Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Journal of Systems Architecture 50 (2004) 697–705

www.elsevier.com/locate/sysarc

Optical transpose k-ary n-cube networks


Khaled Day
Department of Computer Science, Sultan Qaboos University, P.O. Box 36, Al-Khod 123, Muscat, Oman

Received 1 January 2003; received in revised form 19 April 2004; accepted 20 May 2004
Available online 23 July 2004

Abstract

This paper derives a number of results related to the topological properties of OTIS k-ary n-cube interconnection
networks. The basic topological metrics of size, degree, shortest distance, and diameter are obtained. Then results re-
lated to embedding in OTIS k-ary n-cubes of OTIS k-ary (n1)-cubes, cycles, meshes, cubes, and spanning trees are
derived. The OTIS k-ary n-cube is shown to be Hamiltonian. Minimal one-to-one routing and optimal broadcasting
algorithms are proposed. The OTIS k-ary n-cube is shown to be maximally fault-tolerant. These results are derived
based on known properties of k-ary n-cube networks and general properties of OTIS networks.
 2004 Elsevier B.V. All rights reserved.

Keywords: Interconnection networks; k-ary n-cube; Optical transpose interconnection systems; Topological properties

1. Introduction ary n-cubes are the hypercube (where k = 2) and


the torus (where n = 2 or 3). The former has been
The k-ary n-cube has been one of the most com- used in early multicomputers such as the iPSC/2
mon networks for multicomputers due to its desir- [13] and iPSC/860 [20] while the latter has been
able properties, such as ease of implementation adopted in recent systems like the J-Machine
and ability to reduce message latency by exploiting [12], CRAY T3D [9] and CRAY T3E [1].
communication locality found in many parallel Bandwidth limitations imposed by electronic
applications. The k-ary n-cube possesses an n- interconnects such as k-ary n-cubes interconnects
dimensional grid structure with k nodes in each have prompted the need for exploring alternatives
dimension such that every node is connected to that can overcome these limitations. Optics is one
its neighboring nodes in each dimension by direct such alternative that is capable of providing inher-
channels. The two most popular instances of k- ent communication parallelism, high interconnec-
tivity, and large bandwidth.
Optical networks have moved from being a re-
search curiosity to a billion-dollar business, as
E-mail address: kday@computer.org can be seen from several commercial activities in

1383-7621/$ - see front matter  2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.sysarc.2004.05.002
698 K. Day / Journal of Systems Architecture 50 (2004) 697–705

different parts of the world [3,16]. Optical technol- The OTIS architecture has gained considerable
ogies can provide thousands of high bandwidth attention recently. A number of algorithms have
channels at a single communication point (chip, been developed such as BPC permutations [17],
board, or computer). However, electronic inter- routing, selection and sorting [14,15], data rear-
connects perform better when the distance is up rangement [25], matrix multiplication [22], and
to few millimeters [8]. Motivated by these observa- image processing [21]. Several general topological
tions, new hybrid computer architectures utilizing properties of OTIS networks are obtained in [7].
both optical and electronic technologies have been This paper presents a further contribution in
proposed and investigated [11,18,19]. In these sys- this direction by studying the properties of k-ary
tems short interconnects at the chip level are elec- n-cube networks including basic topological prop-
tronic; while optical interconnects are used for erties, embedding, routing, broadcasting, and
larger distances. fault-tolerance. These results are derived based
Optical Transpose Interconnection Systems on known properties of k-ary n-cube networks
(OTIS) present a new optoelectronic computer [2,6] and general properties of OTIS networks [7].
architecture that takes benefits from both optical
and electronic technologies. In this architecture,
processors are divided into groups where elec- 2. Definition and basic properties
tronic interconnects are used to connect processors
within the same group; while optical interconnects Definition 1. The k-ary n-cube Qkn has N = kn nodes
are used for inter-group communication. The each of the form X = xn1xn2. . .x0, where 0 6
OTIS architecture was used in [15,26] to propose xi < k, for all 0 6 i < n. Two nodes X = xn1xn2. . .x0
interconnection networks for multiprocessor sys- and Y = yn1yn2. . .y0 in Qkn are connected if and
tems, by replacing a subset of wires with pairs of only if there exists i, 0 6 i < n, such that xi = yi ± 1
transmitters and receivers. The result is an interest- (mod k) and xj = yj, for i 6¼ j. For clarity we omit to
ing trade-off between electrical and optical inter- write mod k in similar expressions for the remainder
connects in terms of speed, power consumption of the paper.
and space reduction [8]. It has been shown in [10]
that the bandwidth and the power consumption It is shown in [6] that Qkn has degree 2n and
are optimized when the number of processors in diameter n ºk/2ß. Given two nodes X = xn1xn2
a group equals the number of groups. Several stud- . . .x0 and Y = yn1yn2. . .y0 in Qkn , we denote by
ies of OTIS architecture have then been limited to dH (X,Y) the Hamming distance between X and
the case where the group size is equal to the num- Y, i.e., the number of digits in which X and Y dif-
ber of groups in the system [15,17,23,24,26]. In fer. The length of a shortest path between X and Y
such an OTIS system, an inter-group (optical) link is
Pequal to the Lee distance [2] given by: d L ðX ; Y Þ ¼
n1
connects processor p of group g to processor g of i¼0 w i , where wi = min0 6 i < n(jxiyij, kjxiyij).
group p. The intra-group (electronic) links form
an interconnection topology corresponding to a Definition 2. The OTIS-Qkn network is an undi-
k-ary n-cube in the OTIS k-ary n-cube. rected graph (V, E) given by
More recently, Coudert et al. have studied a
more general OTIS architecture where the group V ¼ fhg; pijg; p 2 Qkn g
size is not necessarily equal to the number of
groups. In [4], authors study the relationship be- and
tween the OTIS architecture and the family of de E ¼ fðhg; p1 i; hg; p2 iÞjðp1 ; p2 Þ is an edge in Qkn g
Bruijn digraphs which leads to minimize the total
[ fðhg; pi; hp; giÞjg 6¼ p in Qkn g:
number of lenses of the optical layout. In [5],
authors have proposed an optical layout of several The OTIS-Qkn is composed of N = kn node-disjoint
kðN Þ
graph topologies (including de Bruijn and Kautz sub-graphs Qkð1Þ kð2Þ
n ; Qn ; . . . ; Qn , called groups.
digraphs) using OTIS. Each of these groups is isomorphic to a k-ary n-
K. Day / Journal of Systems Architecture 50 (2004) 697–705 699

cube Qkn . A node Æg, pæ in OTIS-Qkn corresponds to 3. distOTISQkn ðhg1 ; p1 i; hg2 ; p2 iÞ


the node of address p in group QnkðgÞ . We refer to g 8
> distQkn ðp1 ; p2 Þ; if g1 ¼ g2
as the group address of node Æg, pæ and to p as its >
>
< minfdist k ðp ; p Þ þ dist k ðg ; g Þ þ 2;
Qn 1 2 Qn 1 2
processor address. An intra-group edge of the ¼
form (Æg, p1æ, Æg, p2æ) corresponds to an electronic >
> dist k ðp 1 ; g 2 Þ þ dist k ðg 1 ; p 2 Þ þ 1g
>
:
Q n Qn

link. An inter-group edge of the form (Æg, pæ,Æp, gæ) if g1 6¼ g2


corresponds to an optical link. Fig. 1 illustrates the 4. diamðOTIS  Qkn Þ ¼ 2nbk=2c þ 1
OTIS-Q32 network.
The following notations will be used in the
paper. 3. Embeddings

jGj size of a graph G, i.e., the number of We start by showing how to embed disjoint
nodes of G OTIS-Qkn1 sub-graphs in OTIS-Qkn . An example
degG (p) degree of a node p in a graph G showing how three disjoint OTIS-Q31 Õs are embed-
distG (p1, p2) the length of a shortest path between ded in OTIS-Q32 is given in Fig. 2.
the nodes p1 and p2 in G
diam(G) diameter of a graph G, i.e. the maxi- Theorem 1. OTIS-Qkn embeds k disjoint OTIS-Qkn1
mum distG (p1, p2) for all p1 and p2 in G sub-graphs.

The following basic topological metrics of Proof. Qkn can be partitioned into k disjoint Qkn1
OTIS-Qkn are derived from Definitions 1 and 2 sub-graphs. This can be seen by considering the
and from similar results in the literature [17,24]: sub-graph Qk;i k k
n1 of Qn containing all nodes in Qn
k;i
of the form X = ixn2. . .x1x0. Clearly each Qn1 is
1. j OTIS  Qkn j¼ k 2n  isomorphic to Qkn1 . Furthermore, it is shown in
2n; if g ¼ p [7] that if G embeds k disjoint sub-graphs each iso-
2. degOTISQkn ðhg; piÞ ¼
2n þ 1; if g 6¼ p morphic to a graph H, then OTIS-G embeds k dis-
joint sub-graphs each isomorphic to OTIS-H.
Hence OTIS-Qkn embeds k disjoint copies of
OTIS-Qkn1 . h
00 01 02 00 01 02 00 01 02
Now, we present some additional results related
10 11 12 10 11 12 10 11 12 to the embedding of other structures in OTIS-Qkn
such as cycles, meshes, binary cubes, and spanning
20 21 22 20 21 22 20 21 22 trees.
00 01 02
00 01 02 00 01 02 00 01 02 Theorem 2. For any i, 0 6 i 6 n1, OTIS-Qkn
10 11 12 10 11 12 10 11 12
embeds kn + i disjoint cycles each of length kni.

20 21 22 20 21 22 20 21 22 Proof. OTIS-Qkn can be partitioned into kn disjoint


10 11 12
Qkn sub-graphs (groups). Qkn can be partitioned into
00 01 02 00 01 02 00 01 02 k disjoint Qkn1 sub-graphs. This latter partitioning
can be applied recursively i times, 0 6 i 6 n1,
10 11 12 10 11 12 10 11 12
yielding a partitioning of Qkn into ki disjoint Qkni
20 21 22 20 21 22 20 21 22 sub-graphs. Since Qkni is Hamiltonian [2], this im-
plies that Qkn can be partitioned into ki disjoint cy-
20 21 22
cles each of length kni and hence OTIS-Qkn can be
Fig. 1. The OTIS-Q32 network (for clarity only optical links partitioned into kn + i disjoint cycles each of length
connected to group 00 are shown). kni. h
700 K. Day / Journal of Systems Architecture 50 (2004) 697–705

(a) 00 01 02 The next result is related to the embedding of


hypercubes in OTIS-Qkn . It is shown in [2] that Q4n
embeds a 2n-dimensional hypercube with dilation
10 11 12 1. It is shown in [7] that if G embeds an n-cube with
dilation d, then OTIS-G embeds a 2n-cube with
dilation d + 2. Combining these two results yields
20 21 22 the following result.

Theorem 4. OTIS-Q4n embeds a 4n-cube with


(b) dilation 3.

Now we show how to construct a spanning tree


in an OTIS network. Spanning trees are useful for
designing group communication and for simulat-
ing tree-based algorithms. It is shown in [7] that
00 01 02 if there exists in G a spanning tree of height h
rooted at node X, then there exists in OTIS-G a
spanning tree of height 2h + 1 rooted at ÆX, Xæ.
Furthermore, it is shown in [2] how to construct
a spanning tree in Qkn of height n ºk/2ß. We make
use of these results and show how to construct in
10 11 12 OTIS-Qkn a spanning tree of optimal height 2n ºk/
2ß + 1.

Theorem 5. For any node address X in Qkn , OTIS-


Qkn embeds a spanning tree rooted at node ÆX, Xæ of
optimal height 2n ºk/2ß + 1.
20 21 22

Fig. 2. (a) Q32 embeds 3 disjoint copies of Q31 (3 cycles of 3 nodes The spanning tree construction is illustrated on
each). (b) OTIS-Q32 embeds 3 disjoint copies of OTIS-Q31 (for OTIS-Q32 . First, an overview of the spanning tree
clarity only the optical links appearing in the OTIS-Q31 sub- construction for Qkn described in [2] is given. Start-
graphs are drawn). ing with node 00. . .0 as a root, each node
X = xn1xn2. . .x0 has two children for each lead-
ing 0 digit in the address. These two children are
Theorem 3. For any i, 0 6 i 6 n1, OTIS-Qkn em- obtained by keeping all digits unchanged in the ad-
beds, with dilation 3, ki node-disjoint kni · kni dress except for one of the leading zeros which is
wrap-around meshes (tori). replaced by 1 in one child and by k1 in another
child. A node may have one additional child ob-
Proof. We have seen in the proof of theorem 2 tained by incrementing (modulo k) the leftmost
that Qkn can be partitioned into ki disjoint cycles non zero digit (if it exists) if it is less than ºk/2ß
each of length L = kni, for any i, 0 6 i 6 n1. It or by decrementing it if it is larger than ºk/2ß + 1.
is shown in [7] that if G embeds a cycle of length Using this tree as a basis, a spanning tree rooted
L, then OTIS-G embeds an L · L wrap-around at Æ00, 00æ in OTIS-Q32 is obtained as shown in
mesh with dilation 3 with group and processor ad- Fig. 3.
dresses corresponding to node addresses in the cy- The construction of a Hamiltonian cycle in
cle of length L. Therefore OTIS-Qkn embeds, with OTIS-Qkn is addressed next.
dilation 3, ki node-disjoint kni · kni wrap-around
meshes. h Lemma 1. If k is odd then OTIS-Qkn is Hamiltonian.
K. Day / Journal of Systems Architecture 50 (2004) 697–705 701

00 01 02 00 01 02 00 01 02
at start and ending at end. This path is ob-
tained from H by deleting its edge between
10 11 12 10 11 12 10 11 12 the nodes start and end.
(ii) The start node of a group is the group number
20 21 22 20 21 22 20 21 22
of the previous group (cyclically) in the above
00 01 02 sequence and the end node of a group is the
00 01 02 00 01 02 00 01 02
group number of the next group in the se-
10 11 12 10 11 12 10 11 12 quence. Therefore there exists an optical link
in OTIS-Qkn connecting the last node in a
20 21 22 20 21 22 20 21 22 group to the first node of the next group in
10 11 12 the circular list.
00 01 02 00 01 02 00 01 02
We therefore have a sequence of Hamiltonian
10 11 12 10 11 12 10 11 12
paths for the kn groups which can be intercon-
20 21 22 20 21 22 20 21 22 nected by optical links to form a Hamiltonian
cycle in OTIS-Qkn . h
20 21 22

Fig. 3. A Spanning tree (thick lines) rooted at Æ00, 00æ in OTIS- Lemma 2. For n P 2, if k is even then there exists
Q32 . a Hamiltonian cycle H ¼ X 0 ; X 1 ; . . . ; X kn 1 ; X kn ;
2 2
Proof. Since k is odd, kn (size of Qkn ) is also odd. X kn 1 in Qkn such that the two nodes X0 and X kn 1
2
k
Let kn = 2p1 and H = X0, X1, . . .,Xp1, Y0, Y1, . . ., are connected in Qn and the two nodes X kn and
2
Yp2, X0 be a Hamiltonian cycle in Qkn . We con- X kn 1 are also connected in Qkn .
struct a Hamiltonian cycle for OTIS-Qkn by con-
necting Hamiltonian paths in the groups of Proof. (by induction on n) For n = 2, Fig. 4 illus-
OTIS-Qkn (each group is a Qkn ) in the following trates the construction of the desired Hamiltonian
group order: X0, Y0, X1, Y1, . . ., Xp2, Yp2, Xp1. cycle in Q62 . The same dual ÔzigzagÕ construction is
The traversed Hamiltonian path in each of these possible for any even k. Now assume n P 3. Let
groups is obtained by following the Hamiltonian X0, X1, . . ., Xkn11, X0 be a Hamiltonian cycle in
cycle H backward starting at a specific node and Qkn1 such that X0 and X kn 11 are connected and
2
ending at the next node listed in the sequence H also X kn1 and X kn1 1 are connected. Let
n1 2
above. The start node and end node for each of q ¼ k 2  1. Since k is even and n P 3, q must be
the groups X0, Y0, X1, Y1, . . ., Xp2, Yp2, Xp1 in odd and kn11 must also be odd. Consider the
this order are given below in the form [start, end] ordering of the nodes of Qkn shown in Fig. 5.
written under the corresponding group number.

Group number : X0 Y0 X1 Y1 ... X p2 Y p2 X p1


½start; end : ½X p1 ; Y 0  ½X 0 ; X 1  ½Y 0 ; Y 1  ½X 1 ; X 2  . . . ½Y p3 ; Y p2  ½X p2 ; X p1  ½Y p2 ; X 0 

The following can be easily verified about this


The following properties of this ordering can be ordering:
easily verified:
(a) It corresponds to a Hamiltonian cycle in Qkn .
(i) Each of the [start, end] pairs corresponds to a (b) The first node 0X0 of the first half of the list is
pair of adjacent nodes on the Hamiltonian cy- connected to the last node 0Xq of the first half
cle H of Qkn . Therefore there exists a Hamilto- (since X0 and Xq are connected in Qkn1 by
nian path in the corresponding group starting induction hypothesis).
702 K. Day / Journal of Systems Architecture 50 (2004) 697–705

X0 X1 X2 X33 X34 0X0 → 1X0 → … → (k-1)X0


X35 =X k n-1 ↓
0X1 ← 1X1 ← … ← (k-1)X1
X5 X4 X3 X32 X31 X30 ↓
0X2 → 1X2 → … → (k-1)X2
X6 X7 X8 X27 X28 X29 … … …

0Xq ← 1Xq ← … ← (k-1)Xq
X11 X10 X9 X26 X25 X24 ↓
0Xq+1 → 1Xq+1 → … → (k-1)Xq+1
… … …
X12 X13 X14 X21 X22 X23 ↓
0X k n −1−1 ← 1X k n −1 −1 ← … ← (k-1)X k n−1−1

X16 X15 X20 X19


Fig. 5. Recursive Hamiltonian cycle construction in Qkn .
X17 X18
=X k n =X k n
-1
2 2 The following theorem is directly derived from
Lemmas 1–3:
Fig. 4. A Hamiltonian cycle satisfying the conditions of
Lemma 2 for n = 2.
Theorem 6. OTIS-Qkn is Hamiltonian, for any
n P 2 and any k P 2.
(c) The first node 0Xq + 1 of the second half is con-
nected to the last node 0X kn1 1 of the second
half (since Xq + 1 and X kn1 1 are connected in 4. Routing and broadcasting
Qkn1 by induction hypothesis). h
A distributed routing algorithm can be specified
Lemma 3. For n P 2, if k is even then OTIS-Qkn is by a function which, for a given current node and
Hamiltonian. a given destination node, returns the node follow-
ing the current node on a routing path towards the
Proof. H ¼ X 0 ; X 1 ; . . . ; X kn 1 ; X kn ; X kn 1 be a Ham- destination node. Consider the known dimensional
2 2
iltonian cycle in Qkn such that X0 and X kn 1 are routing algorithm in Qkn [2]. The function
2
k
connected in Qn and X kn and X kn 1 are also con- NEXTQkn ðC; DÞ shown below implements this algo-
2
k
nected in Qn (such a Hamiltonian cycle exists by rithm. Assuming a message is currently stored at a
Lemma 2). Similar notations to those used in the node C = Cn1, Cn2, . . ., C0 of Qkn and has to be
proof of Lemma 1 will be used here for construct- routed towards a destination node D = Dn1,
ing a Hamiltonian cycle in OTIS-Qkn as follows: Dn2, . . ., D0 it returns the address of the node fol-

Group number : X0 X kn X1 X kn þ1 X2 ... X kn 1 X kn 1


2 2 2
h i h i h i h i
½start; end : X k n 1 ; X kn ½X 0 ; X 1  X kn ; X kn þ1 ½X 1 ; X 2  X kn þ1 ; X kn þ2 ... ½X kn 2 ; X kn 1  X kn 1 ; X 0
2 2 2 2 2 2

This sequence has the same properties (i) and lowing node C on a minimum routing path to-
(ii) listed in the proof of Lemma 1. We therefore wards node D. It uses a routing vector R = Rn1,
have a sequence of Hamiltonian paths for the kn Rn2, . . ., R0 where Ri encodes both the sign and
groups which can be interconnected by optical magnitude of the minimum routing distance along
links to form a Hamiltonian cycle in OTIS-Qkn . h dimension i.
K. Day / Journal of Systems Architecture 50 (2004) 697–705 703

function NEXT Qkn ðC; DÞ //invoked at node Ægs, pcæ to contribute in broadcast-
//returns the node following C on a minimal routing ing M originating at source Ægs, psæ
path from C to D {
{ if gs6¼pc then initiate BROADCAST Qkn of M in
let R = Rn1Rn2. . .R0 be the routing vector from group pc starting at source gs
C to D for each p in ChildrenQkn (ps, pc) do
let i be the smallest j such that Rj 6¼ 0 invoke BROADCAST OTISQkn (M, Ægs, psæ,
if (i does not exist) then Ægs, pæ)
return null //destination has been reached }
else if (Ri > 0) then
return Cn1Cn2. . .Ci + 1 Ci + 1 Ci1. . .C0 BROADCAST OTISQkn traces the optimal height
else return Cn1Cn2. . .Ci + 1 Ci1Ci1. . .C0 spanning tree constructed in Theorem 5 delivering a
} single message copy to each node in OTIS-Qkn .

A routing function NEXT OTISQkn for the OTIS-Qkn


network can then be obtained as follows: 5. Fault tolerance
function NEXT OTISQkn (Æxc, ycæ, Æxd, ydæ)
//returns the node following Æxc, ycæ on a minimal The node fault tolerance of a network is meas-
routing path from Æxc, ycæ to Æxd, ydæ ured by its node connectivity. A network has
{ node connectivity c if it remains connected in
if (xc = xd and yc = yd) then return null //destina- the presence of less than c faulty nodes. A net-
tion reached work is maximally fault tolerant if its node connec-
if (xc = xd) then return Æxd, NEXT Qkn (yc, yd)æ tivity is equal to the minimum node degree. The
if (distG (yc,yd) + distG (xc,xd) + 2<distG (yc,xd) k-ary n-cube, which is 2n-regular, is known to
+ distG (xc,yd) + 1) then have 2n node-disjoint paths between any two of
if (yc = yd) then return Æyc, xcæ its nodes [2,6]. It is therefore maximally fault-toler-
else return Æxc, NEXT Qkn (yc, yd)æ ant. This section shows that OTIS-Qkn , whose
if (xc = yd) then return Æyc, xcæ minimum node degree is 2n, is maximally fault-
return Æxc, NEXT Qkn (yc,xd)æ tolerant by showing that it contains at least
} 2n node-disjoint paths between any two of its
nodes.
It is shown in [7] that such a routing function The following two results have been proved in
would be optimal (i.e., produces minimum routing [7].
paths) since the routing function NEXT Qkn is opti-
mal in Qkn . Lemma 4. [7] If G is connected and if Æg1, p1æ and
The optimal height spanning tree constructed Æg2, p2æ are two nodes in OTIS-G such that p1 6¼ p2
in Theorem 5 can be used to perform optimal and such that there are d node-disjoint paths between
single node broadcasting in OTIS-Qkn assuming p1 and p2 in G, then there are d node-disjoint paths
multiport I/O. Let BROADCAST Qkn be the broad- between Æg1, p1æ and Æg2, p2æ in OTIS-G.
casting algorithm for Qkn based on the Qkn span-
ning tree described earlier. Let ChildrenQkn (s, c) Lemma 5. [7] If G is connected, g1 6¼ g2 and
denote the set of children of node c in the spanning degG(p) = d, then there are d node-disjoint paths be-
tree of Qkn rooted at node s. An algorithm for tween Æg1, pæ and Æg2, pæ in OTIS-G.
broadcasting a message M originating at a
source node Ægs, psæ in OTIS-Qkn can be obtained These two results along with the know result
as follows. about the existence of 2n node-disjoint paths be-
Algorithm BROADCAST OTISQkn (M, Ægs, psæ, tween any two nodes of the k-ary n-cube [2,6] yield
Ægs, pcæ) the following result.
704 K. Day / Journal of Systems Architecture 50 (2004) 697–705

00 References
01 02 00 01 02 00 01 02

10 11 12 10 11 12 10 11 12 [1] E. Anderson, J. Brooks, C. Grassl, S. Scott, Performance


of the Cray T3E multiprocessor, in: Proceedings of
20 21 22 20 21 22 20 21 22 Supercomputing Conf., San jose, California, 1997.
00 01 02
[2] B. Bose, B. Broeg, Y. Kwon, Y. Ashir, Lee distance and
00 01 02 00 01 02 00 01 02 topological properties of k-ary n-cubes, IEEE Transactions
on Computer 44 (8) (1995) 1021–1030.
10 11 12 10 11 12 10 11 12 [3] S. Chatterjee, S. Pawlowski, All optical networks, Com-
munication of the ACM 42 (6) (1999) 74–83.
20 21 22 20 21 22 20 21 22 [4] D. Coudert, A. Ferreira, S. Perennes, De Bruijn, Isomor-
10 11 12 phisms and free space optical networks, in: IEEE Interna-
00 01 02 00 01 02 00 01 02 tional Parallel and Distributed Processing Symposium,
IPDPS 2000, IEEE Press, 2000, pp. 769–774.
10 11 12 10 11 12 10 11 12 [5] D. Coudert, A. Ferreira, X. Muñoz, Topologies for optical
interconnection networks based on the optical transpose
20 21 22 20 21 22 20 21 22
interconnection system, Applied Optics––IP 39 (17) (2000)
20 21 22 2965–2974.
[6] K. Day, A.E. Al-Ayyoub, Fault diameter of k-ary n-cube
Fig. 6. Four node-disjoint paths between Æ00, 00æ and Æ22, 22æ in networks, IEEE Transactions on Parallel and Distributed
OTIS-Q32 . Systems 8 (9) (1997) 903–907.
[7] K. Day, A. Al-Ayyoub, Topological properties of OTIS-
networks, IEEE Transactions on Parallel and Distributed
Lemma 6. There are at least 2n node-disjoint paths Systems 13 (4) (2002) 359–366.
between any two nodes of OTIS-Qkn . [8] M. Feldman, S. Esener, C. Guest, S. Lee, Comparison
between electronic and free space optical interconnects
based on power and speed consideration, Applied Optics
As an illustration, the node-disjoint path con- 27 (9) (1988) 1742–1751.
struction method of [2] and that of Lemma 5 are [9] R.E. Kessler, J.L. Schwarzmeier, CRAY T3D: A new
combined to obtain the four node-disjoint paths dimension for Cray research, in: CompCon, Houston,
between Æ00, 00æ and Æ22, 22æ in OTIS-Q32 shown Texas, Spring 1993, pp. 176–182.
[10] A. Krishnamoorthy, P. Marchand, F. Kiamilev, S. Esener,
in Fig. 6. The following result is derived from Grain-size considerations for optoelectronic multistage
Lemma 6. interconnection networks, Applied Optics 31 (26) (1992)
5480–5507.
Theorem 7. OTIS-Qkn is maximally fault-tolerant. [11] G. Marsden, P. Marchand, P. Harvey, S. Esener, Optical
transpose interconnection system architectures, Optics
Letters 18 (13) (1993) 1083–1085.
[12] M. Noakes et al., The J-machine multicomputer: an
6. Conclusion architectural evaluation, Proceedings of the 20th Interna-
tional Symposium on Computer Architecture (1993).
[13] S.F. Nugent, The iPSC/2 direct-connect communication
Several topological properties of OTIS k-ary n- technology, Proceedings of 3rd Conference on Hypercube
cube networks have been derived in this paper Concurrent Computers and Applications 1 (1988) 51–60.
including: basic topological metrics of size, degree, [14] A. Osterloh, Sorting on the OTIS-Mesh, Proceedings of the
shortest distance, and diameter; embedding of 14th International Parallel and Distributed Processing
Symposium (IPDPS 2000) (2000) 269–274.
smaller OTIS k-ary (n1)-cubes, cycles, meshes, [15] S. Rajasekaran, S. Sahni, Randomized routing, selection,
cubes, and spanning trees; construction of Hamil- and sorting on the OTIS-Mesh, IEEE Transactions on
tonian cycles; minimal one-to-one routing; optimal Parallel and Distributed Systems 9 (9) (1998) 833–840.
broadcasting; and maximal fault-tolerance. The [16] J. Rayn, WDM: north american development trend, IEEE
obtained results confirm the suitability of OTIS ar- Communications Magazine 32 (2) (1998) 40–44.
[17] S. Sahni, C.-F. Wang, BPC Permutations on the OTIS-
chitectures for multiprocessor interconnection net- Mesh Optoelectronic Computer, Proceedings Fourth Inter-
works offering benefits from both optical and national Conference on Massively Parallel Processing
electronic technologies. Using Optical Interconnections (1997) 130–135.
K. Day / Journal of Systems Architecture 50 (2004) 697–705 705

[18] T. Szymanski, Hypermesh optical interconnection net- [25] C.-F. Wang, Algorithms for the OTIS optoelectronic
works for parallel computing, Journal of Parallel and computer, Ph.D. Thesis, Department of Computer Science,
Distributed Computing 26 (1995) 1–23. University of Florida, 1998.
[19] T. Szymanski, H. Hinton, Architecture of a terabit free- [26] F. Zane, P. Marchand, R. Paturi, S. Esener, Scalable
space intelligent optical backplane, Journal of Parallel and network architectures using the optical transpose intercon-
Distributed Computing 55 (1) (1998) 1–31. nection system (OTIS), Journal of Parallel and Distributed
[20] B. Vanvoorst, S. Seidel, E. Barscz, Workload of an iPSC/ Computing 60 (5) (2000) 521–538.
860, Proceedings of Scalable High-Performance Comput-
ing Conference (1994) 221–228.
[21] C.-F. Wang, S. Sahni, Image processing on the OTIS- Khaled Day received his undergraduate
Mesh optoelectronic computer, IEEE Transactions on degree in computer science from
the University of Tunis in 1986 and the
Parallel and Distributed Systems 11 (2) (2000) 97–109. M.Sc. and Ph.D. degrees from the
[22] C.-F. Wang, S. Sahni, Matrix multiplication on the OTIS- University of Minnesota (USA) in
mesh optoelectronic computer, IEEE Transactions on 1989 and 1992 respectively. Dr. Day is
Computers 50 (7) (2001) 635–646. currently an Associate Professor at the
Department of Computer Science of
[23] C.-F. Wang, S. Sanhi, OTIS optoelectronic computers, in: Sultan Qaboos University in Oman.
K. Li, Y. Pan, S.Q. Zhang (Eds.), Parallel Computation His areas of interest include intercon-
Using Optical Interconnections, Kluwer Academic, 1998, nection networks, parallel algorithms,
pp. 99–116. distributed computing and cluster
computing. He is a senior member of
[24] C.-F. Wang, S. Sahni, Basic operations on the OTIS-mesh the IEEE.
optoelectronic computer, IEEE Transactions on Parallel
and Distributed Systems 9 (12) (1998) 1226–1236.

You might also like