You are on page 1of 16

December 10, 2019 17:31 JOIN S0219265919500105 page 1

Journal of Interconnection Networks


Vol. 19, No. 4 (2019) 1950010 (16 pages)
c The Author(s)
DOI: 10.1142/S0219265919500105

Three-Tier Delaunay Network as a Topology


for Peer-to-Peer Overlay∗
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

SATOSHI FUJITA
Department of Information Engineering, Hiroshima University,
Kagamiyama 1-4-1, Higashi-Hiroshima, 739-8527, Japan
fujita@se.hiroshima-u.ac.jp
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

Received 19 December 2018


Accepted 29 November 2019

This paper considers the problem of improving the routing performance of hierarchical
Delaunay networks. Delaunay network is a network topology for peer-to-peer systems
based on the Delaunay triangulation of a set of points associated with a set of peers.
It is known that Delaunay networks have a favorable property as a topology for peer-
to-peer systems such that a greedy routing scheme always delivers a given message to its
destination without encountering a dead-end. The key idea used in the proposed method is
to apply a hash function to the address of participant peers. More concretely, by applying
a hash function to the coordinate point of the peers and by associating several points to
each peer, we could realize an overlay so that the number of hops to the destination in the
original network could be significantly reduced.

Keywords: Peer-to-peer system; Delaunay triangulation; hierarchical network.

1. Introduction
Development of sustainable publish/subscribe systems is an urgent issue for many
countries and regions including Japan, Korea and Taiwan which frequently suffer
from country-wide disasters such as big earthquake and typhoon. A lessen learnt
from the Great East Japan earthquake happened on March 11, 2011, is that many
network applications such as bulletin board stop their service due to the lost of
servers, regardless of the aliveness of the underlying network infrastructure. This
fact motivates us to realize such network services in a peer-to-peer manner which is
expected to attain a high availability compared with classical client/server systems.

∗ An earlier version of this paper was presented as “Hierarchical Delaunay Network for Peer-to-Peer
Overlay with Address Hashing,” at PDCAT 2017 by the same author.
This is an Open Access article published by World Scientific Publishing Company. It is distributed
under the terms of the Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC) License
which permits use, distribution and reproduction, provided that the original work is properly cited,
the use is non-commercial and no modifications or adaptations are made.

1950010-1
December 10, 2019 17:31 JOIN S0219265919500105 page 2

S. Fujita

To realize a reliable publish/subscribe system covering a wide geographical area


such as countries and provinces without heavily depending on specific servers, we
need to have a distributed system which is tolerant against faults and churn at
the application level, in addition to the robustness at the network level.1,3,4,6,15
In particular, the survivability of a routing mechanism is a crucial factor for such
distributed systems, which requests that a given message is successfully delivered to
its destination even if routing tables held by the peers are disrupted. In this paper
we focus on Delaunay network,10 which is based on the Delaunay triangulation of a
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

point set in a two-dimensional Euclidean space, as the underlying network topology


for such systems. Delaunay network has been applied to many network applications
in the past because of its favorable property such that a greedy routing scheme
which selects the next hop of a message to minimize the Euclidean distance to the
destination, always delivers the message to the destination (a formal definition of
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

the Delaunay network is given in Sec. 2). In addition, as will be explained later, we
can check whether a given network structure is a Delaunay network in a distributed
manner. In other words, it could adjust itself to satisfy the favorable property even
if a part of the network is damaged by artificial attacks and/or natural disasters.
This paper focuses on the routing performance of Delaunay networks. A tri-
angulation of point set P is a partition of the convex hull of P by line segments
so that the shape of each region is a triangle. Since links in a Delaunay network
correspond to line segments used in the corresponding Delaunay triangulation and
the number of such line segments is proportional to |P |, the maximum distance
between two peers in a Delaunay
p network with respect to the number of links on a
path is proportional √ to |P |. In other words, in a Delaunay network consisting of
N peers, it takes Ω( N ) hops to deliver a message to its destination in the worst
case. To overcome such a weakness of Delaunay networks, several methods have
been proposed in the literature, such as the introduction of hierarchical structure
to the overlay and a hybrid with other network topology.17,18 The routing in such
hierarchical networks generally consists of three phases; namely (1) the routing to
a gateway peer participating in both upper and lower layers through a path in the
lower later, (2) the routing to a gateway peer closest to the destination through a
path in the upper layer, and (3) the routing to the destination through the lower
layer. The performance of such a three-phase routing is affected by the ratio of peers
participating in the upper layer; namely if it is too small, it takes long time before
reaching the gateway and if it is too large, it takes long time for the routing in the
upper layer. Theoretically, this value is balanced when the upper layer consists of
O(N 2/3 ) peers, and could not be improved as long as the overlay consists of two
Delaunay networks. In addition, although it could be overcome by using three or
more layers, it incurs additional cost for peers selected for the member of higher
layers.
In this paper we propose a method to improve the routing performance of (hier-
archical) Delaunay networks without incurring additional cost for each participant.

1950010-2
December 10, 2019 17:31 JOIN S0219265919500105 page 3

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

The key idea is to apply a hash function to the coordinate point of each peer. More
concretely, we propose a three-tier network consisting of upper, lower and middle
layers each of which is a Delaunay network of given point set. By applying a hash
function to the coordinate point of peers participating in the upper layer, we could
realize an overlay of the middle layer so that the number of hops to the destination
could be effectively reduced. The results of experimental evaluations indicate that
although we could reduce the average hop count to the destination in the original
Delaunay network to 80% by applying conventional hierarchical Delaunay network,
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

we could further reduce it to 30% by applying the proposed method.


The remainder of this paper is organized as follows. Section 2 gives preliminar-
ies. Section 3 overviews related work concerned with hierarchical Delaunay networks.
Section 4 describes the details of the proposed network and the routing algorithm.
Section 5 summarizes the results of evaluations. Finally Sec. 6 concludes the paper
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

with future works. This paper is an extended version of a conference paper pre-
sented at PDCAT 2017.5 Although it contains several additional contents, the main
difference to the conference version is that it contains experimental results which
were not given in the conference version.

2. Preliminaries
This section describes an overview of Delaunay networks with a formal definition of
Delaunay triangulation.

2.1. Delaunay triangulation


Let P = {p1 , p2 , . . . , pN }(⊂ R × R) be a set of pointsa on a two-dimensional Eu-
clidean space. A triangulation of P is a collection of triangles which is obtained
by dividing the convex hull of P by line segments so that: (1) vertices of each tri-
angle are points in P and (2) any triangle contains no point in P except for its
three vertices. A Delaunay triangulation of P , which will be abbreviated as DT
hereafter, is a triangulation satisfying the following condition which will be referred
to as Condition A: for any triangle in the triangulation, the circumscribed circle
of the triangle contains no points in P except for the vertices of the triangle.b This
definition can be directly applied to the construction of DT for a given point set P ,
as follows:
procedure GENERATE DT
(1) Start from any triangulation of P ;
(2) Repeat the following steps until Condition A holds:

a This paper assumes that points in P are located in “general” position in the sense that any four
points are not on the circumference of a single circle.
b More precisely, there might exist another point in the circle only when those four points are
arranged on the circumference of a circle. Note that we exclude such cases by assuming that points
are located at general positions.

1950010-3
December 10, 2019 17:31 JOIN S0219265919500105 page 4

S. Fujita

circumscribed circle
c

a flip
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

x
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

b
Fig. 1. Flip of line segment in procedure GENERATE DT. In this example, line segment bc is
replaced by ax by the flipping.

(a) Let ∆abc be a triangle which contains x in its circumscribed circle. Without
loss of generality, assume that x is on the opposite side of a with respect to
line bc and if there are several such points, a vertex of a triangle to have b, c
as two vertices is selected as x.
(b) Replace segment bc by segment ax. In the following, we will call this operation
“flip.”

See Fig. 1 for illustration. Note that in Step 2a, such a vertex x can be uniquely
determined since we are assuming a triangulation. As for Step 2b, since the radius
of the circumscribed circle of ∆abx is smaller than that of ∆abc, and since both
circumscribed circles have segment ab as a chord in common, the circumscribed
circle of ∆abx does not contain c. Since similar claim holds for ∆cbx, every flip
reduces the number of violations by at least one. According to the Euler’s formula
concerned with plane graphs, any triangulation of N points contains at most O(N )
triangles. Thus the maximum number of flips necessary to obtain a DT is at most
O(N ). In fact, it is known that DT can be calculated in Θ(N log N ) time since DT
is the dual of Voronoi diagram and the problem of calculating the Voronoi diagram
of a given point set is as hard as the sorting of N elements.

2.2. Delaunay network


The idea of adopting DT as the topology of interconnection networks has attracted
considerable attention during past decades. At an early stage of the research, it
was mainly applied to wireless networks such as mobile ad hoc networks and sensor
networks,8,9,16 but recently the field of application shifts to P2P systems.

1950010-4
December 10, 2019 17:31 JOIN S0219265919500105 page 5

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

Liebeherr et al.10 proposed a P2P overlay such that peers are associated with
points on a two-dimensional space (e.g., geographical location of the peer acquired
through GPS) and communication links are associated with line segments in a DT
of the point set. This overlay has several advantages such that: the physical distance
between two adjacent peers connected by a link is short, and each peer has six neigh-
bors on average. In addition, it has a favorable property so that the greedy routing
of a message towards the direction of the point associated with the destination peer
always reaches the destination without encountering a dead-end.10
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

Let p(i) denote the coordinate point associated with peer i and assume that i
knows the coordinate points of peers in subset N (i), where N (i) denotes the set
of peers adjacent with i. In the greedy routing, after receiving a message m from
a neighbor, peer i reads out the coordinate point p(m) of the destination from the
header of the message, and selects peer j ∗ satisfying the following equation from
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

N (i):
dist(p(j ∗ ), p(m)) = min {dist(p(j), p(m)},
j∈N (i)

where dist(p, q) denotes the Euclidean distance between p and q. It then forwards
the received message to the selected peer j ∗ .
Delaunay network is organized (and maintained) by applying flip operations to
the initial triangulation in a distributed manner,10 while avoiding simultaneous flips
conflicting at a link to guarantee the correctness, which is realized by “locking” all
peers concerned with triangles adjacent with the target triangles of the current flip.
Such a lock can be realized by solving a variant of the maximal independent set
problem in a distributed manner. Since any unlocked flips can be executed concur-
rently, it significantly reduces the convergence time compared with the sequential
execution of flips.

3. Hierarchical Delaunay Networks


There are several previous works concerned with the hierarchical Delaunay network.
Ohnishi et al. proposed a network called LRC (Long Range Contact),13 which is
obtained by connecting each peer in a Delaunay network with (at least) log N remote
peers via shortcut links. More precisely, it connects each peer x to peers at 2i hops
away from x by shortcut links for each 1 ≤ i ≤ log N . The reader might wonder
that the number of peers at distance 2i cannot be bounded by a constant in two-
dimensional case unlike one-dimensional case, since such peers could exist along the
circumference of a circle with radius 2i centered at peer x. To overcome such an issue,
LRC considers “bands” of a constant width parallel to x- and y-axes for each peer,
and selects targets connected via shortcut links from the set of peers intersecting
with those bands. In other words, LRC attaches shortcut links almost parallel to the
x- or y-axis to the original Delaunay network, which makes the resulting graph to
be a two-dimensional version of the skip graph2 rather than a hierarchical version of
Delaunay network. In fact, the improvement of the routing performance is attained

1950010-5
December 10, 2019 17:31 JOIN S0219265919500105 page 6

S. Fujita

Upper layer
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

Lower layer

Fig. 2. Hierarchical Delaunay overlay network consisting of two layers. Red circles represent peers
selected from the lower layer. The Delaunay network in the upper layer is constructed using the
same metric with the lower layer.

by the hyper-cubic shortcut links, and the fact that the underlying graph is organized
according to DT is not used except for the calculation of the intersection of Voronoi
regions with the given bands.13
Skip Delaunay Network (SDN)18 proposed by the same authors is a more explicit
extension of the skip graph. Recall that the basic idea of skip graph is to assign a
random bit sequence to each peer as the ID and to construct ring structure defined
by the lexicographical order of IDs in a hierarchical manner.2 The baseline is a
single ring consisting of all peers and the depth of the hierarchy is determined
by the prefix length. More concretely, the ith level (i ≥ 0) of the hierarchy is a
collection of rings each of which consists of peers to have ID with the same prefix of
length i. Note that by definition, the ith level consists of (at most) 2i independent
rings. In SDN, on the other hand, several peers to have ID with the same prefix
of length i are connected with a Delaunay network instead of a ring. For example,
the 0th level is an ordinary Delaunay network consisting of all peers and the first
level consists of two independent Delaunay networks, one consists of peers to have
ID with prefix 0 and the other consists of peers to have ID with prefix 1. Such a
partitioning of the set of peers proceeds until the resulting set becomes a singleton,
which indicates that the expected number of Delaunay networks relevant with each
peer is log N . SDN is simple and elegant compared with LRC. However, the cost
spent by each peer becomes log N times of the cost of ordinary Delaunay network,
while the improvement of the routing performance is much less than log N times.

1950010-6
December 10, 2019 17:31 JOIN S0219265919500105 page 7

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

Hierarchical Delaunay Overlay Network (HDOV) proposed by Shinomiya et al.


adopts a probabilistic method to select members of the upper level of the hierarchy.
In this scheme, in contrast to LRC and SDN, merely selected peers are connected by
shortcut links at upper levels.17 The probability of being selected as a member of an
upper level is a function of the density of peers in the lower level. More concretely,
(1) peers in a low density region are selected as a member of upper level with high
probability so as to keep the connectivity of the network in the upper level and
(2) peers in a high density region are selected as a member with low probability
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

so as to avoid redundant connections. The density of peers around each peer is


estimated using the size of nearby Voronoi regions (including its own region) and
the average size over all peers. Hence, if the ratio of selected peers is bounded by a
constant, the average cost spent by each peer can be bounded by a constant times
of the cost of original Delaunay network, even when such a probabilistic selection
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

is recursively applied. The authors of Ref. 17 applied HDOV to the information


gathering in sensor networks and propose several protocols including protocol for
processing range queries.

4. Proposed Overlay Network


This section describes the details of the proposed method. The proposed network
consists of three layers. The lower layer consists of all peers participating in the
system and the upper layer consists of several probabilistically selected peers as in
HDOV. The role of the middle layer is to connect remote peers in the upper/lower
layers without significantly increasing the average cost. The structure of the network
in each layer is a Delaunay network consisting of a given point set. To realize an
effective shortcut in the middle layer, we use hash function h from R × R to R × R.
In the following, after describing the details of the structure of upper and middle
layers in Secs. 4.1 and 4.2, respectively, we discuss the design of hash function h in
Sec. 4.3. The routing scheme is described in Sec. 4.4.

4.1. Upper layer


Let V be the set of peers distributed over a two-dimensional space. Let p(i) ∈ R × R
denote the coordinate point of peer i. Let U be a subset of peers selected from V
according to an appropriate probabilistic distribution Ω, such as:
• To select peers from V with an equal probability q (in this case, the expected
size of U becomes q|V |).
• To select peers with a probability calculated by the density of peers so that the
probability of selecting x as a member of U monotonically increases as the the
density of peers around x increases, as in Ref. 17.
• To reduce the variance of the density of selected peers in the first method. More
concretely, it first partitions the two-dimensional space into small squares, and
selects k(≥ 1) peers from each square with a certain probability.

1950010-7
December 10, 2019 17:31 JOIN S0219265919500105 page 8

S. Fujita

Upper layer
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

Middle layer
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

Lower layer

Fig. 3. Proposed overlay network. The middle layer consisting of 2|U | = 8 points is the third
Delaunay network in the overlay network.

In the proposed method, peers in U serve as a “bridge” between different layers,


which implies that the load of selected peers becomes heavy compared with other
(unselected) peers.

4.2. Middle layer


The set of points in the middle layer is given as

{p(i) : i ∈ U } ∪ {h(p(i)) : i ∈ U },

where U is the set of peers participating in the upper layer and h is an appropriate
hash function from R × R to R × R. Those 2|U | points are connected through a
Delaunay network similar to the upper layer. A key point is that each peer in U is
associated with two points in the middle layer which can be used to deliver a message
to its destination (in the middle layer) by repeating greedy forwarding. Of course
those points in the middle layer have specific roles; i.e., points in {p(i) : i ∈ U } play
the role of “bridge” connecting to the other two layers, and points in {h(p(i)) : i ∈ U }
play the role of “tunnel” to a remote point in the middle layer. Namely every message
could pass through the tunnel connecting p(i) and h(p(i)) with no cost even if the
distance of those points in the Euclidean space is very large. The reader should
note that since the average degree of any Delaunay network is bounded by six, the
degree of each peer in the middle layer associated with two points is bounded by 12
on average.

1950010-8
December 10, 2019 17:31 JOIN S0219265919500105 page 9

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

4.3. Hash function


Although we can use any function as h, this paper proposes three concrete hash
functions described below.
The first function h1 maps a given point to a point in R × R which is selected
from a certain region (e.g., convex hull of {p(i) : i ∈ U }) with uniform probability.
Since the distance between mapped points is randomly determined for each pair of
peers, it realizes a random shortcut in the DT of the upper layer. The impact of
such random shortcuts has been certified by the theory of small-world networks.19
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

The second function h2 maps point (x, y) to a point in R × R as follows:


 
1 1
h2 (x, y) = , ,
x y
where we assume that the centroid of the convex hull is set as the origin and the
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

radius of the convex hull is almost twice of the unit length. Intuitively speaking, this
function swaps the inside and the outside of a unit circle. More precisely, it maps
a point far from the origin to a point close to the origin and vise versa. With this
function, peers corresponding to the vertices of the convex hull in the upper layer
are likely to be connected by a link in the middle layer.
The third function h3 realizes a kind of transposition in a higher dimensional
space. Suppose that the coordinate value of each point p(·) is represented by an
integer of 2n bits. Let I be the set of 2n bit integers. Let g be a function from I to
I which is defined as follows: for each b2n−1 b2n−2 · · · b0 ∈ I,
g(b2n−1 b2n−2 · · · b0 )
=bn−1 bn−2 · · · b0 b2n−1 b2n−2 · · · bn .
Note that g realizes a transposition of points in the 2n dimensional space. With this
function g, the third hash function h3 is defined as follows:
h3 (x, y) = (g(x), g(y)).

4.4. Routing scheme


The proposed routing scheme consists of the following three phases, each of which
is executed at most once:
(1) Routing from the source peer to a bridge peer through the overlay of the lower
layer (Phase 1).
(2) Routing to a bridge peer closest to the destination through the overlay of the
upper and middle layers with the help of tunneling (Phase 2).
(3) Routing from the last bridge peer to the destination through the overlay of the
lower layer (Phase 3).
The routing scheme significantly reduces the hop count to the destination by ef-
fectively utilizing shortcut links in the upper and middle layers. The reader should

1950010-9
December 10, 2019 17:31 JOIN S0219265919500105 page 10

S. Fujita

note that when we use an ordinary two-tier Delaunay network without hash func-
tion, the routing time is minimized when the time for Phase 1 and Phase 2 are
balanced; i.e., the effect of shortcut increases as the number of selected peers de-
creases, but it would take long time before reaching the bridge peer if the number of
selected peers is too small. In contrast, since the proposed scheme realizes random
shortcut links with no additional cost, it could significantly reduce the routing time
without reducing the number of selected peers.
There are several issues to be overcome to effectively realize such an idea in
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

actual networks. The first issue is how to deliver a given message to a bridge peer as
soon as possible in Phase 1, and the second issue is how to deliver the given message
to a bridge peer closest to the destination as soon as possible in Phase 2. In the
following, we describe our solution to the above issues in detail.
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

4.4.1. Phase 1
One possible implementation of Phase 1 is to conduct a greedy routing toward the
destination of the message and to switch to Phase 2 when it meets a bridge peer on
the routing path. Although it is simple and natural, such a naive scheme could not
guarantee the usage of shortcut links, which implies that the worst case performance
of the ordinary Delaunay network could not be improved. To overcome this issue,
the proposed routing scheme uses the notion of spread net. More concretely, as a
preprocessing, it firstly spreads a “net of paths” obtained by connecting nearby
bridge peers in the lower layer, which is used to “catch” message transmitted by the
source peer. A message caught by a path in the net is lead to an end vertex of the
path (i.e., to a bridge peer); thus it could always lead a transmitted message to a
bridge peer.
A detailed description of the procedure is as follows. Recall that U denotes the
set of peers selected for the upper layer. Let Nup (u)(⊆ U ) denote the set of adjacent
peers of u in the upper layer. At first, as a preprocessing, each peer conducts the
following steps.
Preprocessing for Phase1
(1) Each u ∈ U transmits spoke message to all peers v in Nup (u) through the lower
layer, which contains: (1) the coordinate point p(u) of the source of the message
and (2) the number of links traveled by the message, where the latter value is
incremented by one for every time of relaying the message.
(2) The transmitted spoke message is delivered to its destination v using a greedy
routing. Each peer w on the delivery path from u to v could learn from the
message that: (1) there is a bridge peer u around w with (2) the number of hops
to u.

The trajectory of spoke messages forms a net of paths. In other words, each peer
w on a path in the net can know the coordinate value and the hop count to its nearby

1950010-10
December 10, 2019 17:31 JOIN S0219265919500105 page 11

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

bridge peers in the lower layer. Thus, if w receives a message to destination v ∗ from
its neighbor and if v ∗ is sufficiently far from w and its nearby bridge peers, i.e., if it
recognizes that the use of shortcut links reduces the hop count to destination v ∗ , it
heads the message to its nearby bridge peer as follows:

Heading to Bridge

(1) Let N ∗ (w) be the set of bridge peers recognized by w. Recall that w knows the
hop count to peers in N ∗ (w) in addition to their coordinate value.
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

(2) w selects a peer x with the smallest hop count from N ∗ (w), sets the tentative
destination of the message to peer x, and transmits it out towards the tentative
destination x.
(3) The transmitted message is delivered to its tentative destination using a greedy
routing.
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

(4) After x receiving the message, x starts the delivery of the message in Phase 2
toward a bridge peer closest to its actual destination.

Since each path organized by a spoke message emulates a link in the upper layer,
any message issued by a peer will always be “caught” by the net of paths.

4.4.2. Phase 2
Recall that the set of points in the middle layer is a union of two subsets {p(i) : i ∈
U } and {h(p(i)) : i ∈ U }. For brevity, in the following exposition, we call points in
the former subset red and points in the latter subset white. Note that every bridge
is red.
The objective of Phase 2 is to deliver a message given to a red point in the
middle layer to another red point closest to the destination as quickly as possible.
To this end, we design a routing scheme so that: it firstly delivers the message to
a point which is closest to the destination in the middle layer and is either red or
white (Step 1); if the arrived point is not red, then it forwards the message to a
nearby red point through the middle layer (Step 2); and delivers the message to a
red point closest to the destination in the upper layer (Step 3). Step 3 is the same
as the routing in ordinary Delaunay network. In contrast, the routing in Step 1 is
accelerated by effectively utilizing the tunneling and Step 2 is realized by keeping
a pointer to nearby red points for each white point. A concrete description of the
procedure is as follows:

Routing for Phase2

(1) Two peers u and v are said to be adjacent in the middle layer if their corre-
sponding points are connected by a link in the middle layer. Let Nmid (u) be the
set of peers adjacent to u in the middle layer. Assume that each peer u in U
knows Nmid (u) with their corresponding coordinate values.

1950010-11
December 10, 2019 17:31 JOIN S0219265919500105 page 12

S. Fujita

(2) After receiving a message m from a neighbor, peer w refers to the coordinate
point p(m) of the destination. It then selects a peer x in Nmid (w) such that:
min{dist(h(p(x)), p(m)), dist(p(x), p(m))}
is the smallest among peers in Nmid (w) and it is smaller than
min{dist(h(p(w)), p(m)), dist(p(w), p(m))}.
If there is no such x, then go to Step 3. Otherwise, w forwards the message to
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

x through a link in the middle layer.


(3) Now the message arrives at a point in the middle layer closest to the destination.
If the current point is red, then it simply switches to Phase 3 since it has arrived
at a bridge peer closest to the destination. Otherwise, it moves to a nearby red
point by traversing a link connecting from white point to nearby red points
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

prepared at the middle layer. It switches to Phase 3 if it successfully arrives at


a red point.

5. Evaluation
5.1. Setup
In this section, we evaluate the routing performance of the proposed method by
experiments. In the experiments, we measure the average hop count over all pairs
of peers in each of the following four networks and compare them (note that in
general, the hop count required for the routing of a message from a peer to another
peer in an overlay is different from the distance between those two peers in the
overlay):
(1) ordinary Delaunay network SIMPLE constructed over a given set of peers which
are randomly distributed on a two dimensional plane,
(2) HDOV17 consisting of two layers,
(3) an enhancement of HDOV with a technique given in Sec. 4.4.1 denoted PHASE1,
and
(4) an enhancement of PHASE1 with a technique given in Sec. 4.4.2 denoted
PROPOSED.
The number of peers N is selected from [100,1600], and peers in the upper layer
are selected through random sampling with probability 1 − p, so that a larger p
indicates fewer peers in the upper layer. The reader should note that such a random
sampling would be enough for evaluating the best performance of HDOV since we
are assuming that the location of each peer is randomly given.

5.2. Results
At first, as a baseline of the discussion, we measure the average hop count of SIMPLE.
Figure 4 summarizes the results. The average hop count monotonically increases as

1950010-12
December 10, 2019 17:31 JOIN S0219265919500105 page 13

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

20

18

16 SIMPLE
Average number of hops

14

12

10

8
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

0
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

100 300 500 700 900 1100 1300 1500


Number of nodes

Fig. 4. The average hop count of SIMPLE.


N increases, which is almost proportional to N , and takes about 14 hops when
N = 1000, indicating that there remains enough room for the improvement.
Next we compare the performance of PHASE1 with SIMPLE and HDOV. Figure 5
summarizes the results, where the vertical axis is the hop count normalized by the
hop count of SIMPLE. Figure 5(a) and (b) show the impact of parameters N and
p to the normalized performance of each scheme, respectively. By Fig. 5(a), we can
find that although HDOV beats PHASE1 for small N ’s less than 300, when N ≥ 400,
PHASE1 significantly improves the performance of other schemes; for example, when
N = 1500, HDOV reduces the hop count of SIMPLE to 80%, which is further reduces
to 60% by applying PHASE1.
The above effect is apparently due to the reduction of the hop count taken during
Phase 1; i.e., the reduction of the length of Phase 1. To verify this, we compare the
length of Phase1 taken by each scheme, and found that although the length of Phase
1 in HDOV monotonically increases as N increases, that of PHASE1 is bounded by
a constant independent of N , which gradually increases as p increases, while the
constant grows more slowly than HDOV for any fixed N .
Finally we compare the performance of PROPOSED with PHASE1. Figure 6 sum-
marizes the results when we use random function h1 in the middle layer. Although
we tried several candidates for the hash function in the experiments (see Sec. 4.3
for the details), we could not find significant difference between them, which is
probably because the (original) location of peers is randomly given in our setting.
A theoretical analysis of the class of hash functions which maximizes the routing
performance for a given distribution of the location of peers, is left as an interesting
open problem. As the figure shows, the introduction of the middle layer with a hash

1950010-13
December 10, 2019 17:31 JOIN S0219265919500105 page 14

S. Fujita

P=0.95
1.4
Average # of hops normalized by simple Delaunay
HDOV
1.2
PHASE1
1

0.8
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

0.6

0.4

0.2
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

0
100 300 500 700 900 1100 1300 1500
Number of node
(a) When p = 0.95.

500 nodes
0.9

HDOV
Average # of hops normalized by simple Delaunay

0.85 PHASE1

0.8

0.75

0.7

0.65
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
probability p

(b) When N = 500.


Fig. 5. Reduction of the average hop count of SIMPLE by HDOV and PHASE1.

function significantly reduces the hop count especially for large N ’s; for example,
when N = 1000, PROPOSED reduces the average hop count of SIMPLE to 30%,
indicating that we could realize a four or five step routing in the resulting overlay,
on average.

1950010-14
December 10, 2019 17:31 JOIN S0219265919500105 page 15

Three-Tier Delaunay Network as a Topology for Peer-to-Peer Overlay

P=0.95
1.4
Average # of hops normalized by simple Delaunay
PHASE1
1.2
PROPOSED
1

0.8

0.6
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

0.4

0.2

0
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

100 300 500 700 900 1100 1300 1500


Number of nodes

Fig. 6. Reduction of the average hop count of SIMPLE by PROPOSED.

6. Concluding Remarks
This paper proposes a hierarchical Delaunay network for peer-to-peer overlays. The
key idea of the proposed method is to associate each peer with two addresses, one is
the given coordinate in the original Delaunay network and the other is the address
obtained by applying an appropriate hash function. We propose a routing scheme
for the resulting hierarchical network. A future work is to build a concrete network
application using the resulting overlay.

Acknowledgment
This work was partially supported by KAKENHI, the Grant-in-Aid for Scientific
Research (B), Grant Number 16H02807.

References
1. F. Araujo and L. Rodrigues, “GeoPeer: A Location-Aware Peer-to-Peer System,” In
Proc. the 3rd IEEE International Symposium on Network Computing and Applications
(2004), pages 36–46.
2. J. Aspnes and G. Shah, “Skip graphs,” ACM Trans. Algorithms 3(4) (2007).
3. H. M. N. D. Bandara and A. P. Jayasumana, “Community-Based Caching for Enhanced
Lookup Performance in P2P Systems,” IEEE Trans. Parallel and Distributed Systems
24(9) (2013) 1201–1217.
4. R. Baraglia et al., “GoDel: Delaunay overlays in P2P networks via Gossip,” In Proc. the
12th IEEE International Conference on Peer-to-Peer Computing (2012), pages 1–12.
5. S. Fujita, “Hierarchical Delaunay Network for Peer-to-Peer Overlay with Address Hash-
ing,” In Proc. PDCAT 2017, pages 305–312.

1950010-15
December 10, 2019 17:31 JOIN S0219265919500105 page 16

S. Fujita

6. M. Ghaffari et al., “A Dynamic Networking Substrate for Distributed MMOGs,” IEEE


Emerging Topics in Computing 3(2) (2015) 289–302.
7. J. Kleinberg, “The Small-World Phenomenon: An Algorithmic Perspective,” Cornell
Computer Science Technical Report (1999), pages 99–1776.
8. X.-Y. Li, I. Stojmenovic and Y. Wang, “Partial Delaunay Triangulation and Degree
Limited Localized Bluetooth Scatternet Formation,” IEEE Trans. Parallel and Dis-
tributed Systems 15(4) (2004) 350–361.
9. X.-Y. Li et al., “Localized Delaunay Triangulation with Application in Ad Hoc Wireless
Networks,” IEEE Trans. Parallel and Distributed Systems 14(10) (2003) 1035–1047.
by 116.102.32.160 on 04/22/23. Re-use and distribution is strictly not permitted, except for Open Access articles.

10. J. Liebeherr, M. Nahas and W. Si, “Application-layer multicasting with Delaunay


triangulation overlays,” IEEE Journal on Selected Areas in Communications 20(8)
(2002) 1472–1488.
11. A. Montresor and M. Jelasity, “PeerSim: A Scalable P2P Simulater,” In Proc. the 9th
IEEE International Conference on Peer-to-Peer Computing (2009), pages 99–100.
J. Inter. Net. 2019.19. Downloaded from www.worldscientific.com

12. M. Ohnishi, R. Nishide and S. Ueshima, “Incremental Construction of Delaunay Over-


laid Network for Virtual Collaborative Space,” In Proc. the 3rd International Conference
on Creating, Connecting and Collaborating through Computing (2005), pages 75–82.
13. M. Ohnishi et al., “Distributive Generation Algorithm of Long Range Contact for Re-
mote Spatial-data Access on P2P Delaunay Network,” In Proc. the 5th International
Conference on Creating, Connecting and Collaborating through Computing (2007), pages
145–152.
14. S. Ratnasamy et al., “A Scalable Content-Addressable Network,” In Proc. SIGCOMM
(2001), pages 161–172.
15. S. Ratnasamy et al., “GHT: A Geographic Hash Table for Data-Centric Storage,” In
Proc. the 1st ACM International Workshop on Wireless Sensor Networks and Applica-
tions (2002), pages 78–87.
16. Z. Shen and J. Thomas, “Location Triggered Routing Protocol in Mobile Ad Hoc Net-
works,” In Proc. the 3rd IEEE International Symposium on Network Computing and
Applications (2004), pages 261–268.
17. J. Shinomiya et al., “An Examination of Sensor Data Collection Method for Spatial
Interpolation on Hierarchical Delaunay Overlay Network,” In Proc. the 11th Interna-
tional Conference on Mobile Data Management (2010), pages 407–412.
18. S. Tsuboi et al., “Generating Skip Delaunay Network for P2P Geocasting,” In Proc.
the 6th International Conference on Creating, Connecting and Collaborating through
Computing (2008), pages 179–186.
19. D. J. Watts and S. H. Strogatz, “Collective dynamics of ’small-world’ networks,” Nature
393(6684) (1998) 440–442.

1950010-16

You might also like