer to peer
load speed is below 128kbit/s even in most traffic can be expected to increase with the upstream
broadband access lines. Then the burstiness speed. This shows that enforcing a close relation
of traffic on aggregation links is reduced in between up- and download volume of each user is
short and medium time scales, since statis- essential in order to achieve a high throughput of
tical multiplexing has a smoothing effect on peer to peer traffic on broadband IP platforms.
the variability of the traffic rate.
When the user population and thus the number
4 Routing on application layer
of sessions in parallel is constant and transfer
activities in the sessions are independent of Peer to peer networks establish their own routing on
each other, then the central limit theorem is the application layer. Different routing principles
applicable to the rate statistics of aggregated are in use with consequences for the efficiency and
traffic being composed of a large number of scalability. Earlier versions of the Gnutella network
simply forwarded broadcast search requests to all
small and independent flows. Remote access
neighbours within a limited distance measured by
routers often aggregate the traffic of several
the hop count, i.e. the number of intermediate nodes.
thousand users at the boundary of IP platforms. In this way a large amount of messages are ex-
As a consequence, the coefficient of variation, changed for each search, which imposed restrictions
i.e. the ratio of the variance to the square of the on the Gnutella network scalability [9].
mean, becomes smaller with the level of aggre- Presently, peer to peer protocols often introduce
gation and the traffic rate approaches a Gaus- hierarchically structured search nodes to reduce the
sian distribution. Simple link dimensioning overhead together with functions known from IP
rules are available for Gaussian traffic with routing like hello messages to confirm connectivity
regard to quality of service aspects [4][5]. to the peers. Nevertheless, a significant messaging
overhead can be observed in large peer to peer net- Depending on the protocol, patterns can be recog-
works. The implementation of two independent nized, which indicate data transmitted in peer to
routing layers causes inefficiency at the extend of peer search phases or the initiation of data ex-
the mismatch between different underlying network changes. Afterwards they can be traced at the trans-
structures. port layer until the termination of a corresponding
TCP connection.
5 Monitoring and analysis of peer to The analysis on this layer is complex and cannot be
peer applications performed permanently for the complete traffic on
high speed links. Recognition pattern have to be
Monitoring and analysis of peer to peer traffic found for each relevant peer to peer protocol and, in
is essential for Internet service providers in or- addition, the analysis has to be updated whenever
der new protocols or new variants of existing protocols
emerge. Although there are approaches for the
to determine the components of the traffic analysis of several currently used protocols [8], a
mix, which may indicate shifts and trends in complete classification of peer to peer traffic re-
the usage of applications, mains difficult since some protocols have developed
techniques to disguise their purpose also on the ap-
to estimate the overhead in peer to peer plication layer, e.g Freenet [2].
messaging, Besides the analysis of the components of complete
to analyze the peer to peer network structure data flows, another approach of application layer
and transmission paths to be compared to analysis has been carried out by crawling into peer
to peer protocols [10]. Therefore a node is inserted
the routing on the network layer.
into a peer to peer network, which can collect data
about the connectivity and structure of the network
5.1 Monitoring on network & transport layer as well as the status of other nodes.

The components of traffic can be analyzed on the

network and transport layer by evaluating informa- 5.3 Caches
tion in the header of the IP and TCP packets, i.e. IP Web caches provide an opportunity to optimize
addresses, TCP ports and information about the TCP traffic flows. Usual web caches do not apply to peer
signaling and connection status. Peer to peer proto- to peer traffic and have become inefficient. On the
cols are often associated with a standard port or a other hand, caches can be set up especially for peer
range of specific ports, which allow to recognize to peer traffic. Therefore they act as a node within a
them on the transport layer. When the search phase peer to peer network, which stores a large amount of
relies on a small set of servers then the correspond- data.
ing requests can be traced by an analysis of IP ad-
dresses. The analysis on this layer is supported by Web caches are not intended to play an active role
sampling mechanisms provided by the router equip- in peer to peer networks. They should represent a
ment and can be performed at line speed even on transparent proxy to shorten transmission paths to
backbone links. distinct nodes, but should not be directly addressed
as a peer to peer network nodes. When a download
However, most widely used peer to peer protocols is requested for some data chunk available in the
can be configured to use any arbitrary TCP port and cache, then the cache can respond instead of another
thus an essential part of peer to peer traffic cannot source, which has been selected in the search phase.
be categorized on this layer, which runs over non- For transparency, the data should be transferred as if
standard TCP ports or even standard ports of other it would originate from the selected source with
applications. In recent time, the portion of unknown regard to
TCP ports is increasing on transport layer statistics,
as can be observed e.g. in the Internet2 NetFlow the IP addresses,
statistics [6]. the upstream access bandwidth of the source,
possible changes in the status of the source, e.g.
5.2 Application layer analysis accounting to balance up- and download volume
To get better insight into peer to peer traffic compo- of peer to peer network nodes.
nents, application level analysis is necessary which
A data transfer from the cache usually will have
also inspects the payload of transmitted packets.
shorter transmission delay and will not be able to
match the available upstream rate of the original forwarding prioritiy. Even applications like web
source, including time varying bottlenecks on the browsing and email can be included in a 20% por-
transmission path beyond the cache. But at least the tion of traffic with most serious requirements.
online availability and the access speed of the origi- When the network is dimensioned for an essentially
nal source should be taken into account. In fact, the larger traffic volume than generated only by the
upload capacity of the caches substitute a part of the preferred traffic classes, then they will not suffer
upload capacity of nodes in the peer to peer network from bottlenecks and queueing delay in normal op-
with consequences for the total data throughput. In eration whereas, vice versa, the impact of a small
this way, caches for peer to peer traffic have to be premium traffic portion on the peer to peer traffic is
adapted for each protocol in use and do not reduce moderate.
messaging overhead in the search phase.
In addition, the delivery for premium traffic classes
The efficiency of caches depends on the source se- can even be assured in some failure situations, e.g.
lection by the peer to peer protocol. In principle, for single link breakdowns, provided that restoration
unnecessary load on backbone and expensive inter- links are available in an appropriate network design.
national links can be avoided by the application Since overload may occur on those links, the best
protocol, if one of the nearest sources in terms of the effort traffic will then often be affected.
IP routing distance is preferred. But a proper distri-
Nowadays peer to peer protocols can cope with
bution of the data with regard to the underlying IP
temporal disconnections on the application layer and
network is only a secondary focus of peer to peer
recover transmission from the previous state after-
protocols.
wards. When file transfers via FTP or HTTP proto-
cols are interrupted, an essential part of the trans-
mission is often lost and a complete restart is re-
7 Implications for quality of service quired. Since a segmentation and reassembly of
The Internet has developed from data transfer appli- large data files into small chunks is included on the
cations to a service integrating platform with stead- application layer of peer to peer protocols, an im-
ily increasing variety of service types including file proved reliability and efficiency of transfers is
transfer, email, web browsing, voice over IP, peer to achieved, which is valuable especially for non-
peer data exchange etc. Each service type has its assured QoS of best effort transmission.
specific quality of service (QoS) demands regarding As pointed out in section 5, it is difficult to classify
the required bandwidth, the transmission time e.g. peer to peer traffic and virtually impossible to
real time constraints, as well as tolerance for trans- clearly distinguish it from other traffic on the net-
mission errors and failure situations. Peer to peer work and transport layer. A treatment with lower
data exchange is usually of the best effort service priority based on TCP port numbers will increase
type without strict QoS demands. Often a download the tendency to disguise peer to peer applications by
may run for several hours or days in the back- using randomly chosen ports for unknown protocols
ground. or by transporting peer to peer data exchange e.g.
Although shorter transfers or even real time trans- over the HTTP port for web browsing. Therefore the
missions would be desirable, users are aware that only efficient way to classify traffic seems to be a
affordable tariffs in a mass market impose access declaration and marking of premium traffic with
bandwidth limitations such that video files require strict quality of service requirements by the users
considerable transfer time even with broadband himself or by an originating server, which will be
access. accepted on service provider platforms in combina-
tion with a corresponding tariff scheme.
On the other hand, the impact of peer to peer traffic
on other services has to be taken into account. The
present traffic profile in IP networks with a domi-
nant peer to peer traffic portion of best effort type 7 Uncertain prediction of Internet
suggests that the differentiated services architecture traffic
[1] is sufficient as a simple scheme to support QoS
by introducing traffic classes to be handled with For the future development there are requirements
different priority. for larger bandwidths in telecommunication net-
works, but especially the role of peer to peer net-
If no more than 20% of the delivered traffic has working is difficult to predict.
strict QoS requirements including voice over IP and
virtual private networks (VPN), then sufficient QoS Most video files are currently represented using
can be guaranteed for those traffic types by strict MPEG compression. In order to achieve an opti-
mum resolution e.g. in HDTV (High definition tele- [2] I. Clarke, S. Miller, T. Hong, O. Sandberg and B.
vision) quality together with coding schemes with- Wiley, Protecting free expression online with
out loss of information, an essentially larger data Freenet, IEEE Internet Computing, Special issue on
volume has to be transmitted. Thus improved qual- peer to peer networking (2002) 40-49
ity for video and other broadband applications will [3] MPEG (Moving pictures expert group) stan-
continue to demand for larger bandwidths. dardization, www.mpeg.org
With regard to peer to peer applications, the ille- [4] F. Hartleb and G. Haßlinger, Comparison of
gitimate usage of copyright protected content and link dimensioning methods for TCP/IP net-
the future effect of countermeasures are unknown works, Proc. IEEE GLOBECOM, (2001) 2240-247
factors of influence. On the other hand, it cannot be
[5] G. Haßlinger, Quality-of-service analysis for
foreseen whether and when video streaming and statistical multiplexing with Gaussian and auto-
downloads will be offered over client server archi- regressive input modeling, Telecommunication
tectures via Internet at acceptable conditions for Systems 16 (2001) 315-334
mass market, which then may satisfy at least a part
of the requirements of the present peer to peer users. [6] Internet2 netflow statistics, weekly reports,
In addition, scalability, security including reliability netflow.internet2.edu (2001-2003)
against denial of service attacks and other aspects [7] B. Krishnamurthy and J. Wang, Traffic classification
seem not to be handled with much care by current for application specific peering, Proceedings of
peer to peer protocols. Nevertheless, peer to peer ACM SIGCOMM Internet Measurement Workshop,
networking has established a new principle for the Marseille, France (2002)
distribution of information and content, which is [8] H. de Meer, K. Tutschku and P. Tran-Gia, Dynamic
more efficient than the client server communication operation in peer-to-peer overlay networks, PIK
for many applications. Journal (2003)
[9] M. Ripeanu and A. Iamnitchi, Mapping the Gnutella
network, IEEE Internet Computing, Special issue on
peer to peer networking (2002) 50-57
Conclusion [10] S. Saroiu, K. Gummadi and S. Gribble, A Measure-
Despite of increasing the traffic volume on the ment Study of Peer-to-Peer File Sharing Systems,
Internet, peer to peer traffic profiles on broadband Proceedings of Multimedia Computing and Network-
platforms have a smoothing effect on the variability ing, San Jose, CA, USA (2002)
of the traffic rate in short and medium time scales [11] S. Sen and J. Wang, Analyzing Peer-to-Peer Traffic
and hot spots as compared to most other traffic Across Large Networks, Proceedings of ACM SIG-
types. They contribute a continuous load of long COMM Internet Measurement Workshop, Marseille,
lasting background file transfers. France (2002) extended version to appear in
ACM/IEEE Transactions on Networking
While these properties partly facilitate the network
dimensioning and planning for service providers,
monitoring, analysis and control of the peer to peer
components and the prediction of their future devel-
opment is difficult and requires further investiga-
tion.
The present traffic mix with a dominant portion of
best effort type data exchanges in the background
has implications for the quality of service concept,
suggesting that differentiated services [1] are suffi-
cient to support other traffic types which generate
smaller traffic volume. This could even include
applications like web browsing, if a clear classifica-
tion of service types would be possible.

References
[1] M. Carlson, W. Weiss, S. Blake, Z. Wang, D.
Black and E. Davies, An Architecture for Dif-
ferentiated Services, RFC 2475 (1998)