Incentives - Peer-to-Peer Computing: Applications, Architecture, Protocols, and Challenges - page 76

tions are stateless, reciprocity generally refers to stateful and history-
based interactions. Specifically, a peer A may serve another peer B at
time t 1 and does not get an immediate return. However, the transac-
tion is recorded in some history database (centralized in some external
entity or distributed in both A and B). At a later time t 2 > t 1 , peer B
serves peer A, possibly because peer B selects peer A as the client due
to the earlier favor from A. That is, as peer A has served peer B before,
peer B would give a higher preference to serve peer A. A critical prob-
lem is: how to tackle a special form of free-riding behavior, namely the
“whitewashing” action (i.e., a user leaves the system and rejoins with a
different identity), which enables the free-rider to forget about his/her
obligations.
5. Reputation-Based Mechanisms: A reputation-based mechanism is
a generalized form of reciprocity. Specifically, while a reciprocity record
is induced by a pair of peers (or a circular list of more than two peers), a
reputation system records a score for each peer based on the assessments
made by many peers. Each service provider (or consumer, depending
on the application) can then consult the reputation system in order to
judge whether it is worthwhile or safe to provide service to a particular
client. Reputation-based mechanism is by nature globally accessible and
thus, peer selection can be done easily. However, the reputation scores
must be securely stored and computed, or otherwise, the scores can-
not truly reflect the quality of peers. In some electronic market places
such as eBay, the reputation scores are centrally administered. But such
an arrangement would again need an external entity and some signif-
icant overhead. On the other hand, storing the scores in a distributed
manner at the peers would induce problems of fraud. Finally, similar
to reciprocity-based mechanisms, whitewashing is a low cost technique
employed by selfish users to avoid being identified as low quality users
which would be excluded from the system.
The different techniques mentioned above are suitable for different appli-
cations. Generally speaking, there are two mainstream applications in P2P
environments: sharing of discrete data, and sharing of continuous data. Ex-
amples of the former include file sharing systems (e.g., Napster), data sharing
systems (e.g., sharing of financial or weather reports), etc. A notable example
of the latter is P2P video streaming. Indeed, there is an important difference
between file sharing and media streaming systems. In the former, a user needs
to wait until a file (or a discrete unit of shared information) is completely
received before it can be consumed or used. Thus, there could be a signif-
icant delay between service request and judgement of service quality. In an
extreme case, a user may not discover that a shared file is indeed the one
requested or just a piece of junk. By contrast, in a media streaming applica-
tion, a user would quickly discover if the received information is good enough.
The quality of service (QoS) metric used is also different in these two differ-