1036 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 8, NO.

10, OCTOBER 1997

Honeycomb Networks: Topological Properties

and Communication Algorithms
Ivan Stojmenovic

Abstract—The honeycomb mesh, based on hexagonal plane tessellation, is considered as a multiprocessor interconnection
network. A honeycomb mesh network with n nodes has degree 3 and diameter ª 1.63 n - 1, which is 25 percent smaller degree
and 18.5 percent smaller diameter than the mesh-connected computer with approximately the same number of nodes. Vertex and
edge symmetric honeycomb torus network is obtained by adding wraparound edges to the honeycomb mesh. The network cost,
defined as the product of degree and diameter, is better for honeycomb networks than for the two other families based on square
(mesh-connected computers and tori) and triangular (hexagonal meshes and tori) tessellations. A convenient addressing scheme for
nodes is introduced which provides simple computation of shortest paths and the diameter. Simple and optimal (in the number of
required communication steps) routing, broadcasting, and semigroup computation algorithms are developed. The average distance
in honeycomb torus with n nodes is proved to be approximately 0.54 n . In addition to honeycomb meshes bounded by a regular
hexagon, we consider also honeycomb networks with rhombus and rectangle as the bounding polygons.

Index Terms—Computer architecture, interconnection topology, routing, broadcasting, mesh-connected computer, hexagonal
tessellation.

—————————— ✦ ——————————

1 INTRODUCTION

V ARIOUS research and development results on how to

interconnect multiprocessor components have been
reported in literature. Several surveys of parallel computing
ous applications. Examples are cellular phone station
placement, the representation of benzenoid hydrocarbons,
computer graphics, and image processing. Here, we pro-
architectures exist (e.g., [1], [10]). One of most popular ar- pose to study the honeycomb mesh interconnection net-
chitectures is the mesh-connected computer (see Fig. 1), in work (Fig. 3), based on the hexagonal tessellation. By add-
which processors are placed in a square or rectangular grid, ing wraparound edges, we define the corresponding hon-
with each processor being connected by a communication eycomb torus network. We will study the basic topological
link to its neighbors in up to four directions. Tori are properties and communication algorithms for honeycomb
meshes with wraparound connections to achieve vertex and meshes and tori.
edge symmetry. Meshes and tori are among the most fre- A topology is evaluated in terms of a number of pa-
quent multiprocessor networks available today on the rameters. In this paper, we are interested in the symmetry,
market. diameter, degree, bisection width, recursive decomposition,
It is well known that there are three possible tessellations and routing, broadcasting, and semigroup computation
of a plane with regular polygons of the same kind: square algorithms. The remaining aspects of honeycomb meshes
(Fig. 1), triangular (Fig. 2), and hexagonal (Fig. 3), corre- and tori are currently open research problems. Although all
sponding to dividing a plane into regular squares, triangles, characteristics are important, we shall consider the network
and hexagons, respectively. The square tessellation is the cost, defined as the product of the degree and diameter
basis for mesh-connected computers, which are widely (measured with respect to the number of nodes) as the
studied in literature. The triangular tessellation is used to main parameter in our comparison. Table 1 summarizes the
define hexagonal mesh (Fig. 2) multiprocessor, studied in degree, diameter, cost, and bisection widths of the consid-
[3], [5], [6], [7], [8], [9], [12]. Several sources [5], [6], [8] refer ered networks, each with n processors. All data are ap-
to the network as being “honeycomb” architecture. They proximate (coefficients are rounded and constants are not
begin with hexagonal tesselation but use cells (instead of included). The total number of links is approximately de-
vertices) as processors. Ben-Natan and Barak [2] mentioned gree * n/2 and is therefore not included in the table. Table 1
the honeycomb array as a processor network but studied also includes the corresponding data [4], [10] for hypercu-
only the contractions of square, triangular, and hexagonal bic networks.
graphs. Honeycomb and hexagonal meshes and mesh-connected
Hexagonal tessellations were used in literature for vari- computers clearly belong to the same family of networks.
Their comparison in all criteria should give the same as-
———————————————— ymptotic performance. For example, they all have fixed
• The author is with the Computer Science Department, University of Ot- d i
degree and O n diameter. Thus, honeycomb meshes
tawa, Ottawa, Ontario K1N 9B4, Canada. E-mail: [email protected].
http://www.csi.uottawa.ca/~ivan. compare to other networks (e.g., hypercube) asymptotically
Manuscript received 26 Apr. 1995; revised 21 Sept. 1996. as well as mesh-connected computers (for such criteria).
For information on obtaining reprints of this article, please send e-mail to: Hypercubic networks are better than mesh-like ones in
[email protected], and reference IEEECS Log Number 104524.

1045-9219/97/$10.00 © 1997 IEEE

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio
 STOJMENOVIC: HONEYCOMB NETWORKS: TOPOLOGICAL PROPERTIES AND COMMUNICATION ALGORITHMS 1037

broadcasting, and semigroup computation) have simple

and optimal solutions which are faster than on the mesh-
connected computers with the same number of nodes (because
of reduced diameter, despite reducing degree). Further re-
search is needed to answer numerous other questions.
The paper is organized as follows. Section 2 of this paper
describes topological properties of honeycomb meshes and
introduces a convenient addressing scheme for nodes. In
Fig. 1. Square mesh. Section 3, optimal routing, broadcasting, and semigroup
computation algorithms for honeycomb meshes are pre-
sented. Honeycomb torus network is defined in Section 4;
its topological properties and routing/broadcasting algo-
rithms are easily derived as extensions of the corresponding
results for honeycomb meshes. In Section 5, we introduce
honeycomb rhombic and rectanglular (square) meshes and
tori by defining the border of honeycomb network in a dif-
ferent way. Section 6 lists some of the numerous open
problems on the honeycomb networks.

2 TOPOLOGICAL PROPERTIES OF HONEYCOMB

MESHES
Honeycomb meshes can be built from hexagons in various
Fig. 2. Hexagonal mesh. ways. The degree of any such network is three. The sim-
plest way to define them is to consider the portion of the
hexagonal tessellation which is inside a given convex poly-
gon. We shall consider three types of meshes which differ
by their boundary. Honeycomb hexagon mesh (HHM) is
inside a regular hexagon (Fig. 3), honeycomb rhombic mesh
(HRoM) is inside a rhombus (Fig. 7), and honeycomb rec-
tangular mesh (HReM) is inside a rectangle (Fig. 8).
We have chosen to study in detail the honeycomb hex-
agonal meshes; for brevity, the network will be simply
called honeycomb mesh in the sequel. The analysis of the
other two types of networks is quite similar, and their
properties will be outlined in Section 5.
To maximize symmetry, honeycomb (hexagonal) meshes
can be built as follows: One hexagon is a honeycomb mesh
of size one, denoted HM1. The honeycomb mesh HM2 of
size two is obtained by adding six hexagons to the bound-
ary edges of HM1. Inductively, honeycomb mesh HMt of
Fig. 3. Honeycomb mesh. size t is obtained from HMt-1 by adding a layer of hexagons
around the boundary of HMt-1. For instance, Fig. 2 is hon-
eycomb mesh of size three, i.e., HM3. Alternatively, the size
terms of diameter, average distance, and product of di-
t of HMt is determined as the number of hexagons between
ameter and degree. Still, mesh is the only one of two to-
the center and boundary of HMt (inclusive). Two alternate
pologies manufacturers have used. Why? One of the main
types of honeycomb meshes are discussed in Section 5. Sec-
reasons, certainly, is because meshes are planar graphs,
tions 2-4 deal only with just introduced type of honeycomb
therefore, they have easy physical layout.
networks. Proofs of some of results given below can be
One should, therefore, compare honeycomb meshes with found in [11], which is a preliminary version of this paper.
the networks within the family of planar graphs composed The full version of this paper is available from the author.
of regular polygons, and the comparison should involve 2
The number of nodes and links of HMt are 6t and 9t -
2
exact (as opposed to asymptotic) values of the considered 3t, respectively.
parameters. The comparison based on degree, diameter, We shall introduce a convenient coordinate system for
total number of links, cost, and bisection width favor hon- nodes and edges of honeycomb meshes. Let x-, y-, and z-axes
eycomb meshes over mesh-connected computers. Some start at the center of the honeycomb mesh and be parallel to
criteria (e.g., how applications and algorithms are easily three edge directions, respectively (see Fig. 4). The honey-
mapped on the proposed structure as compared to on the comb mesh is a bipartite graph. All nodes can be subdivided
mesh) may not be expressed in analogous way. This paper into two groups, which will be called black and white
shows that three important basic algorithms (routing, nodes, such that any edge joins a black and a white node.

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio
 1038 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 8, NO. 10, OCTOBER 1997

TABLE 1
COMPARISON OF NETWORKS
network degree diameter cost bisection width
mesh-connected computer 4 2 n 8 n n
hexagonal mesh 6 1.16 n 6.93 n 2.31 n
honeycomb mesh 3 1.63 n 4.90 n 0.82 n
torus 4 n 4 n 2 n
hexagonal torus 6 0.58 n 3.46 n 4.61 n
honeycomb torus 3 0.81 n 2.45 n 2.04 n
honeycomb rhombic mesh 3 2.83 n 8.49 n 0.71 n
honeycomb square mesh 3 2 n 6 n 0.5 n
honeycomb rhombic torus 3 1.06 n 3.18 n 1.41 n
honeycomb square torus 3 n 3 n n
2
hypercube log n log n log n n/2
cube-connected cycles 3 O(log n) O(log n) O(n/log n)
butterfly 4 O(log n) O(log n) O(n/log n)
deBruijn 4 O(log n) O(log n) O(n/log n)

Fig. 4 indicates black and white nodes in HM3. Edges going + w = 2 for each white one. Two nodes are connected
from a black node to a white node have vector coordinates by an edge iff they differ in exactly one coordinate. †
+ + +
x = (1, 0, 0), y = (0, 1, 0), or z = (0, 0, 1), while edges going
out of white vertices toward black ones have vector coordi-
- - -
nates x = (-1, 0, 0), y = (0, -1, 0), or z = (0, 0, -1). Note
+ - + - + -
that x and x , y and y , z and z are parallel to x-, y-, and
z-axes, respectively. Once edge vectors are defined, con-
sider what are coordinates of nodes. The z-coordinate of all
nodes reachable from a given node A by traveling along
edges (±1, 0, 0) or (0, ±1, 0) (i.e., parallel to x- or y-axis) is
the same as the z-coordinate of A. All nodes with a fixed
(say) z-coordinate belong to a zigzag chain. The z-
coordinate will change only by following an edge parallel
+ -
to z-axis, i.e., edge z or z . If two central chains are given z-
coordinates 1 and 0, the six zigzag chains with respect to z-
axis are z = -2, z = -1, z = 0, z = 1, z = 2, and z = 3; they are
drawn in bold in Fig. 4. Similar zigzag chains can be de-
fined for x- and y-coordinates. For instance, the coordinates
of node A in Fig. 4 are (1, 2, -1). If each zigzag chain with a
fixed z-coordinate is drawn as a vertical line, the honey-
comb mesh receives the shape of bricks in a wall network Fig. 4. Z-coordinates of nodes.
(Fig. 5). The network drawn in this way clearly shows that
the number of edges parallel to z-axis in a shortest path
between two nodes is greater or equal to the difference in
their z coordinate. It easily follows that the coordinates of
all nodes are integers (x, y, z) such that -t + 1 ≤ x, y, z ≤ t.
Moreover, we can prove the following lemma.
LEMMA 1. Nodes of HMt can be coded by integer triples (u, v, w)
such that - t + 1 ≤ u, v, w ≤ t, and 1 ≤ u + v + w ≤ 2. Two
nodes (u¢, v¢, w¢) and (u¢¢, v¢¢, w¢¢) are connected by an
edge iff |u¢ - u¢¢| + |v¢ - v¢¢| + |w¢ - w¢¢| = 1.
PROOF. Edges going from a black node to a white node join
(u, v, w) to either (u + 1, v, w), (u, v + 1, w), or (u, v,
w + 1) and, thus, the sum u + v + w increases by one.
Analogously, edges joining from a white node to a black
one join (u, v, w) to one of (u - 1, v, w), (u, v - 1, w), or
(u, v, w - 1), thus decreasing the sum u + v + w by
one. Since u + v + w alternatively increases (decreases)
by one for each white (black) node, it obviously fol-
lows that u + v + w = 1 for each black node and u + v Fig. 5. Brick drawing of HM3.

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio
 STOJMENOVIC: HONEYCOMB NETWORKS: TOPOLOGICAL PROPERTIES AND COMMUNICATION ALGORITHMS 1039

Using mathematical induction, we can prove that the edges parallel to z-axis. In short, the routing algorithm
distance between nodes (u¢, v¢, w¢) and (u¢¢, v¢¢, w¢¢) of a checks at each current node which of the edge directions x,
y, or z (in this order) would reduce the distance to the des-
hexagonal mesh is |u¢ - u¢¢| + |v¢ - v¢¢| + |w¢ - w¢¢|, and
tination, and will send the message on that edge. At least
that the diameter of HMt is 4t - 1 (nodes (t, -t + 1, 0) and one of the edge directions would lead to a node closer to
(-t + 1, t, 1) are at distance |2t - 1| + |2t - 1| + 1 = 4t - 1). the destination. The algorithm is simple, and details and
Thus, honeycomb mesh network with n nodes has degree code are omitted. The algorithm is optimal in the number of
three and diameter 4 n6 - 1 ª 1.63 n - 1 and its network communication steps, and asymptotically optimal in the
n
number of computation steps (i.e., has constant number of
cost is 12 6
- 3 = 2 6 n - 3 ª 4.90 n . The number of instructions between any two communication steps). The
2
nodes of a hexagonal mesh of size t is n = 3t - 3t + 1 [3], shortest path between two nodes can be explicitly described
(with constant number of parameters) on the basis of their
while its diameter is 2t - 2. The degree of nodes is six.
2 3
addresses and consists of two zigzag chains.
Therefore, n ª 3t and t ª 3
n . The network cost is there- The broadcasting process should satisfy some desirable
fore ª 6 2e 3
j
n = 4 3 n ª 6.93 n . The diameter of a properties. A node should not receive the message simulta-
3
neously from more than one of its neighbors. Moreover, it
square mesh-connected computer with n nodes is 2 n - 2 is desirable that a node receives the message exactly once
and its degree is four. Thus, the cost of a mesh-connected for the whole duration of the broadcasting process. This is
computer is ª 8 n . achieved by a broadcasting algorithm that follows the same
Therefore, honeycomb meshes have 25 percent smaller paths generated by the corresponding routing scheme. The
degree and 18.5 percent smaller diameter than the mesh- source will send the message to all three neighbors. Each
connected computer with the same number of nodes. This other node will examine the incoming message and for-
is an important advantage of honeycomb meshes over ward it to zero, one, or two of its neighbors (it will never be
square ones. The advantage can be explained as follows. It sent back to the neighbor the message was received from).
is clear from Fig. 5, which also shows that honeycomb mesh The algorithm counts the number of times the message
can be embedded onto the square grid, that the distance used edge parallel to the x-, y-, and z-axes, respectively. The
between two nodes remains virtually the same if a quarter counters count edges in positive directions by positive
of edges is eliminated from the grid. Further, by making a numbers (i.e., by increasing their value by one) and edges
hexagonal rather than square boundary, the boundary is in negative directions (opposite to axis orientations) by
closer to a circular shape, thus reducing its diameter. negative numbers (i.e., by decreasing their value), and do
Let us compare the network costs of the networks based not allow forwarding messages in opposite directions. The
on regular plane tessellations. The hexagonal mesh has ap- number of communication steps between a source and any
proximately 13 percent smaller cost than a mesh-connected node is equal to the distance between them. The details of
computer with the same number of nodes. Honeycomb the algorithm are omitted. In the one-port communication
mesh, however, has smaller cost than either of these two model (in which nodes can forward the message to only
networks that are used in practice for designing multiproc- one neighbor at a time), the number of communication
essors. The cost of a honeycomb network is approximately steps necessary for the message to reach each particular
29 percent smaller than the cost of a hexagonal mesh and destination is equal to the length of the shortest path be-
about 39 percent smaller than the cost of a mesh-connected tween the source and a given node plus the delay that cor-
computer. The latter is a significant reduction in the cost responds to messages being sent on other edges (leading to
compared to (arguably) the most popular network on the other destinations) of intermediate nodes before it being
market. sent on the edge of the shortest path. We can show that the
delay is, at most, three communication steps. If all port
communication model (in which each node is allowed to
3 ROUTING, BROADCASTING, AND SEMIGROUP
send the same message to all or some of its neighbors at
COMPUTATION IN HONEYCOMB MESHES once) is used, then the number of communication steps
The problem of finding a path from a source to destination between the source and any node is equal to the distance
and forwarding a message along the path is known as the between them.
routing problem. The broadcasting task is to send a mes- The semigroup computation problem is to compute x1 ©
sage from a source to all other nodes. Routing and broad- x2 © ... © xn, where © is any binary associative and com-
casting are the basic data communication problems for mutative operation (e.g., the sum, product, minimum,
every network. In this section, we will describe routing and maximum, or, and, etc.) and the data are distributed one
broadcasting algorithms for the honeycomb mesh net- per processor on a given network. The problem can be
works. solved by running the broadcasting algorithm from any
Suppose that source = (u¢, v¢, w¢) is the source node node in the reverse order, such that data are being collected
while the destination node is node dest = (u¢¢, v¢¢, w¢¢). Let (each node on the path applies “©” to reduce the informa-
Dx = u¢¢ - u¢, Dy = v¢¢ - v¢, and Dz = w¢¢ - w¢. The vector tion forwarded to the next node to a single message).
(Dx, Dy, Dz) is the “translation” vector for the message. The Broadcasting can be then applied to send the result to other
shortest path between the two nodes consists of |Dx| edges nodes. The performance is the same as that of the broad-
parallel to x-axis, |Dy| edges parallel to y-axis, and |Dz| casting algorithm.

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio
 1040 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 8, NO. 10, OCTOBER 1997

4 HONEYCOMB TORUS NETWORK

Honeycomb torus network can be obtained by joining pairs
of nodes of degree two (i.e., their unused ports) of the hon-
eycomb mesh. In order to achieve edge and vertex symme-
try, the best choice for wrapping around seems to be the
pairs of nodes that are mirror symmetric with respect to
three lines, passing through the center of hexagonal mesh,
and normal to each of three edge orientations. Fig. 6 shows
how to wrap around honeycomb mesh of size three (HM3)
to obtain HT3, the honeycomb torus of dimension three.
The three lines of mirror symmetry for wraparound edges
are shown in dashed lines. The same edge is given the same
number. Wraparound edges are actually drawn in Fig. 6;
for clarity, they will not be drawn in subsequent figures.
For example, edges 4ed5gf in Fig. 6 make a hexagon, and
edges 7a1b6c also make a hexagon. Thus, by wrapping
around as indicated, we receive new hexagons. In general,
the nodes that are joined by an edge are the following pairs:
(-t + 1, t + 1 - p, p) and (t, -p + 1, p - t), (p, -t + 1, t + 1 - p)
and (p - t, t, -p + 1), (t + 1 - p, p, -t + 1) and (-p + 1, p - t, t)
for 1 £ p £ t. Note that vertices are joined are on the bound-
ary and the difference between them is either (2t - 1, -t, -t), Fig. 6. Honeycomb torus of size three.
(-t, 2t - 1, -t), or (-t, -t, 2t - 1), depending on the direction
of their missing edge.
One can observe that it is possible to move from chain
z = 0 to chain z = t, using edge 1 (see Fig. 6), without mak-
ing any step in the z direction. Thus the chains z = p and
z = p - t for each p, 1 £ p £ t, can be considered as one
chain. Each pair of chains contains exactly 6t nodes, i.e., the
same number of nodes.
2
The number of links in HTt is 9t . The diameter of hon-
eycomb torus HTt is 2t. Therefore, the diameter of honey-
comb tori is approximately twice as small as the diameter of
the corresponding honeycomb meshes. This is the same
ratio as in the case of tori that expand mesh-connected
computers and hexagonal meshes. Therefore, the cost of all
three tori is two times smaller than the cost of the corre-
sponding meshes, and the comparison of costs remains the
same as given earlier for meshes.
Let us compare now the bisection widths (the minimum
numbers of links that, when cut, separate the network into
two parts with equal number of processors) of the six net-
works under consideration. The bisection widths of mesh-
connected computer, torus, hexagonal mesh, hexagonal Fig. 7. Honeycomb rhombic mesh.
torus, honeycomb mesh, and honeycomb torus with n
nodes are approximately n , 2 n , 2.31 n , 4.61 n , Therefore, any hexagon can be considered as the center of
0.82 n , and 2.04 n , respectively. We can observe that the the honeycomb torus.
bisection width of honeycomb torus is higher that the bi- It can be proven that the distance between two nodes
section width of (square) torus with the same number of (u¢, v¢, w¢) and (u¢¢, v¢¢, w¢¢) of honeycomb torus of size t is
nodes (in spite of 25 percent reduction in the number of min(|u¢¢ - u¢| + |v¢¢ - v¢| + |w¢¢ - w¢|, |u¢¢ - u¢ + 2t| + |v¢¢
edges). On the other hand, hexagonal meshes and tori have - v¢ - t| + |w¢¢ - w¢ - t|, |u¢¢ - u¢ - 2t| + |v¢¢ - v¢ + t| +
better bisection widths than honeycomb ones. |w¢¢ - w¢ + t|, |u¢¢ - u¢ - t| + |v¢¢ - v¢ + 2t| + |w¢¢ - w¢ -
Suppose that HTt is drawn on a (larger) hexagonal grid t|, |u¢¢ - u¢ + t| + |v¢¢ - v¢ - 2t| + |w¢¢ - w¢ + t|, |u¢¢ - u¢ -
such that there are multiple copies of the same node. Con- t| + |v¢¢ - v¢ - t| + |w¢¢ - w¢ + 2t|, |u¢¢ - u¢ + t| + |v¢¢ - v¢
sider the addresses of these multiple copies. The difference + t| + |w¢¢ - w¢ - 2t|).
of their addresses is one of vectors (2t, -t, -t), (-t, 2t, -t), A simple routing algorithm can be described as fol-
(-t, -t, 2t), (-2t, t, t), (t, -2t, t), or (t, t, -2t). The edge and lows: Find the distances from the source to multiple cop-
vertex symmetry of honeycomb tori can be shown by using ies of the destination, choose the closest copy, and send
multiple copies of the same node on the hexagonal grid. the message using the routing algorithm for the mesh.

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio
 STOJMENOVIC: HONEYCOMB NETWORKS: TOPOLOGICAL PROPERTIES AND COMMUNICATION ALGORITHMS 1041

Because of symmetry, the source can be considered as the

center of hexagonal torus (and the destination address can
be modified accordingly). The broadcasting algorithm for
honeycomb tori resembles the broadcasting algorithm for
honeycomb meshes. In fact, it is the same procedure in
which different boundary conditions are used.
THEOREM 1. The average distance between any two nodes in HTt
is
4 1 2
t- ª 6 n ª 0.54 n .
3 6t 9
PROOF. Because of symmetry, it suffices to determine the
average distance from a source S to all other nodes.
There are three nodes at distance 1 from S, six nodes
at distance 2, nine nodes at distance 3, ..., 3(2t - 1)
nodes at distance 2t - 1, and 3(t - 1) + 2 = 3t - 1 nodes
Fig. 8. Honeycomb rectangular mesh.
at distance 2t. The sum of all distances from S to other
nodes is
brick drawing of HSMt is, in fact, the mesh-connected com-
a f a f a f
0 ◊ 1 + 1 ◊ 3 + 2 ◊ 6 + 3 ◊ 9 + K + 2t - 1 ◊ 3 2t - 1 + 2t ◊ 3t - 1 = 2
puter with n = t nodes, in which some nodes are deleted
3e1 + 2 + 3 + K + a 2t - -1f j + 6t - 2t =
2 2 2 2 2 such that each node has at most three, instead of four,
neighbors. Thus, HSMs can be embedded into mesh-
1 connected computer with the same number of nodes. Inter-
3 a 2t - 1fa2t fa 4t - 1f + 6t - 2t = 8t - t.
2 3
6 estingly, the diameter of the mesh remains unchanged with
The average distance is 8 t 2- t = 43 t - 61t . Note that the this reduction in degree. Thus, 25 percent reduction in de-
3

6t gree is achieved for HSMs keeping the same diameter.

average distance in honeycomb tori is almost the HSMs and HRoMs can be recursively decomposed in the
same as the average distance in square tori (ª 0.5 n ).† same way as mesh-connected computers. For example,
dashed lines in Fig. 8 subdivide HRoM(6, 6) into four
5 HONEYCOMB RHOMBIC AND RECTANGULAR HRoM(3, 3) networks. Fig. 9 shows how to add wrap-
around edges to HRoM(6, 6). It can be easily generalized
MESHES AND TORI and applied to HRoMs of any size. Fig. 10 shows how to
We shall adopt the number of nodes on each line parallel to add wraparound edges to HReM(6, 6). The generalization
the sides of the boundary rhombus (rectangle, respectively) to any size HReM is trivial. The obtained tori will be called
and passing through nodes as the measure of the size of honeycomb rhombic tori (HRoT) and honeycomb rectan-
HRoMs and HReMs. HRoM of size t, abbreviated HRoMt, gular tori (HReT), respectively. Edge and vertex symme-
has t nodes on each such line; for instance, Fig. 7 shows try of HRoTs and HReTs can be easily shown. The num-
HRoM6. The definition can be easily extended to honey- ber of links in HRoT and HReT(t¢, t¢¢) is 3t2, and 3t¢t¢¢, re-
t
comb parallelogramic meshes if the two lines are allowed a
distinct number of nodes. HReMs have two parameters, the spectively. The diameter of HRoTt and HSTt is 23 t and t,
“lengths” of two rectangle sides. For instance, Fig. 8 shows respectively.
HReM(6, 6). For simplicity and other reasons (e.g., torus
expansion), we will consider only the case when the num-
6 CONCLUSION
ber of nodes is the same for both families of lines that are
parallel to a given side of rhombus or rectangle; other cases There are a number of problems on honeycomb meshes and
can be analyzed in an analogous way. We also assume that tori that remain for further investigation. In fault tolerant
t¢ is even in HReM(t¢, t¢¢). applications it is important to construct the maximal possi-
This section will summarize the topological properties of ble number of edge-disjoint paths between any two nodes.
HRoMs (HRoTs) and HReMs (HReTs). Proofs will be In the case of honeycomb networks, the node degree three
omitted. Their degrees, diameters, costs, and bisection limits the number of such paths to three. It is not difficult to
widths are given in Table 1. observe that three disjoint paths between any two nodes (of
2
The number of links of HRoMt and HReM(t¢, t¢¢) are 3t - degree three) always exist. The length of the longest of
2t and 3t¢t¢¢ - t¢ - t¢¢ (respectively). The diameter of HRoMt these paths should be minimized, which is less trivial. The
is 4t - 3. The diameter of HReM(t¢, t¢¢) is 2t¢¢ + t¢ - 2 for 2t¢¢ ≥ t¢ embeddings of rings (i.e., the construction of Hamiltonian
and 2t¢ - 2, otherwise. Thus, the smallest diameter of HReM paths), trees, hypercubes, star, and other networks into
with n = 2t¢t¢¢ nodes is achieved for 2t¢¢ = t¢, and is equal to honeycomb meshes and tori (and vice versa) shall be stud-
the diameter of the mesh-connected computer with n nodes. ied (few straightforward results in this direction are
This HReM will be referred to as the honeycomb square known). Fig. 5 suggests the study of a possible simulation
mesh of size t and is defined as HSMt = HReM(t, t/2). The of algorithms devised for mesh-connected computers to run

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio
 1042 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 8, NO. 10, OCTOBER 1997

REFERENCES
[1] S.G. Akl, Parallel Computation: Models and Methods. Prentice Hall,
1996.
[2] R. Ben-Natan and A. Barak, “Parallel Contractions of Grids for Task
Assignment to Processor Networks,” Networks, vol. 22, pp. 539-562,
1992.
[3] M.S. Chen, K.G. Shin, and D.D. Kandlur, “Addressing, Routing,
and Broadcasting in Hexagonal Mesh Multiprocessors,” IEEE
Trans. Computers, vol. 39, no. 1, pp. 10-18, Jan. 1990.
[4] F.T. Leighton, Introduction to Parallel Algorithms and Architectures:
Arrays, Trees, and Hypercubes. Morgan Kaufmann, 1992.
[5] D. Milutinovic, V. Milutinovic, and B. Soucek, “The Honeycomb
Architecture,” Computer, vol. 20, no. 4, pp. 81-83, Apr. 1987.
[6] V. Milutinovic, “Mapping of Neural Networks on the Honey-
comb Architecture,” Proc. IEEE, vol. 77, no. 12, pp. 1,875-1,878,
1989.
[7] G. Rote, “On the Connection Between Hexagonal and Unidirec-
tional Rectangular Systolic Arrays,” Lecture Notes in Computer Sci-
ence, vol. 227, pp. 70-83. Springer-Verlag, 1986.
[8] B. Robic and J. Silc, “High-Performance Computing on a Honey-
comb Architecture,” Proc. Second Int’l ACPC Parallel Computation
Conf., pp. 22-28. Austria: Springer-Verlag, 1993.
[9] K.G. Shin, “HARTS: A Distributed Real-Time Architecture,” Com-
Fig. 9. Honeycomb rhombic torus. puter, vol. 24, no. 5, pp. 25-35, May 1991.
[10] I. Stojmenovic, “Direct Interconnection Networks,” Parallel and
Distributed Computing Handbook, A.Y. Zomaya, ed., pp. 537-567.
McGraw-Hill, 1996.
[11] I. Stojmenovic, “Honeycomb Networks,” Proc. Math. Foundations
of Computer Science MFCS ’95, Lecture Notes in Computer Science,
vol. 969, pp. 267-276. Prague: Springer-Verlag, 1995.
[12] H.Y. Youn and J.Y. Lee, “An Efficient Dictionary Machine Using
Hexagonal Processor Arrays,” IEEE Trans. Parallel and Distributed
Systems, vol. 7, no. 3, pp. 266-273, Mar. 1996.

Ivan Stojmenovic received the BS and MS

degrees in 1979 and 1983, respectively, from
the University of Novi Sad and PhD degree in
mathematics in 1985 from the University of
Zagreb. In 1980, he joined the Institute of
Mathematics, University of Novi Sad. During
the winter of 1985-1986, he was a visiting
researcher at the Electrotechnical Laboratory,
Tsukuba, Japan. In the fall of 1987 and spring
Fig. 10. Honeycomb rectangular torus. of 1988, he was a visiting assistant professor
in the Computer Science Department, Wash-
ington State University, Pullman, Washington, and in the Department
on honeycomb meshes. The existence of a deadlock-free of Mathematics and Computer Science, University of Miami, Miami,
routing procedure shall be searched. The basic data com- Florida, respectively. In the fall of 1988, he joined the faculty of the
munication techniques (parallel prefix, sorting, merging, Computer Science Department at the University of Ottawa, Ottawa,
Canada, where he currently holds the position of a full professor. He
personalized communication, etc.) need to be developed. has published three books and more than 130 different papers in jour-
The network scalability seems to be a problem, especially nals and conferences. His research interests are parallel computing,
for honeycomb tori (it faces similar scalability problems as multiple-valued logic, combinatorial algorithms, neural networks, com-
the popular torus network). putational geometry, and graph theory. He is currently managing editor
of Multiple-Valued Logic, an international journal, and an editor of the
The properties of the direct products of honeycomb following journals: Parallel Processing Letters, Parallel Algorithms and
meshes with itself or other types of meshes may be investi- Applications, IASTED International Journal of Parallel and Distributed
gated to find new attractive networks. It might be of inter- Systems, Journal of Computing and Information, and Tangenta.
est to consider the square grid network with a quarter of
edges eliminated to obtain a honeycomb-like network. Fi-
nally, the generalization(s) of honeycomb meshes to three
and more dimensions should be defined and studied.

ACKNOWLEDGMENTS
The author is grateful to the referees whose careful reading
has greatly improved the clarity and conciseness of the
presentation, and to Professor Michael A. Palis for his
timely and professional handling of my submission. This
research is partially supported by NSERC.

d licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on September 12,2022 at 07:50:44 UTC from IEEE Xplore. Restrictio