Node-disjoint paths in incomplete WK-recursive
networks
Ming-Yang Su
a, Hui-Ling Huang
b, Gen-Huey Chen
c,*,
Dyi-Rong Duh
daDepartment of Electronic Engineering, St. John's and St. Mary's Institute of Technology,
Taipei, Taiwan, ROC
bDepartment of Information Management, Southern Taiwan University of Technology,
Tainan, Taiwan, ROC
cDepartment of Computer Science and Information Engineering, National Taiwan University,
Taipei 10764, Taiwan, ROC
dDepartment of Computer Science and Information Engineering, National Chi Nan University,
Nantou, Taiwan, ROC
Received 23 July 1996; received in revised form 11 May 1998; accepted 19 April 2000
Abstract
The incomplete WK-recursive networks have been recently proposed to relieve the re-striction on the sizes of the WK-recursive networks. In this paper, a maximal set of node-disjoint paths is constructed between arbitrary two nodes of an incomplete WK-recursive network. The eectiveness of the constructed paths is veri®ed by both theoretic analysis and extensive experiments. A tight upper bound on the maximal length is suggested. On the other hand, experimental results show that for arbitrary two nodes, the expected maximal length is not greater than twice their distance and about equal to the diameter. When the two nodes are the farthest pair, the maximal length is not greater than twice the diameter and the expected maximal length is not greater than 1.5 times the diameter. Ó 2000 Elsevier Science B.V. All rights reserved.
Keywords: Graph-theoretic interconnection network; Incomplete WK-recursive network; Node-disjoint paths; WK-recursive network
*Corresponding author. Tel.: +886-2-23625336; fax: +886-2-23628167.
E-mail address: ghchen@csie.ntu.edu.tw (G.-H. Chen).
0167-8191/00/$ - see front matter Ó 2000 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 8 1 9 1 ( 0 0 ) 0 0 0 6 5 - X
1. Introduction
The WK-recursive networks [23] own two attractive topological properties: ex-pansibility and equal degree. A network is expansible if no changes to node con-®guration and link connection are necessary when it is expanded, and of equal degree if its nodes have the same degree no matter what its size is. A network with these two properties will gain the advantages of easy implementation and low cost when it is manufactured. A VLSI implementation of a 16-node WK-recursive network has been realized at the Hybrid Computing Research Center [23]. This prototype net-work was further extended to 64 nodes later [24]. Recently two variants of the WK-recursive networks have been proposed in [5,6].
Although the WK-recursive networks own many favorable properties (see [1,3,4,7±9,11,23,24]), there is a rigorous restriction on their sizes. As will become clear in Section 2, the number of nodes contained in a WK-recursive network must satisfy dt, where d > 1 is the size of the basic building block and t P 1 is the level of
expansion. Thus, as d 4, extra 3 47 49152 nodes are required to expand from a
7-level WK-recursive network to an 8-level one. Almost all announced networks have suered from the same restriction. In order to relieve this restriction, some incomplete networks have been proposed recently. Among them, incomplete hy-percubes [12], incomplete star networks [14,17], clustered-star graphs [13], incom-plete rotator graphs [16], and incomincom-plete WK-recursive networks [19] are some examples. Previously several results on the incomplete WK-recursive networks were obtained; topological properties were investigated in [21], a shortest-path routing algorithm appeared in [22], and a broadcasting algorithm was proposed in [20].
Given a network, it is both theoretically interesting and practically important to ®nd node-disjoint paths (disjoint paths for short) between any two of its nodes. With disjoint paths, transmission rate can be accelerated and transmission reliability can be enhanced. In the past, a maximal set of disjoint paths was constructed for many (complete) networks, e.g., hypercubes [18], star graphs [2], and WK-recursive works [3]. However, the same problem remained unsolved for all incomplete net-works but the clustered-star graphs (see Ref. [10]). In this paper, a maximal set of disjoint paths is constructed for the incomplete WK-recursive networks.
In the next section, the incomplete WK-recursive networks are ®rst reviewed. A prerequisite step for constructing disjoint paths is presented in Section 3. By its aid a maximal set of disjoint paths is constructed in Section 4. Moreover, a tight upper bound on their maximal length is suggested. In Section 5, extensive experiments are further made to verify their eectiveness. Finally, we conclude this paper with some remarks in Section 6.
2. Incomplete WK-recursive networks
The WK-recursive networks can be built incrementally with basic building blocks. Any complete graph can serve as a basic building block. Throughout this paper we use graph and network interchangeably. Let K d; t denote a WK-recursive network
of level t whose basic building block is a d-node complete graph, where d > 1 and t P 1. K d; 1, which is the basic building block, is the d-node complete graph, and K d; t for t P 2 is a d-supernode complete graph, where each supernode is a K d; t ÿ 1. Each node of a K d; t is assigned with a unique identi®er which consists of a d-ary sequence of length t. The following de®nition is due to Chen and Duh [1]. De®nition 2.1. The node set of a K d; t is denoted by fatÿ1atÿ2 a1a0j
ai2 f0; 1; . . . ; d ÿ 1g for 0 6 i 6 t ÿ 1g. Node adjacency is de®ned as follows:
atÿ1atÿ2 a1a0is adjacent to (1) atÿ1atÿ2 a1b, where 0 6 b 6 d ÿ 1 and b 6 a0, and
(2) atÿ1atÿ2 aj1ajÿ1ajjif aj6 ajÿ1and ajÿ1 ajÿ2 a0for some 1 6 j 6 t ÿ 1,
where ajjrepresents j consecutive ajs. The links of (1) are called substituting links, and
are labeled 0. The link of (2), if existing, is called j-flipping link (or simply flipping link), and is labeled j. Besides, if atÿ1 atÿ2 a0, there is a link, called open
link, incident to node atÿ1atÿ2 a1a0. The open link, which is labeled t, is reserved
for further expansion.
The structures of K 4; 1 and K 4; 3 are illustrated in Fig. 1. The links within basic building blocks are substituting links, and those connecting two embedded K d; js are j-¯ipping links. For example, the link between nodes 311 and 133 is a 2-¯ipping link, and the other links incident to node 311 are all substituting links. The open links are incident to nodes 000, 111, 222, and 333.
De®nition 2.2. De®ne ctÿ1ctÿ2 cm K d; m to be the subgraph of a K d; t induced
by fctÿ1ctÿ2 cmamÿ1 a1a0jaj2 f0; 1; . . . ; d ÿ 1g for 0 6 j 6 m ÿ 1g, where
1 6 m 6 t ÿ 1 and ctÿ1; ctÿ2; . . . ; cmare all integers from f0; 1; . . . ; d ÿ 1g.
For example, refer to Fig. 1(b), where 31 K 4; 1 is the subgraph of K 4; 3 in-duced by {310, 311, 312, 313}.
De®nition 2.3. Node atÿ1atÿ2 a1a0 is a k-frontier if akÿ1 akÿ2 a1 a0,
where 1 6 k 6 t.
By de®nition a k-frontier is automatically an l-frontier for 1 6 l 6 k ÿ 1. Both end nodes of a k-¯ipping link are k-frontiers. For 1 6 m 6 t ÿ 1, an embedded K d; m contains one m 1-frontier and d ÿ 1 m-frontiers. These d frontiers are 2mÿ 1
distant from each other.
The incomplete WK-recursive networks, which were originally de®ned in [19], are induced subgraphs of the WK-recursive networks. If we number the nodes of a K d; t according to their lexicographical order, then an N-node incomplete WK-recursive network is the subgraph of the K d; t induced by the ®rst N nodes. Throughout this paper we use IK d; t to denote an N-node incomplete WK-recur-sive network, where dtÿ1< N < dtand N is a multiple of d.
Associated with an N-node IK d; t is a coefficient vector (btÿ1; btÿ2; . . . ; b1), where
0 6 bj6 d ÿ 1 for all 1 6 j 6 t ÿ 1 and N btÿ1dtÿ1 btÿ2dtÿ2 b1d. It
embedded K d; js: btÿ1btÿ2 bj10 K d; j; btÿ1btÿ2 bj11 K d; j; . . . ; btÿ1btÿ2
bj1 bjÿ 1 K d; j. For example, the structure of an IK 4; 3 with coecient
vector (3, 2) is shown in Fig. 2. It contains three embedded K 4; 2s, i.e., 0 K 4; 2; 1 K 4; 2, and 2 K 4; 2, and two embedded K 4; 1s, i.e., 30 K 4; 1
Fig. 1. The structrue of: (a) K 4; 1 and (b) K 4; 3. This ®gure also shows a heuristic routing path and the shortest routing path between nodes 033 and 133.
and 31 K 4; 1. In the rest of this paper, coecient vector (btÿ1; btÿ2; . . . ; b1) is
written as (btÿ1; btÿ2; . . . ; bi; ), provided bi6 0 and biÿ1 biÿ2 b1 0, where
1 6 i 6 t ÿ 1. For example, (2, 0, 4, 0, 0) is written as (2, 0, 4, ).
Given an IK d; t with coecient vector (btÿ1; btÿ2; . . . ; bi; ), let Sm represent the
subgraph induced by the nodes of btÿ1btÿ2 bm10 K d; m; btÿ1btÿ2 bm11
K d; m; . . . ; btÿ1btÿ2 bm1 bmÿ 1 K d; m, where i 6 m 6 t ÿ 1. For example,
given an IK 5; 7 with coecient vector (4, 2, 4, 3, 1, 4, ), S6 contains
0 K 5; 6; 1 K 5; 6; 2 K 5; 6, and 3 K 5; 6; S5 contains 40 K 5; 5 and
41 K 5; 5, and so on. We note that the embedded K d; ms within Sm join one
another through m-¯ipping links. That is, Sm is a bm-supernode complete graph
with each supernode being a K d; m. If each Sm is regarded as a stage, then the
structure of the IK d; t forms a (t ÿ i)-stage graph, denoted by Stÿ1 Stÿ2 Si. Refer to Fig. 3, where three examples are shown. For the
sake of simplicity, each embedded K d; m within Sm is drawn as a circle, and the
one btÿ1btÿ2 bm1j K d; m is denoted by Cmj, where 0 6 j 6 bmÿ 1. All the links
within Sm are omitted.
There are minfbm; bmÿ1g m-¯ipping links between Smand Smÿ1that connect Cmj and
Cj
mÿ1 for all 0 6 j 6 minfbm; bmÿ1g ÿ 1. Besides, there may exist a u-¯ipping link
between Suand Sv, where i 6 v < u 6 t ÿ 1 and u ÿ v > 1. Such a link, if it exists, is
called a jumping u-flipping link. A necessary and sucient condition for the existence of jumping ¯ipping links is presented below.
Theorem 2.1. Given an IK d; t with coefficient vector (btÿ1; btÿ2; . . . ; bi; ), one
jumping u-flipping link exists between Su and Sv if and only if bu> buÿ1
buÿ2 bv1< bv, where i 6 v < u 6 t ÿ 1 and u ÿ v > 1. Moreover, this jumping
flipping link connects Ce
uand Cev, where e buÿ1 buÿ2 bv1.
Proof. (() According to the de®nition of IK d; t, there are e u-¯ipping links between Suand Suÿ1that connect Cujand Cuÿ1j for all 0 6 j 6 buÿ1ÿ 1. Besides, there exists one
jumping u-¯ipping link connecting Ce
u and Cve whose two end nodes are
btÿ1btÿ2 bu1ebuu2 Cue and btÿ1btÿ2 bu1bueu btÿ1btÿ2 bu1buee eev12 Cve.
For buÿ1< j 6 buÿ 1, the jumping ¯ipping link btÿ1btÿ2 bu1jbuu; btÿ1btÿ2
bu1buju) does not exist because btÿ1btÿ2 bu1buju is not a node of the IK d; t.
()) Without loss of generality, assume the jumping u-¯ipping link emits from Ca u
to Sv for some 0 6 a 6 buÿ 1. We ®rst show bu> buÿ1 by contradiction. Suppose
bu6 buÿ1. There is a u-¯ipping link between Cju and Cuÿ1j for all 0 6 j 6 buÿ 1.
Re-garding Suÿ1 Suÿ2 Sias an embedded IK d; u, there are two u-¯ipping links
between Ca
uand the embedded IK d; u: one is between Cuaand Cuÿ1a and the other is
between Ca
u and Sv. This is a contradiction because at most one u-¯ipping link may
exist between any two embedded K d; us and an IK d; u is a subgraph of a K d; u. Similarly, a P buÿ1 can be proved.
We continue to show that a buÿ1; buÿ1 buÿ2 bv1< bv, and the jumping
u-¯ipping link connects Ca
uand Cav. According to the de®nition of u-¯ipping link, the
end node of the jumping u-¯ipping link in Su(actually in Cua) is btÿ1btÿ2 bu1abuu. Thus
the other end node in Sv is btÿ1btÿ2 bu1buau. It is not dicult to see that
btÿ1btÿ2 bu1buaudoes not belong to the IK d; t if a > buÿ1. Consequently, we have
a buÿ1 and the end node in Sv is btÿ1btÿ2 bu1bubuuÿ1 btÿ1btÿ2
bu1bubuÿ1buÿ1 buÿ1 buÿ1v1. The latter further implies that buÿ1 buÿ2
bv1< bvand the end node is located in Cav Cbvuÿ1.
In the rest of this paper we use Je
u;v to denote the jumping u-¯ipping link that
connects Ce
uand Cve(refer to Fig. 3 for illustration). Theorem 2.1 provides a fast way
to determine all jumping ¯ipping links from the coecient vector btÿ1; btÿ2; . . . ; bi; . We only need to examine btÿ1; btÿ2; . . . ; bi; from the left to
the right so that Je
u;vexists if bu> buÿ1 buÿ2 bv1< bv, where u ÿ v > 1 and
e buÿ1 buÿ2 bv1.
We note that for i 6 n 6 m 6 t ÿ 1; bm6 0, and bn6 0; Sm Smÿ1 Sn
forms an embedded IK d; m 1 with coecient vector bm; bmÿ1; . . . ; bn; whose
each node has its identi®er pre®xed with btÿ1btÿ2 bm1. For example, refer to
Fig. 3(c), where S3 S2 S1forms an embedded IK 5; 4 with coecient vector (3, 1,
4, ) whose each node has its identi®er pre®xed with 424. Theorem 2.1 can be applied to Sm Smÿ1 Snas well.
3. A prerequisite step
Suppose X and Y are arbitrary two nodes of an IK d; t with coecient vector btÿ1; btÿ2; . . . ; bi; . Without loss of generality, we assume X 2 Cma and Y 2 Cbn,
where i 6 n 6 m 6 t ÿ 1; 0 6 a 6 bmÿ 1 and 0 6 b 6 bnÿ 1. In this section, an
algo-rithm that groups stages Sm; Smÿ1; . . . ; Siinto blocks is proposed. Each block contains
one or more consecutive stages, and every two adjacent blocks intersect with one stage. The union of all blocks is the set of all stages. The algorithm will be invoked when we construct disjoint paths between X and Y in Section 4.
With input bm; bmÿ1; . . . ; bi; and a, the algorithm produces a sequence of
integers m0; m1; . . . ; mk, where k P 0 and m P m0> m1> > mk i. These
integers de®ne k 1 blocks, i.e., Sm Smÿ1 Sm0; Sm0 Sm0ÿ1 Sm1; ;
Smkÿ1 Smkÿ1ÿ1 Smk. The algorithm, as shown below, takes O m time.
Algorithm (Stage_Grouping bm; bmÿ1; . . . ; bi; ; a: = 0 6 a 6 bmÿ 1 =.
(1). Scan bm; bmÿ1; . . . ; bi; from the left to the right and determine in sequence
Jx1
y1;z1; Jyx22;z2; . . . ; Jyxcc;zc, so that a > x1> x2> > xc. That is, Jyx11;z1 is the ®rst jumping
¯ipping link encountered in the scanning which has x1< a. Each Ju;ve between Jyxjj;zj
and Jxj1
yj1;zj1 has e P xj, where 1 6 j < c, and each Ju;ve after Jyxcc;zc has e P xc. Let
L fx1; x2; . . . ; xcg. If no feasible jumping ¯ipping link is found in the scanning, L is
empty.
For example, refer to Fig. 3(c) again. If X 2 C3
6, then (4, 2, 4, 3, 1, 4, ) and 3 are
taken as input. Since J2
6;4 and J3;11 are found in the scanning, L f2; 1g. Similarly, if
X 2 C2
6, only J3;11 is found and thus L f1g. On the other hand, if X belong to C61or
C0
6; L is empty because no feasible jumping ¯ipping link can be found. We also note
that m P y1> z1P y2> z2P P yc> zcP i. By the aid of Theorem 2.1, this step
can be completed in O m time.
(2). Determine m0 minfr jbr> a and bjP a for all m P j > rg. If L is not
empty, determine m1; m2; . . . ; mc sequentially as follows: m1 minfr jbr> x1 and
bjP x1 for all m0> j > rg, m2 minfr jbr> x2 and bjP x2 for all m1> j > rg; . . . ;
mc minfr jbr> xc and bjP xc for all mcÿ1> j > rg.
By examining bm; bmÿ1; . . . ; bi; from the left to the right, this step can be
completed in O m time. For example, refer to Fig. 3(c) again. If a 3, we have m0 6; m1 3, and m2 1. If a 2, we have m0 3 and m1 1. If a 1 or 0, we
have m0 1. We note that m P m0P yl> z1P m1P y2> z2P P mcÿ1P
yc> zcP mcP i.
(3). Output m0; m1; . . . ; mk i, where k c or c 1, according to the following
four cases:
Case 1. L is empty and m0 i. Output m0.
Case 2. L is empty and m0> i. Set m1 i and output m0; m1.
Case 3. L is not empty and mc i. Output m0; m1; . . . ; mc.
Case 4. L is not empty and mc > i. Set mc1 i and output m0; m1; . . . ; mc; mc1.
Refer to Fig. 3(c) again. The algorithm will output (6, 3, 1), (3, 1), (1), and (1) if a 3; 2; 1, and 0, respectively. The output m0; m1; . . . ; mk de®nes k 1 blocks,
denoted by B0; B1; . . . ; Bk, where B0 Sm Smÿ1 Sm0 and Bl Smlÿ1 Smlÿ1ÿ1
Sml for all 1 6 l 6 k. Every two adjacent blocks Blÿ1 and Bl contain one
common stage Smlÿ1. Lemma 3.1 [22]. Let Jx1 y1;z1; J x2 y2;z2; . . . ; J xc
yc;zc and m0; m1; . . . ; mcbe defined as in Algorithm
Stage_Grouping. Then, for all 1 6 j 6 c, 1. mjÿ1P yj;
2. if mjÿ1> yj, then bmjÿ1 > xjÿ1P bmjÿ1ÿ1P P byj;
4. if zj> mj, then bzj> xj; bmj > xj, and bqP xjfor all zj> q > mj;
5. xj minfbmjÿ1; bmjÿ1ÿ1; . . . ; byj; . . . ; bzj; . . . ; bmjg;where x0 a is assumed.
This lemma is illustrated in Fig. 4, where mjÿ1> yj and zj> mj are assumed.
According to Lemma 3.1, Jx1
y1;z1; Jyx22;z2; . . . ; Jyxcc;zc are the leftmost and upmost jumping
¯ipping links in B1; B2; : . . . ; Bc, respectively (the smaller the value xjis, the upper Jyxjj;zj
is). That is, for any Je
u;vin Bjwe have u 6 yjand e P xj(actually u 6 zjif Ju;ve 6 Jyxjj;zj).
We note that B0 may or may not contain jumping ¯ipping links and Bc1, if it
exists, does not contain any jumping ¯ipping link. For Bc1 we have bmc >
xcP bmcÿ1P P bmc1 mc1 i. We also note that B0 contains at least one
stage, Bjfor 0 < j 6 c contains at least three stages, and Bc1, if it exists, contains at
least two stages.
4. Construction of disjoint paths
In this section, disjoint paths are constructed between arbitrary two nodes X and Y of an IK d; t with coecient vector btÿ1; btÿ2; . . . ; bi; . Without loss of
generality, we assume X 2 Ca
m and Y 2 Cnb, where i 6 n 6 m 6 t ÿ 1; 0 6 a
6 bmÿ 1; and 0 6 b 6 bnÿ 1. The disjoint paths have maximal length not greater
than 2m1 2mÿ 1. The construction time is O d D
m, where Dm 2mÿ 1 is the
diameter of a K d; m.
First we consider a trivial case of m n and a b. Since X and Y belong to the same embedded K d; m; d ÿ 1 disjoint paths between X and Y can be obtained by Duh and Chen's work [3]. These d ÿ 1 paths are all within Ca
m Cbn, and their
maximal length is not greater than 3Dmÿ1 2 2m 2mÿ1ÿ 1. It should be
men-tioned that there may exist one more disjoint path between X and Y that goes outside Ca
m. This path, if it exists, is much longer than those d ÿ 1 paths within Cam. We
exclude this path from our discussion.
In the rest of this section, a maximal set of disjoint paths between X and Y is constructed within Sm Smÿ1 Sifor: (i) m n and a 6 b and (ii) m > n. It is
still possible that there exists one additional disjoint path going outside Sm Smÿ1 Si. This path, if it exists, is excluded from our discussion because it
is too long as compared with those within Sm Smÿ1 Si.
To begin, we have to review Vecchia and Sanges' routing algorithm [23] for a K d; t because it is necessary to the construction of the disjoint paths. Suppose A and B are arbitrary two nodes of a K d; t. We de®ne A rB if they belong to the
same embedded K d; r, and A 6rB otherwise, where 1 6 r 6 t. For example, refer to
Fig. 1(b), where 033 3133, but 033 62133. A routing path from A to B within a
K d; t can be obtained by the following procedure.
1. Determine the level r so that A rB but A 6rÿ1B, where 1 6 r 6 t.
2. Determine the ¯ipping link, say (W ; Z), so that A rÿ1W and Z rÿ1B.
3. Determine a routing path from A to W and a routing path from Z to B, recursively. A routing path from A to B is obtained by concatenating the routing path from A to W, the ¯ipping link (W ; Z), and the routing path from Z to B. For example, a routing path from node 033 to node 133 within K 4; 3 is shown with bold lines in Fig. 1(b). When a message is transmitted from A to B, it is ®rst routed to the nearest r ÿ 1-frontier, say Z, with Z rÿ1B, then routed to the nearest r ÿ 2-frontier, say
Z0, with Z0
rÿ2B, and so on. In other words, when the message is going along the
routing path, the identi®ers of the traversed nodes are gradually equalized with B from the left to the right. For example, let us consider the routing path from node 033 to node 133 that is indicated with bold lines in Fig. 1(b). The left digit is equalized at node 100, the middle digit is equalized at node 130, and ®nally all the three digits are equalized at the destination node 133. The following observation is immediate.
Observation 4.1. When routing a message according to Vecchia and Sanges' algo-rithm, the identi®ers of the traversed nodes are gradually equalized with B from the left to the right.
Vecchia and Sanges' algorithm, although simple, does not guarantee the shortest path. For example, the shortest path from node 033 to node 133 is shown with dashed lines in Fig. 1(b). Let p A; B denote the routing path from node A to node B within a K d; t that is produced by Vecchia and Sanges' algorithm. The following two lemmas have been proved in [3].
Lemma 4.1 [3]. Suppose A and B are arbitrary two nodes of a K(d; t). If A rB and
either of them is an r-frontier, then p A; B is the shortest, where 1 6 r 6 t. Moreover, it takes at most O Dt time to determine p A; B, where Dt 2tÿ 1 is the diameter of the
Lemma 4.2 [3]. Suppose A is an arbitrary node of a K d; t, and let Vl;0; Vl;1; . . . ; Vl;dÿ1
be the dl-frontiers of an embedded K d; l that contains A, where 1 6 l 6 t. Then, the d paths p A; Vl;0; p A; Vl;1; . . . ; p A; Vl;dÿ1 are mutually disjoint, exclusive of A.
By Lemma 4.1, p A; Vl;0; p A; Vl;1; . . . ; p A; Vl;dÿ1 are all the shortest and they can
be determined in O d Dl time. Now we are ready to construct disjoint paths
between X 2 Ca
m and Y 2 Cnb. First we consider the situation of m n and a 6 b.
Within Smthere are bmÿ 1 disjoint paths between X and Y. Besides, there may exist
one more disjoint path that passes through the embedded IK d; m formed by Smÿ1 Smÿ2 Si. These bm paths are pictorially expressed in Fig. 5, where
a 62 f0; bmÿ 1g; b 62 f0; bmÿ 1g, and a < b are assumed. In Fig. 5, each thin line
represents a ¯ipping (or jumping ¯ipping) link and each bold line represents a subpath obtained by Vecchia and Sanges' algorithm. According to Lemmas 4.1 and 4.2, these bm paths are mutually disjoint (exclusive of X and Y) and they can be
determined in O bm Dm 6 O d Dm time.
We also note that the path passing the embedded IK d; m exists only if the two nodes btÿ1btÿ2 bm1bmam and btÿ1btÿ2 bm1bmbm belong to the embedded
IK d; m (refer to Fig. 5). According to Observation 4.1, p btÿ1btÿ2 bm1bmam;
btÿ1btÿ2 bm1bmbm, which is constructed within btÿ1btÿ2 bm1bm K d; m, can be
expressed as follows: btÿ1btÿ2 bm1bmam! ! btÿ1btÿ2 bm1bmabmÿ1; ! btÿ1btÿ2 bm1bmbamÿ1! ! btÿ1btÿ2 bm1bmbabmÿ2; ! btÿ1btÿ2 bm1bmb2amÿ2! ! btÿ1btÿ2 bm1bmb2abmÿ3; ! ... ! btÿ1btÿ2 bm1bmbmÿ2a2! btÿ1btÿ2 bm1bmbmÿ2ab; ! btÿ1btÿ2 bm1bmbmÿ1a ! btÿ1btÿ2 bm1bmbm:
Each node in the subpath from btÿ1btÿ2 bm1bmam to btÿ1btÿ2 bm1bmabmÿ1
has leading digits btÿ1btÿ2 bm1bma, each node in the subpath from
btÿ1btÿ2 bm1bmbamÿ1 to btÿ1btÿ2 bm1bmbabmÿ2 has leading digits
btÿ1btÿ2 bm1bmba, and so on. That is, every node in p btÿ1btÿ2 bm1bmam;
btÿ1btÿ2 bm1bmbm precedes btÿ1btÿ2 bm1bmbm lexicographically if a < b,
and precedes btÿ1btÿ2 bm1bmam lexicographically if a > b. Hence,
p btÿ1btÿ2 bm1bmam; btÿ1btÿ2 bm1bmbm is entirely contained in the embedded
IK d; m, provided the two end nodes are contained in the embedded IK d; m. Then we consider the situation of m > n. With input bm; bmÿ1; . . . ; bi; and
a, Algorithm Stage_Grouping is ®rst executed to produce m0; m1; . . . ; mk so
Smk, where k P 0 and mk i. Suppose Y 2 Cbn belongs to Bl, where 0 6 l 6 k.
If Y 2 Sml, we consider Y 2 Bl but 62 Bl1. Two cases are discussed below.
Case 1 (l 0). We have m > n P m0. Let j minfbm; bmÿ1; . . . ; bng if
minfbm; bmÿ1; . . . ; bng bm or bnand j minfbm; bmÿ1; . . . ; bng 1 else. There are j
disjoint paths between X and Y, denoted by P0; P1; . . . ; Pjÿ1, where Pr passes
Ca
m; Crm; Crmÿ1; . . . ; Crn; Cbn in sequence for all 0 6 r 6 j ÿ 1 Cma Crm as a r, and
Cr
n Cbn as r b). Besides, there is an additional disjoint path, denoted by Pj, if
j bn< bm and there is a jumping ¯ipping link, say Ju;vbn, under Sn, where
m > u > n > v. It is still possible that Pj contains other jumping ¯ipping links
Fig. 5. bm Disjoint paths between X and Y, where a 62 f0; bmÿ 1g; b 62 f0; bmÿ 1g, and a < b are
between Sm and Su. Without loss of generality, we assume Pj contains only Ju;vbn, and
so Pj passes Cam; Cjm; Cmÿ1j ; . . . ; Cuj; Cvj; Cvb; Cbv1; . . . ; Cbn, in sequence. According to
Lemmas 4.1 and 4.2, these Pjpaths are mutually disjoint (exclusive of X and Y) and
they can be determined in O j Dm 6 O d Dm time.
For example, let us consider X 2 C2
6 and Y 2 C51 in an IK 5; 7 with coecient
vector (4, 2, 4, 3, 1, 4, ) (refer to Fig. 6). The execution of Algorithm Stage_Grouping produces B0 S6 S5 S4 S3 and B1 S3 S2 S1. We have
j minf4; 2g 2. Since j b5< b6 and one jumping ¯ipping link goes under
S5, there are j 1 3 disjoint paths between X and Y. These three paths are
shown in Fig. 6, where each thin line represents a ¯ipping (or jumping ¯ipping) link and each bold line represents a subpath obtained by Vecchia and Sanges' algorithm.
Case 2 (0 < l 6 k). We have mlÿ1> n P ml. Let a >x1> x2> > xc be
de-®ned as in Algorithm Stage_Grouping, where c k or k ÿ 1. Recall the discussion in the last paragraph of Section 3. There is at least one jumping ¯ipping link in Bl unless l c 1. We ®rst assume l < c 1 and let Jyxll;zl represent the leftmost
and upmost jumping ¯ipping link in Bl. Three subcases have to be discussed
below.
Subcase 1 (mlÿ1> n P yl). There are bn disjoint paths between X and Y, denoted
by P0; P1; . . . ; Pbnÿ1, where Pr passes Cma; Crm; Cmÿ1r ; . . . ; Crn; Cnb in sequence for all
0 6 r 6 bnÿ 2 and Pbnÿ1 passes Cma; Cmÿ1a ; . . . ; Cma0; Cxm10; C x1 m0ÿ1; . . . ; Cmx11; Cxm21; C x2 m1ÿ1; . . . ; Cx2 m2; Cmx32; . . . ; Cmxlÿ1lÿ1; Cmbnlÿ1ÿ1; C bnÿ1 mlÿ1ÿ1; . . . ; Cbnnÿ1; Cnb in sequence.
For example, refer to Fig. 7, where an IK 5; 7 with coecient vector (4, 2, 4, 3, 1, 4, ) is shown. Suppose X 2 C3
6 and Y 2 C31. The execution of Algorithm
Stage_Grouping produces B0 S6; B1 S6 S5 S4 S3, and B2 S3 S2 S1.
There are b3 3 disjoint paths between X and Y as shown in Fig. 7(a).
Subcase 2 (yl> n P ml and bn xl). There are bn 1 disjoint paths between X
and Y. Since bn xl minfbmlÿ1; bmlÿ1ÿ1; . . . ; bmlg (by Lemma 3.1), bmlÿ1 > xlÿ1> xl,
and bml > xl, there exists a jumping ¯ipping link under Sn, say Ju;vbn, where
mlÿ1P u > v P ml. We use P0; P1; . . . ; Pbnto denote the bn 1 disjoint paths, where Pr
passes Ca
m; Crm; Cmÿ1r ; . . . ; Cnr; Cnb in sequence for all 0 6 r 6 bnÿ 1 and Pbn passes
Ca m; Camÿ1; . . . ; Cam0; C x1 m0; C x1 m0ÿ1; . . . ; C x1 m1; C x2 m1; C x2 m1ÿ1; . . . ; C x2 m2; C x3 m2; . . . ; C xlÿ1 mlÿ1; C xl mlÿ1; Cxl mlÿ1ÿ1; C xl mlÿ1ÿ2; . . . ; C xl u; Cvxl; Cvb; Cv1b ; . . . ; Cnb in sequence.
For example, suppose X 2 C3
6 and Y 2 C20belong to the same IK 5; 7. There are
b2 1 2 disjoint paths between X and Y as shown in Fig. 7(b).
Subcase 3 (yl> n P ml and bn> xl). Actually we have zlP n P ml and bn> xl
because bylÿ1 bylÿ2 bzl1 xl can be assured by Theorem 2.1. There are
xl 1 disjoint paths between X and Y, denoted by P0; P1; . . . ; Pxl, where Pr passes
Fig. 7. Disjoint paths between X and Y for Case 2. (a) mlÿ1> n P yl; (b) yl> n P mland bn xlÿ 1;
Ca
m; Crm; Crmÿ1; . . . ; Cnr; Cbn in sequence for all 0 6 r 6 xlÿ 1 and Pxl passes
Ca m; Camÿ1; . . . ; Cm0 a; Cxm10; C x1 m0ÿ1; . . . ; Cxm11; Cmx21; Cm1ÿ1 x2; . . . ; Cmx22; Cxm32; . . . ; Cmxlÿ1lÿ1; Cmxllÿ1; Cxl mlÿ1ÿ1; C xl mlÿ1ÿ2; . . . ; Cxyll; Czxll; . . . ; Cxnl; Cbn in sequence.
For example, suppose X 2 C3
6 and Y 2 C12 belong to same IK 5; 7. There are
xl 1 2 disjoint paths between X and Y as shown in Fig. 7(c).
On the other hand, if l c 1, we have bmc> xcP bmcÿ1P P bnP P
bmc1 mc1 i. There are bn disjoint paths between X and Y whose construction is
similar to Subcase 1.
According to Lemmas 4.1 and 4.2, the paths obtained for Case 2 are mutually disjoint (exclusive of X and Y), and they can be determined in O maxfbn 1;
xl 1g Dm 6 O d Dm time.
The following theorem holds as a consequence of our discussion above. Theorem 4.1. Suppose X 2 Ca
mand Y 2 Cbnbelong to an IK d; t with coefficient vector
btÿ1; btÿ2; . . . ; bi; , where i 6 n 6 m 6 t ÿ 1; 0 6 a 6 bmÿ 1, and 0 6 b 6 bnÿ 1. Then
the disjoint paths between X and Y can be determined in O d Dm time.
Let len X ; Y be the maximal length of the disjoint paths between X and Y. In the following we show that len X ; Y has an upper bound of 2m1 2mÿ 1.
Theorem 4.2. Suppose X 2 Ca
mand Y 2 Cbnbelong to an IK d; t with coefficient vector
btÿ1; btÿ2; . . . ; bi; , where i 6 n 6 m 6 t ÿ 1; 0 6 a 6 bmÿ 1, and 0 6 b 6 bnÿ 1. Then
len X ; Y 6 2m1 2mÿ 1.
Proof. If m n, there are at most bm disjoint paths between X and Y. By the aid
of Lemma 4.1, the path passing the embedded IK d; m has length at most Dm 1 Dm 1 Dm 2m1 2mÿ 1, and the others each have length at most
Dm 1 Dm 1 Dm 2m1 2mÿ 1. Hence, len X ; Y 6 2m1 2mÿ 1. In the
rest of the proof, we assume m > n and Y 2 Blfor some 0 6 l 6 k.
If l 0, there are at most j 1 disjoint paths P0; P1; . . . ; Pj between X and Y. The
path Pj, if it exists, has length at most Dm 1 Dm 1 Dmÿ1 1 Du1
1 Du 1 Dv 1 Dv 1 Dv1 1 Dn< 2m1 2mÿ 1, and the
oth-ers each have length at most Dm 1 Dm 1 Dmÿ1 1 Dmÿ2 1
Dn1 1 Dn 1 Dn 2m1 2mÿ 1. Hence, len X ; Y 6 2m1 2mÿ 1.
If 0 < l 6 k, there are bn or bn 1 or xl 1 disjoint paths between X and Y,
ac-cording to three subcases. Similarly, len X ; Y 6 2m1 2mÿ 1.
5. Experiments and results
In this section the eectiveness of the disjoint paths is veri®ed by extensive ex-periments. The following two algorithms were implemented for the need of our experiments.
· Su, Chen, and Duh's algorithm that computes the diameter of an IK d; t [21]. · Su, Chen, and Duh's algorithm that computes the shortest path between arbitrary
Remarks. Adopting the prune-and-search technique [15], the algorithm of Su et al. [21] can compute the diameter of an IK d; t and the farthest pair of nodes in O t time. Although the diameter of an IK d; t can be computed by Su, Chen, and Duh's algorithm, no formula is available for computing it.
First we compare len X ; Y with dis X ; Y , where dis X ; Y is the distance between X and Y. Fig. 8 shows the average ratios of len X ; Y to dis X ; Y for IK d; ts with 4 6 d 6 6 and 2 6 t 6 10. The values of dis X ; Y were computed by the algorithm of Su et al. [22]. For each pair of d and t; 106 random instances were run and their
average ratio was computed. The averages got stable after running as many as 106
instances. A randomly generated coecient vector combined with two nodes, also randomly generated, of an IK d; t forms an instance. Experimental results showed that len X ; Y is not greater than twice dis X ; Y in average.
Then we compare len X ; Y with the diameter. Fig. 9 shows the average ratios of len X ; Y to the diameter for IK d; ts with 4 6 d 6 6 and 2 6 t 6 10. The diameters were computed by the algorithm of Su et al. [21]. Like Fig. 8, the average ratio for 106 random instances was taken for each pair of d and t. Experimental results
showed that len X ; Y is smaller than the diameter in average both as d 4 and t P 2 and as d 5 and t P 5. Besides, len X ; Y tends to the diameter as d 6 and t increases.
When X and Y were selected to be the farthest pair of nodes (i.e., dis X ; Y is equal to the diameter of the IK d; t), the average ratios of len X ; Y to the diameter were shown in Fig. 10. The farthest pair of nodes can be determined by the algorithm of Su et al. [21]. The average ratios each were obtained by running 105 random
in-stances because the averages got stable after running as many as 105 instances.
Experimental results showed that most of the averages fall in the range of 1.2±1.3. The maximal average ratio does not exceed 1.5.
It is worth mentioning that for the experiments of Fig. 10, no instance has len X ; Y exceeding twice the diameter. The distributions of the 105ratios obtained
for an IK 4; 8, an IK 5; 8, and an IK 6; 8 were shown in Figs. 11(a)±(c), respec-tively. For example, for the IK 4; 8 there are 54,566 ratios (about 54%) fall in the range of 1.0±1.1 and there are 4044 ratios (about 4%) fall in the range of 1.1±1.2.
Fig. 10. Average ratios of len X ; Y to dis X ; Y for IK d; t, where X and Y are the farthest pair nodes. Fig. 9. Average ratios of len X ; Y to diameter for IK d; t.
6. Concluding remarks
In this paper, we have constructed a maximal set of disjoint paths between X and Y, which are arbitrary two nodes of an IK d; t. The construction time is bounded by O d Dt. We have shown that the disjoint paths have maximal length not greater
than 2m1 2mÿ 1, where X 2 Ca
m; Y 2 Cnb, and i 6 n 6 m 6 t ÿ 1 are assumed. The
eectiveness of the disjoint paths was further veri®ed by extensive experiments. Experimental results showed that the disjoint paths have expected maximal length not greater than twice their distance and about equal to the diameter. Besides, when X and Y are the farthest pair, the disjoint paths have maximal length not greater than
twice the diameter and expected maximal length not greater than 1.5 times the di-ameter.
References
[1] G.H. Chen, D.R. Duh, Topological properties, communication, and computation on WK-recursive networks, Networks 24 (6) (1994) 303±317.
[2] M. Dietzfelbinger, S. Madhavapeddy, I.H. Sudborough, Three disjoint path paradigms in star networks, in: Proceedings of the International Symposium on Parallel and Distributed Processing, 1991, pp. 400±406.
[3] D.R. Duh, G.H. Chen, Topological properties of WK-recursive networks, Journal of Parallel and Distributed Computing 23 (3) (1994) 468±474.
[4] D.R. Duh, G.H. Chen, On the Rabin number problem, Networks 30 (3) (1997) 219±230.
[5] R. Fernandes, Recursive interconnection networks for multicomputer networks, in: Proceedings of the International Conference on Parallel Processing, vol. I, 1992, pp. 76±79.
[6] R. Fernandes, A. Kanevsky, Hierarchical WK-recursive topologies for multicomputer systems, in: Proceedings of the International Conference on Parallel Processing, vol. I, 1993, pp. 315±318. [7] R. Fernandes, A. Kanevsky, Substructure allocation in recursive interconnection networks, in:
Proceedings of the International Conference on Parallel Processing, vol. I, 1993, pp. 319±322. [8] R. Fernandes, D.K. Griesen, A. Kanevsky, Ecient routing and broadcasting in recursive
interconnection networks, in: Proceedings of the International Conference on Parallel Processing, 1994, pp. 51±58.
[9] R. Fernandes, D.K. Friesen, A. Kanevsky, Embedding rings in recursive networks, in: Proceedings of the International Symposium on Parallel and Distributed Processing, 1994, pp. 273±280.
[10] O.P. Gu, S. Peng, Algorithms for node dispoint paths in incomplete star networks, in: Proceedings of the International Conference on Parallel and Distributed Systems, 1994, pp. 296±303.
[11] A. Iazzetta, C. Sanges, U. Scafuri, Special multicast wormhole routing for WK-recursive networks, manuscript.
[12] H.P. Katse, Incomplete hypercubes, IEEE Transactions on Computers C-37 (5) (1988) 604±608. [13] S. Lati®, N. Bagherzadeh, On clustered-star graph and its properties, manuscript.
[14] S. Lati®, P.J. Bagherzadeh, Incomplete star: An incrementally scalable network based on the star graph, IEEE Transactions on Parallel and Distributed Systems 5 (1) (1994) 97±102.
[15] D.T. Lee, F.P. Preparata, Computational geometry ± a survey, IEEE Transactions on Computers C-33 (12) (1984) 1072±1101.
[16] S. Ponnuswamy, V. Chaudhary, Embedding of cycles in rotator and incomplete rotator graphs, in: Proceedings of the International Symposium on Parallel and Distributed Processing, 1993, pp. 603± 610.
[17] C.P. Ravikumar, A. Kuchlous, G. Manimaran, Incomplete star graph: An economical fault-tolerant interconnection network, in: Proceedings of the International Conference on Parallel Processing, vol. I, 1993, pp. 83±90.
[18] Y. Saad, M.H. Schultz, Topological properties of hypercubes, IEEE Transactions on Computers 37 (7) (1998) 867±872.
[19] M.Y. Su, Incomplete WK-recursive networks, Ph.D. dissertation, Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan, 1997.
[20] M.Y. Su, G.H. Chen, D.R. Duh, Broadcasting on incomplete WK-recursive networks, Journal of Parallel and Distributed Computing 57 (1999) 271±294.
[21] M.Y. Su, G.H. Chen, D.R. Duh, Topological properties of incomplete WK-recursive networks, Networks 32 (1) (1998) 23±45.
[22] M.Y. Su, G.H. Chen, D.R. Duh, A shortest-path routing algorithm for incomplete WK-recursive networks, IEEE Transactions on Parallel and Distributed Systems 8 (4) (1997) 367±379.
[23] G.D. Vecchia, C. Sanges, A recursively scalable network VLSI implementation, Future Generation Computer Systems 4 (3) (1988) 235±243.
[24] G.D. Vecchia, C. Sanges, An optimized broadcasting technique for WK-recursive topologies, Future Generation Computer Systems 4 (3) (1989/1990) 353±357.