PeerSim: A Scalable P2P Simulator
∗
Alberto Montresor
University of Trento, Italy
alberto.montresor@unitn.it
M
´
ark Jelasity
University of Szeged and HAS, Hungary
jelasity@inf.u-szeged.hu
1. Introduction
The key features of peer-to-peer (P2P) systems are scala-
bility and dynamism. The evaluation of a P2P protocol in
realistic environments is very expensive and difficult to re-
produce, so simulation is crucial in P2P research.
PEERSIM is an extremely scalable simulation environ-
ment that supports dynamic scenarios such as churn and
other failure models. Protocols need to be specifically im-
plemented for the PEERSIM Java API, but with a reasonable
effort they can be evolved into a real implementation. Test-
ing in specified parameter-spaces is supported as well.
PEERSIM started out as a tool for our own research. Af-
ter releasing it under the LGPL open source license, we
were pleased to discover that our pragmatic design choices
appealed to many other independent research groups: at
the time of writing, PEERSIM has been downloaded over
12,000 times and has been used in more than 150 scientific
papers, of which only a small fraction has been written by
the PEERSIM authors.
It is often said that “eating your own dog food” can be
the key to success in an open source project: our own re-
search in diverse P2P application areas ranging from aggre-
gation to topology maintenance, and from synchronization
to global optimization has been carried out using PEER-
SIM [1, 2, 5, 6].
2. Modularity and Configuration
PEERSIM was designed with modularity and ease of config-
uration in mind. The network is modeled as a list of nodes;
a node has a list of protocols, and the simulation has initial-
izers and controls.
Initializers are executed before the simulation, while
controls are executed during the simulation. They may
modify or monitor every component. For example, they
may add new nodes or destroy existing ones; or they may act
∗
Available from http://peersim.sourceforge.net. In Proc.
IEEE P2P 2009, pp 99–100, doi:10.1109/P2P.2009.5284506 . A. Montre-
sor was supported by the European Commission through the NAPA-WINE
Project (Grant No. 214412). M. Jelasity was supported by the Bolyai
Scholarship of the Hungarian Academy of Sciences.
network.size 10000
simulation.cycles 100
D 20
protocol.news Newscast
protocol.news.cache D
init.rand WireKOut
init.rand.protocol news
init.rand.k D
control.conn Clustering
control.conn.protocol news
Figure 1. A simple PEERSIM configuration file
at the level of protocols providing them with external input
or modifying their parameters. Controls can also be used
to passively monitor the simulation; for example, they can
report the variance reduction rate during the execution of
a diffusion-based aggregation protocol [4], or they may re-
port graph-theoretical properties of overlay topologies, such
as diameter, clustering, and so on.
The simulation engine can be cycle-based (protocols are
executed is some specified order) or event-based. These
components can be fully configured and customized. PEER-
SIM provides simple components with basic functionalities,
but users are allowed to replace them with their own alter-
native implementations based on their preferences.
Each simulation is specified by a plain text configuration
file similar to a Java property file. Properties define im-
plementations (Java classes) of components, and they also
specify numeric or string parameters for these components.
Configuration files fully specify an experiment, so they can
be bundled with simulation code for reproducibility.
PEERSIM has been implemented in Java. This allows us
to build experiments at run-time via reading configuration
files and dynamically loading classes using Java reflection.
The example configuration file in Figure 1 defines a
network composed of 10,000 nodes. Figure 2 illustrates
the resulting peersim components. The simulation is run
using the cycle-based engine for 100 cycles. Each node
runs a protocol called news that is implemented by class
Configuration
Manager
Simulation
Engine
Initializers
Controls
Network
Load
Avg
load balancing
Neighbors
Newscast
aggregation
Node
Clustering
...
WireKOut
...
Figure 2. PEERSIM components. Additional
protocol layers and components can easily
be added (shown in light color).
Newscast [6]. The only parameter of Newscast is the
cache size cache, which gets the value of D (i.e., 20). The
same constant is used in the initializer rand, implemented
by class WireKOut that initializes the overlay links man-
aged by news to create a random overlay topology with
constant out-degree k=D. Finally, observer Clustering
periodically reports the clustering coefficient of the overlay
network managed by news.
PEERSIM reads the configuration file and loads the spec-
ified classes at runtime. Based on the configuration file,
either the cycle-driven or the event-driven simulation en-
gines are loaded. The former, to allow for scalability, uses
some simplifying assumptions such as ignoring the details
of the transport layer in the communication protocol stack.
The latter is less efficient but more realistic. Among other
things, it supports transport layer simulation as well. Cycle-
based protocols can also be run by the event-based engine,
but not vice versa.
3. Feature Highlights
Scalability. Internal data structures have a very small
memory footprint. To give a few anecdotic examples: with
simple protocols and the cycle-based engine, the limit of
network size is practically the entire available memory; net-
works of more than 10
7
nodes have been simulated in 4G
memory. In the other extreme case, the most complex pro-
tocols we have simulated using the event-based engine still
scale up to 10
5
nodes or more.
Modularity. All components of the simulated system, as
well as the observer and initializer components can be freely
configured. In addition, the implementation of the compo-
nents of the simulator itself can be customized. It is pos-
sible to replace key components such as the network node
and the event queue of the event-based engine, simply by
implementing the appropriate interface and adding a line to
the configuration file.
Graph abstraction. One strong feature of PEERSIM is its
graph abstraction. It can treat overlay networks as graphs
and can provide various initializers (random and small-
world models, etc.), as well as observers including net-
work diameter, clustering, and connectivity. Overlay net-
work graphs can be exported in various popular formats for
drawing and analysis.
Vector abstraction. Existing PEERSIM modules allow
developers to enrich their simulations by simply writing a
few lines of text in the configuration file. For instance, the
vector package allows a set of protocol instances located
at each of the nodes to be treated as a vector. Vector opera-
tions such as calculating the angle between two vectors, or
initializing vectors, are supported.
Transport layer and churn. To augment the realism of
simulations, PEERSIM can be configured to use trace-based
datasets. This feature is supported only in the event-based
engine. The transport layer is modeled via a special pro-
tocol that provides a message sending service. For exam-
ple, the King dataset is supported, which models the latency
among a collection of geographically distributed nodes [3].
Churn models are available as well, for the cycle-based as
well as the event-based model.
References
[1] O. Babaoglu, T. Binci, M. Jelasity, and A. Montresor. Firefly-
inspired heartbeat synchronization in overlay networks. In
First IEEE International Conference on Self-Adaptive and
Self-Organizing Systems (SASO 2007), pages 77–86, Boston
(MA), USA, 2007. IEEE Computer Society.
[2] M. Biazzini, B. B
´
anhelyi, A. Montresor, and M. Jelasity. Dis-
tributed hyper-heuristics for real parameter optimization. In
Proceedings of the 11th Annual Conference on Genetic and
Evolutionary Computation (GECCO ’09), 2009.
[3] K. P. Gummadi, S. Saroiu, and S. D. Gribble. King: Estimat-
ing latency between arbitrary internet end hosts. In Internet
Measurement Workshop (SIGCOMM IMW), 2002.
[4] M. Jelasity, A. Montresor, and O. Babaoglu. Gossip-based
aggregation in large dynamic networks. ACM Transactions
on Computer Systems, 23(3):219–252, August 2005.
[5] M. Jelasity, A. Montresor, and O. Babaoglu. T-Man: Gossip-
based fast overlay topology construction. Computer Net-
works, 2009. in press, DOI: 10.1016/j.comnet.2009.03.013.
[6] M. Jelasity, S. Voulgaris, R. Guerraoui, A.-M. Kermarrec, and
M. van Steen. Gossip-based peer sampling. ACM Transac-
tions on Computer Systems, 25(3):8, August 2007.
