Dynamic Layer Management in Super-Peer Architectures

Dynamic Layer Management in Super-peer Architectures
Zhenyun Zhuang, Yunhao Liu, Li Xiao
Department of Computer Science and Engineering,
Michigan State University, East Lansing, MI 48824, USA
{zhuangz1, liuyunha, lxiao}@cse.msu.edu
Abstract
The emerging peer-to-peer (P2P) model has recently
gained a significant attention due to its high potential of
sharing various resources among networked users.
Super-peer unstructured P2P systems have been found
very effective by dividing the peers into two layers,
super-layer and leaf-layer, in which message flooding is
only conducted among super-layer. However, current
super-peer systems do not employ any effective layer
management schemes, which means the transient and
low-capacity peers are allowed to act as super-peers.
Moreover, the lack of an appropriate size ratio
maintenance mechanism on super-layer to leaf-layer
makes the system’s search performance far from being
optimal. We propose a Dynamic Layer Management
algorithm, DLM, which can maintain the optimal layer
size ratio, and adaptively adjust peers between
super-layer and leaf-layer. DLM is completely
distributed in the sense that each peer decides to be a
super-peer or a leaf peer independently without the
global knowledge. DLM could effectively help a
super-peer P2P system maintain the optimal layer size
ratio, and designate peers with relatively long lifetime
and large capacities as super-peers, and the peers with
short lifetime and low capacities as leaf-peers under
highly dynamic network situations. We demonstrate that
the quality of a super-peer system is significantly
improved under DLM scheme by comprehensive
simulations.
1. Introduction
Since the development of Napster, millions of peers
join the network by connecting to at least one active peer
in the peer-to-peer overlay network. Each peer acts as
both a client who requests information and services, and
a server who produces and/or provides information and
services.
This work was partially supported by the US National
Science Foundation (NSF) under grant ACI-0325760, by
Michigan State University IRGP Grant 41114.
Today, both unstructured P2P systems, such as
Gnutella [5], and structured P2P systems, such as Chord
[16] and Cycloid [14], are under intensive study [3] [4]
[7] [19]. While unstructured P2P systems are most
commonly used in today's Internet, an early generation is
pure unstructured P2P system. Gnutella is an example of
pure P2P systems, where all the peers are involved in
query flooding process. In these systems, any peer could
be a query source issuing queries to its neighbors, and
the peers receiving the queries may check its local
storage, and response or further relay queries to their
neighbors.
However, peers can be very different from each other
in bandwidth, CPU power, duration times, shared files,
and interests [6]. In pure P2P systems, all peers,
regardless of their capacities, act equal roles and take the
same responsibilities on all the operations. As the
network size increases, the weak peers will seriously
limit the scalability of P2P systems.
To address this problem, super-peer architectures
were proposed, which have been attracting more and
more users in unstructured P2P community. For example,
KaZaA and early Morpheus, based on FastTrack P2P
stack, have dominated the top downloads lists for most
of 2001 and are still increasing in popularity in 2002 [6]
[10]. Schemes are also proposed to introduce Ultra-peers
into Gnutella protocol [2].
The advantages of super-peer systems are well
discussed in [17]. However, current super-peer systems
do not exploit any effective layer management schemes;
thereby leave the following two problems unsolved. First,
assume an optimal value of layer size ratio is given for a
network, how to maintain this layer size ratio? Current
super-peer approaches lack an appropriate size ratio
maintenance mechanism, and make the system’s search
performance far from being optimal. Second, what types
of peers should be elected to super-layer? As far as we
know, no efficient algorithm is given to ensure that the
super-peers have large-capacity and long-lifetime. Since
no global knowledge exist in distributed P2P systems, it
is impossible to tell what values can be “high” or “long”
enough, especially under highly dynamic environments.
This problem is even more difficult to solve when the
network needs to consider more metrics besides capacity
and duration-time.
Proceedings of the 2004 International Conference on Parallel Processing (ICPP’04)
0190-3918/04 $20.00 © 2004 IEEE

In this paper, we propose a dynamic layer
management algorithm to solve these two problems,
called DLM, which is a fully distributed dynamic layer
management algorithm; it can adaptively elect peers and
adjust them between super-layer and leaf-layer. DLM is
completely distributed in the sense that each peer
determines to be a super-peer or a leaf-peer
independently without global knowledge. DLM could
effectively help a super-peer P2P system maintain an
given layer size ratio, and designate peers with relatively
long lifetime and large capacities as super-peers, and the
peers with short lifetime and low capacities as leaf-peers
under highly dynamic network situations.
The rest of this paper is organized as follows. In
Section 2, we discuss the related work. In Section 3, we
describe the importance of layer size ratio in super-peer
systems and network assumptions. We then present the
design rational of DLM in Section 4. In Section 5, we
evaluate the effectiveness of DLM through simulations.
In Section 6, we discuss the side effects of DLM, and we
conclude this paper in Section 7.
2. Related Work
Applications based on Super-peer architectures are
consuming a large portion of Internet traffic. For
example, in June 2002, KaZaA consumed approximately
37% of all TCP traffic, which was more than twice the
Web traffic on the University of Washington campus
network [6]. Sandvine also estimates that 76% of P2P
file sharing traffic is KaZaA/FastTrack traffic and only
8% is Gnutella traffic in the US [9]. Thus, super-peer
architectures are attracting more and more attentions in
P2P research communities.
To well understand the behavior of super-peer
systems, authors in [17] examined the performance
tradeoffs in super-peer systems by considering
super-peer redundancy and topology variations. They
also studied the potential drawbacks of super-peer
networks and reliability issues. To make Gnutella
network more scalable, Singla and Rohrs [2] describe
how Ultrapeers work in an ideal network with a static
topology and a handshaking mechanism based on
Gnutella v0.6 protocol. Some requirements for
super-peers were proposed, such as not fire walled,
suitable operating system, sufficient bandwidth and
sufficient uptime. Authors in [15] present some
incentives to deploy super-peers and propose a
topic-based search scheme to increase the effectiveness
of super-peers.
However, because KaZaA, the most popular
application based on super-peer architectures, is
proprietary and uses encryption, little has been known
about the protocols, architectures, and behaviors of
KaZaA. With over millions of users at anytime, KaZaA
has neither been documented nor analyzed. Since
super-peers and leaf-peers take different responsibilities,
peers need to be assigned to appropriate layers through
some well-designed layer management mechanism. But
as far as we know, no such mechanism is proposed to date.
3. Impact of Layer Size Ratio
We first discuss the importance of maintaining an
appropriate layer size ratio in a super-peer network.
In a super-peer network, the search tasks are mainly
performed by super-peers, which actually form the
“backbone” of the P2P network. Indeed, super-peer
systems take advantage of peers’ heterogeneity by
dividing peers into two layers- super-layer and leaf-layer,
thereby scale better by reducing the number of query
paths [17].
The peers in super-layer are called super-peers and
are responsible for processing and relaying the queries
from the leaf-peers and other super-peer neighbors. Each
super-peer behaves like a proxy or agent of its leaf-peers,
and keeps an index of its leaf-peers’ shared data. The
peers in leaf-layer are called leaf-peers, and keep a small
number of connections to super-peers for the purpose of
reliability.
In super-peer systems, both super-peers and
leaf-peers can submit queries, but only super-peers relay
queries and query responses. A super-peer may forward
an incoming query to its neighboring super-peers. When
receiving a query, a super-peer first checks if the queried
data is stored in local or in its leaf-peers (by checking the
index of its leaf-peers’ objects). If some results are found
in a peer, it will send a QueryHit message back to the
query source along the inverse query path.
Compared with pure P2P systems, super-peer
systems have higher search efficiency because instead of
all the peers, only super-peers are involved in search
processes. Intuitively, an appropriate layer size ratio, i.e.
the ratio of the number of leaf-peers to the number of
super-peers, is of great importance in the sense that, on
one extreme, if most of the peers are in the super-layer,
the system will be more like a pure P2P system since too
many peers take part in searching; on the other extreme,
if too few super-peers are available, the system is more
like a centralized P2P system. A major drawback of a
centralized P2P system is single point of failure.
Some layer management mechanisms use
pre-configured values as the thresholds to select
super-peers. For example, the Ultra-peer Proposal in
Gnutella 0.6 [1] recommends at least 15KB/s
downstream and 10KB/s upstream bandwidth. We argue
that this approach cannot maintain an appropriate size
ratio. Let us look at the example shown in Figure 1.
0190-3918/04 $20.00 © 2004 IEEE

3422
47
19
100
200
100
12
18
8
30
272
260
120
80
290
120
189
27
48
150 280
31
26
47
13
29
160
100
34
Super-peer
Leaf-peer
Figure 1(b) Too M any Super-peers
140
6
20
16
30778
422
Figure 1(c) Too Few Super-peers
Figure 1(a) A Normal Network
Figure 1 Inappropriate Layer Size Ratio
Suppose that a system sets the bandwidth threshold to
50KB/s. Peers with bandwidths larger than 50KB/s can
be selected as super-peers. A network with good size
ratio is shown in Figure 1(a), where the number inside a
peer denotes its bandwidth value. However, after a
certain time period, if most new joining peers have high
bandwidths, the system will soon have too many
super-peers, as shown in Figure 1(b). Or if most joining
peers are weak ones with low bandwidths, the system
will be more like a centralized P2P as shown in Figure
1(c).
Consider a P2P network with n participating peers, in
which ns peers are super-peers and nl peers are leaf-peers.
Assume each leaf-peer connects to m super-peers and
each super-peer connects to ks other super-peers and kl
leaf-peers, on the average. We use η to denote the layer
size ratio of a super-peer network, and it is given by:
η l
s
n
n
= .
Consider all the connections between super-layer and
leaf-layer. The out-degree from super-layer to leaf-layer
is nsk1, and the out-degree from leaf-layer to super-layer
is nlm. Since these two out-degrees are equal, so we
have,
ηm
n
mn
k
s
l
l == . (Equation a)
Also, since ns+nl=n, we have
η+
=
1
n
ns . (Equation b)
Therefore, in a network with n, ns, nl, m defined
above, on the average each super-peer connects to
ηmkl = leaf-peersˈand
η+
=
1
n
ns .
The size ratio of these two layers depends on the
specific systems and applications, since each specific
P2P system has its own targets and functions. Due to
page limit, we leave the discussion of the appropriate
size ratio to our other papers. In this paper, we just
assume that the value of η is given by the protocol, and
every participating peer of the network knows this value.
We also assume that new peers randomly select
active peers as neighbors based on the bootstrapping and
joining mechanisms currently used, although this
randomness does not necessarily lead to totally random
topology [11]. Some studies also show that keeping
high-capacity peers with high degree can make the
system more scalable [18], but it also imports unfairness
among peers, and thereby hinders the peers accurately
report the capacities, as well. In this paper, we assume
that all the peers in the same layer prefer to be treated
equally.
4. Design of DLM
We propose a Dynamic layer management algorithm
(DLM) that intends to deal with the two problems
described in Section 1. These problems are not trivial to
deal with, since no peer knows the global knowledge of
the network. In other words, DLM should achieve two
goals: (1) maintains the size ratio of super layer to leaf
layer; (2) keeps the peers with larger lifetimes and
capacities as super-peers, and keeps the peers with
shorter lifetimes and capacities as leaf-peers.
Unfortunately, these two goals are not always
compatible. For example, at some time, the system may
have a highly skewed layer size ratio, and more
super-peers are heavily needed, but all the joining peers
have low bandwidths. In this case, DLM needs to
consider the two goals simultaneously, and adaptively
promote some leaf-peers, though not “powerful” enough,
to super-layer. Moreover, to perform well under both
stable and highly changeable network situations, DLM
needs to dynamically adjust the promotion and demotion
policies according to the changing environments.
Ideally, the super-peers should be more powerful and
with longer lifetime. To measure the eligibility of a peer,
we define two metrics, capacity and age.
Definition 1: We define capacity as the ability of a
peer to process and relay queries and query responses.
We use capacity(d) to refer to the capacity value of a
peer d. The value of capacity can be computed as,
capacity(d)= ∑=
r
i
ii dvw
1
)(* ,
where r is the number of metrics affecting the peer’s
capacity, vi(d) is the value of ith metric, and wi(d) is the
weight of the corresponding metric. Examples of the
metrics affecting the capacity are bandwidth, CPU
powerfulness and storage space. We assume that a peer’s
capacity value does not change throughout its session
and can be known when it connects the network. For
simplicity, we omit the computation details of a peer’s
capacity, and just use the bandwidth of a peer as its
0190-3918/04 $20.00 © 2004 IEEE

capacity, which does not affect the presentation of DLM
algorithm.
Definition 2: We define age as the length of time up
to now since a peer joins the network up to present.
The lifetime of a peer is the time the peer participates
in the P2P network. It is the gap from the time the peer
joins the network to the time it leaves the network. The
age is less or equal to the lifetime of a peer by the
definition. DLM will automatically promote the leaf
peers with larger ages to super layer, and demote the
super-peers with smaller ages to leaf layer. Although the
capacity value can be set at the time when the peer joins
the system, there are no means to know the lifetime of a
peer, since a peer may leave the network at any time.
One practical way to infer the lifetime is by monitoring
its age: the longer the peer lives, more likely the peer
will lives in the future. So we use the age of a peer in
DLM to denote its lifetime. We use age(d) to denote peer
d’s age value.
In DLM, each peer first collects its neighboring
peers’ information, which includes the capacities, ages,
and the number of neighboring leaf-peers (for a
super-peer neighbor). After processing the collected
information, the peer will determine its own status. We
describe DLM in four phases.
Phase 1: Information Collection
Peers first exchange information with its neighbors
using messages. We design two pairs of messages, which
are employed by a pair of leaf-peer and super-peer. The
formats of these two pairs of messages are shown in
Table 1.
Table 1 Message Formats
Messages Value
Neigh_num_request Null
Neigh_num_response lnn
Value_request Null Null
Value_response capacity age
The first pair of the messages includes
neighbor_num_request and neighbor_num_response
Messages; neighbor_num_request is sent from a
leaf-peer d to a super-peer s to request the leaf neighbor
number of s, and we use lnn(s) to refer to this number.
Message neighbor_num_response is the response from
the super-peer. The second pair of the messages is
value_request and value_response, which are sent
between a super-peer and a leaf peer to query and
respond the latter peer’s capacity and age. Note that
these two pairs of messages can also be piggybacked in
other messages in some P2P protocols.
One issue deserves some words here is how often that
the peers exchange this information. Obviously, higher
frequency means higher accuracy, while more traffic
overhead as well. In this design we employ an
event-driven policy, in which information exchange is
invoked whenever a peer finds that a new connection is
created. In simulation, we have also evaluated other
policies, such as time interval based policy where peers
exchange information periodically. Our results show that
event-driven performs the best in the sense that it
incurred smaller overhead when having the same
performance. Also, the amount of this overhead is trivial
compared with other traffic costs in P2P systems, which
we will discuss in Section 6.
Phase 2: Maintaining Appropriate
Layer-Size-Ratio
One of the goals of DLM is to maintain appropriate
layer size ratio. To achieve this goal, DLM first needs to
estimate the extent of appropriateness of current layer
size ratio.
The basic idea of DLM is based on three observations.
First, current layer size ratio can be easily calculated if
the global knowledge is known, but in reality no peer
knows this global knowledge. Second, it is trivial for a
peer to collect its neighbors’ information. Third, due to
the randomness of the neighbor selection mechanism in
super-peer systems, to some extent, the current numbers
of leaf neighbors of super-peers can reflect the current
layer size ratio. That is, if the super-layer size is too
small, the average number of leaf neighbors of a
super-peer will be larger than kl, the “optimal” leaf
neighbor number; otherwise it will be less than kl. Since
each peer knows the optimal value of η, then the value of
kl can be computed using Equation a. Therefore, lnn, the
leaf neighbor number of some super-peer, can be used as
a flag of the current layer size ratio when compared with
the value of kl.
To illustrate the comparison method, we now define
the related set of a peer.
Definition 3: we use G to represent the related set of
a peer in the other layer. For a super-peer s, we define
G(s) as the set of its current neighboring leaf-peers.
While for a leaf-peer l, we define G(l) as the set of
super-peers that it has connected within a period of time
Tl.
The definitions of G on super-peers and leaf-peers
are different as a super-peer normally keeps connections
to many leaf-peers, while a leaf-peer normally only
connects to a few super-peers. To more accurately
estimate the network condition, we hope that the size of
G is large enough, so that we consider all the super-peers
that a leaf-peer has contacted in a recent period of time.
In simulation, G contains all the super-peers that a
leaf-peer has connected since it joins the network.
0190-3918/04 $20.00 © 2004 IEEE

For a super-peer s, since it connects to some
leaf-peers, it can directly use lnn(s). While for a leaf-peer
l, it uses the average lnn value of the super-peers in G(l).
We let µ to denote the extent of inappropriateness of
current layer size ratio compared to the optimal layer size
ratio η, and it is computed as,
µ=log(lnn/k1).
We can see that a positive value of µ, where lnn is
larger than k1, means that there are too few super-peers in
the system, since super-peers have more leaf-peer
neighbors than normal. While a negative value of µ,
where lnn is smaller than k1, means that there are too
many super-peers. Furthermore, a larger absolute value
of µ means a larger degree of inappropriateness. Thereby
the value of µ reflects the system requirement, and it is
used to adjust some other parameters, which determine
the possibility of a peer’s being promoted or demoted.
Phase 3: Scaled Comparisons of Capacity and
Age
DLM automatically promotes the leaf-peers with
large capacities and longer lifetimes to super-layer, and
demotes the super-peers with small capacities and shorter
lifetimes to leaf-layer. The decision of promotion or
demotion is based on the comparison results with other
peers.
One straightforward way of comparison is directly
comparing the metric values of a peer with other peers.
However, comparing in such a direct way may fail to
maintain the layer size ratio successfully. Consider a
scenario in which the system needs more super-peers, but
the leaf-peers all have larger metric values than the
current super-peers. The results of simple comparisons
would forbid any leaf-peer to be promoted; thereby the
system cannot adjust the layer size ratio at all.
DLM improves the direct comparison by using
“scaled-comparison”. Since the capacity and age metrics
are disjoint, that is, a peer with high-capacity not
necessarily has larger age and vice versa, we analyze
these two metrics individually. In scaled-comparison,
DLM introduces two scale parameters, Xcapa and Xage,
corresponding to the two metrics respectively. These two
scale parameters are adjusted dynamically by DLM
based on the value of µ to reflect the system
requirements.
For each peer that runs DLM, it uses two counting
variables, Ycapa and Yage, corresponding to Capacity and
Age metrics, respectively. A peer sets the value of these
two counting variables by comparing its metric values
with the peers in its related set, G. For a peer d, the
pseudo codes of the scaled-comparison are listed below.
for all peer di in G(d)
if (capacity(di)* Xcapa >capacity(d))
Ycapa +=1/ (size of G(d));
if (age(di)* Xage >age(d))
Yage +=1/ (size of G(d));
We can see that Ycapa and Yage store the fractions of
peers that have “larger” metric values than those of d in
G(d). Ycapa reflects the relative capacity value of a peer
compared to the peers in the other layer; while Yage
reflects the relative age value of one peer compared to
the peers in the other layer. These two variables will be
used to determine the eligibility of the peer’s promotion
or demotion.
The values of Xcapa and Xage are adjusted according to
the value of µ. For a super-peer, if it finds that the system
needs more super-peers, it will decrease the possibility of
its demotion by decreasing the two scale parameters.
Otherwise, it will increase the possibility of its demotion
by increasing the scale parameters. While for a leaf-peer,
if it finds that more super-peers are needed, it will
decrease the scale parameters in hoping to increase the
promotion possibility; otherwise it will increase the scale
parameters to decrease the promotion possibility.
Phase 4: Promotion or Demotion
For a leaf-peer l, if Ycapa and Yage are small enough, it
means that many super-peers in G(l) have metric values
“smaller” than it. Thus, l may assume that it has
comparatively large metric values and may decide to be
promoted to super-layer. While for a super-peer s, if Ycapa
and Yage are large enough, it means that many leaf-peers
in G(l) have metric values “larger” than it. Thus, s may
assume that it has comparatively small metric values and
may decide to be demoted to leaf-layer.
We use two threshold variables, Zcapa and Zage, in the
determination. For a leaf-peer l, if Ycapa and Yage are
smaller than Zcapa and Zage, respectively, it will be
promoted to be a super-peer. In promotion, the leaf-peer
keeps its current connections to other super-peers. The
scenarios before and after peer l’s promotion are
illustrated in Figure 2(a) and 2(b).
H
S2
DS1
L I F G H
S2 D
S1
L
I F G
H
D
Fig 2(a) Before Promotion Fig 2(b) After Promotion
Super-peer
Leaf-peer
Figure 2 Promotion of a Leaf Peer
S2
S3S1
SI G HH
S2
S3S1
S
I F G
H
D
Fig 3(b) After DemotionFig 3(a) Before Demotion
Super-peer
Leaf-peer
F
Figure 3 Demotion of a Super-peer
0190-3918/04 $20.00 © 2004 IEEE

5. Performance Evaluation
For a super-peer s, if capa and Yage are larger than Zcapa
and Zage, respectively, it will be demoted to be a leaf-peer.
In demotion, the super-peer only keeps m of its current
connections to other super-peers and drops the
connections to leaf-peers. The scenarios before and after
peer s’s demotion are illustrated in Figure 3(a) and 3(b).
The values of threshold variables, Zcapa and Zage, are
also adjusted according to the value of µ. When more
super-peers are needed, super-peers will increase the
values of the threshold variables to reduce the demotion
tendencies, and leaf-peers will reduce the values of the
threshold variables to increase the promotion tendencies.
For the cases that there are too many super-peers, inverse
measures will be taken accordingly.
Essentially, to achieve the two goals we set in the
beginning of Section 4, DLM needs to consider two
factors: the layer-size-ratio factor (µ), and the
capacity-age factor. The first factor reflects the
requirement of keeping the optimal layer size ratio, and
the second factor reflects the requirement of keeping
large-capacity and large-age peers in the super-layer.
Since the two scale parameters, Xcapa and Xage, are
adjusted based on the values of the first factor, DLM can
achieve the first goal: maintaining the optimal layer size
ratio. Similarly, by using scaled comparisons method,
DLM achieves the second goal: keeping large-capacity
and large-age peers on the super-layer.
We use simulation to evaluate DLM and compare it
to preconfigured algorithms.To collect real data, we
implement two Gnutella clients based on the publicly
available Mutella [8]. One client participates into the
network as an ultra-peer, and the other acts as a leaf-peer.
Using these two clients we collect some first-hand data,
such as the lifetimes of ultra-peers and leaf-peers, the
number of open connections and average query
frequencies of peer The collected data are consistent with
the data presented in previous studies [6, 12, 13]. We
configure our simulation environments based on these
observations. The values of the parameters are shown in
Table 2.
Table 2. Simulation Parameters
Parameter Value Description
n 50000 # of peers in the network
nl 48780 # of preferred leaf-peers
ns 1220 # of preferred super-peers
η 40.0 Layer size ratio
m 2 Super-peer neigh. # of a leaf-peer
kl 80 Aver. leaf neigh. # of super-peers
ks 3 Aver super neigh. # of super-peers
We evaluate DLM under various simulation
environments. We first simulate a stable network by
assigning new peers with capacity and lifetime values
based on previous studies [12]. We then simulate a
dynamic network with dynamic means of the
distributions. For these two cases, the network sizes are
not changed. We also evaluate DLM on networks with
different sizes. The results are consistent, so we only
present the results in dynamic environments, in Figures 4,
5, and 6.
In a stable network, the simulation starts “cold”, i.e.
without any peer. The size of the network increases with
new peers joining until reaches the designated size. Then
with time going, whenever a peer dies, a new peer is
created and joins the network, thereby the network size
does not change. The new peer is always assigned to leaf
layer first, and DLM will promote eligible leaf peers to
super layer. The lifetime and capacity values of a new
peer are generated based on the distributions we
observed and shown in previous studies.
We simulate the dynamic networks by varying the
means of the capacity and age distributions for new
joining peers. Starting from the 300th
time unit, the
lifetime values of new joining peers are generated using
the method as before, but with half mean values. Thus
from that time, the new peers have smaller lifetime
values. Figure 4 plots the average ages of each layer. As
expected, the age of super-layer is much larger than that
of leaf-layer, regardless the changing environments.
$YHUDJH $JH &RPSDULVRQ

6LPXODWLRQ 7LPH 0LQXWH

$YHUDJH
$JH

6XSHU /DHU
/HDI /DHU
Figure 4 Average Age
$YHUDJH DSDFLW RPSDULVRQ

6LPXODWLRQ 7LPH 0LQXWH

Dynamic Layer Management in Super-Peer Architectures

More Related Content

What's hot

Viewers also liked

Similar to Dynamic Layer Management in Super-Peer Architectures

More from Zhenyun Zhuang

Recently uploaded

Dynamic Layer Management in Super-Peer Architectures