On Rearrangeable and Non-Blocking Switching Networks: Journal of Computer and System Sciencfs
On Rearrangeable and Non-Blocking Switching Networks: Journal of Computer and System Sciencfs
On Rearrangeable and Non-Blocking Switching Networks: Journal of Computer and System Sciencfs
NICHOLAS PIPPENGER
Switching networks of the type used in telephone exchanges are studied, with emphasis
on a particular class of networks possessing great structural symmetry. This class contains
rearrangeable networks carrying N calls with roughly 6Nlogs N contacts, and non-
blocking networks with roughly 16N(log, N)* contacts; these results are the best obtainable
by the methods used, We also show, by an impractical argument, that there are non-
blocking networks with roughly 9ON log, N contacts.
1. INTRODUCTION
Many parts of a telephone exchange have a cost proportional to the number of subscri-
bers, but the switching network in a telephone exchange shows a diseconomy of scale: a
network that serves twice as many subscribers costs more than twice as much. Thus as
larger and larger exchanges are considered, the cost of the network becomes more and
more important.
We shall study the problem of building a switching network for a telephone exchange
with the minimum possible cost. This study will involve several idealizations that
deserve discussion. First, we shall examine the problem from a combinatorial rather than
a probabilistic point of view. Thus we shall seek networks without blocking rather than
networks with a small probability of blocking. The questions studied combinatorially
here are studied probabilistically in Pippenger [lo, 111. Though non-blocking networks
are not used in practice to the extent that seldom-blocking networks are, they clearly
would be preferable if they could be built at the same cost. Whether the apparent differ-
ence in cost between these two types of networks is real or illusory is an open question that
will probably require further study of both types for its resolution.
Second, we assess the cost of a network simply as the number of contacts it contains, and
concern ourselves not with its numerical value but only with its asymptotic behavior.
These idealizations are traditional, and are used in most of the literature we cite. Further-
more, they (or other equally idealistic assumptions) are necessary if analysis is to be
carried out; without them we would be reduced to tabulating the results of numerical
optimizations. Thus they will enable us to observe a number of interesting qualitative
phenomena against a background of confusing quantitative detail. Once these phenomena
are understood, they can be used heuristically in the search for networks whose true costs
are numerically optimal.
145
0022~OOOO/78/0172-Ol45$02.OO/O
Copyright Q 1978 by Academic Prrss. Inc.
All rights of reproduction in any form reserved.
146 NICHOLAS PIPPENGER
2. SWITCHING NETWORKS
inputs,
outputs, and
contacts.
JJ’
Suppose .F is a network with I inputs and J outputs and 9 is a square network with K
lines. The extendedproduct .F x x 9 is a network obtained by taking K copies of S,
J copies of 9, and K copies of the mirror-image F* of F, and by interconnecting them
as shown in Figure 3. There is one link interconnecting each copy of F with each copy
of ‘9, and one link interconnecting each copy of 9 with each copy of Y*. The resulting
network has IK lines; if 9 has F contacts and Y has G, then F x x 9’ has 2FK + JG.
The sandwich W(A, , B,; .. . . A, , B,; M) is the network W,( constructed according to
the following scheme.
% = WW,
Wl = %?(A*, B,) x x WO ,
IK
contacts.
It is easy to see that, if the interconnections are suitably made, we have
F x (P’ x F) g (9 x F) x r,
FXX(FXXY)~(~XF)XX9,
where g means “is the same network as” (or, more formally, “is isomorphic to,” as in
Cantor [5]). From these facts we derive two propositions, both of which are easily
established by induction on k.
PROPOSITION 2.1. V(A,, B,;...; A,; BJ si W(A, , Bl) x 9’-(A,, B,;...; B, , B,).
PROPOSITION 2.2. W(A, , B,;...; A, , B,; M) g V(A, , B,;...; A,, B,) x x V(M)).
The purpose of a network is to establish paths from inputs to outputs. We shall
consider only paths of minimum length. In a crossbar these paths have length 1; in an
open-face sandwich V(A, , B, ;. . . ; AI, , Bk) they have length k; in a sandwich W(A, ,B,;...;
A, , B,; M) they have length 2k + 1. This minimum length will be called the number
of stages.
At any moment in time, many paths may be established simultaneously, but no two can
have a terminal in common (lest there be “crosstalk”). Such a set of paths will be called
a state.
In any state, all the terminals that are involved in established paths will be called bzcsy;
all others will be called idle. A path will be called busy if any of the terminals involved in it
is busy, and idle if all of them are idle.
SWITCHING NETWORKS 149
3. REARRANGEABLE NETWORKS
Rearrangeable networks can establish any set of connections from inputs to outputs.
An additional request for connection, however, may require a complete rearrangement
of the state. A request for disconnection, of course, presents no problems. Because of
the great effort that may be required to satisfy a request, rearrangeable networks are not
currently used in telephone exchanges, though they may find applications in other
large-scale systems such as reconfigurable computers. It will be worthwhile for us to
examine them, however, since this will display some useful techniques in a simple setting.
150 NICHOLAS PIPPENGER
The crossbar Q(N) is obviously a rearrangeable network with N lines and N2 contacts.
All extant methods for building rearrangeable networks with fewer contacts depend on
CRITERION 3.2. The sandwich ‘YY(A, , B,;. . . ;A, , B,; M) is yearrangeable ;f Bl >
A 1 ,..., B/c 2 4s .
We shall prove two theorems on the cost of rearrangeable sandwiches.
THEOREM 3.3. The minimum possible cost of a (2k + I)-stage sandwich with at least N
lines that satisjies Criterion 3.2 is asymptotic to P~N~+~I(~+~),
where
pk = 2(k + l)(+)l/‘k+l).
n Aj M > N,
( lsj$k 1
W>, 2 c AjfM N.
( LQ$k 1
This can be minimized by droping the integrality constraints and using elementary
calculus; the result is
W 2 2(k + 1)(Nk+2/2)1/(lc+1),
To obtain an upper bound we set Aj = Q for 1 < j < k, M = 2Q, and of course
Bj = A, for 1 < j < k, where
Q = [(N/2)li’“+l’].
THEOREM 3.4. The minimum possible cost of a sandwich (of any number of stages) with
at least N lines that satisjies Criterion 3.2 is asymptotic to 6N log, N.
Bj = A, (1 <j<k)
and
W& 2 c A&M N.
( l&k 1
W 3 N.
W 3 6N log, N + O(N),
M = [N/3”],
4. NON-BLOCKING NETWORKS
Non-blocking networks, like rearrangeable networks, can establish any set of connec-
tions from inputs of outputs. In contrast, however, an additional request for connection
can be satisfied without disturbing existing connections and irrespective of which state
the history of connections and disconnections has left the network in. Non-blocking
networks are thus ideal telephone exchanges. We shall examine them in the remainder of
this paper.
The crossbar g(N) is obviously a non-blocking network with N lines and N2 contacts.
All extant methods of building non-blocking networks with fewer contacts depend on
the notion of a majority-access network. Consider a network in some state. We shall say
that an input has access to an output if there is an idle path connecting them. A network
is a majority-access network if, in any state, each idle input has access to more than half
of the outputs. The point of this notion consists in
CRITERION 4.2. The sandwich %‘-(A,, B,;...; A,, B,; M) is non-blocking if B, :>
2(A, - I),..., BI, > 2(A, - 1).
THEOREM 4.3. The minimum possible cost of a (2k + 1)-stage sandwich with at least iV
lines that satisfies Criterion 4.2 is asymptotic to ~~N~i-ll(~+l’, where
Bj = 2Aj - 1 (I <jjk).
For the purpose of proving the lower bound we may assume that W = O(Nl+l(kt”).
Let us show that we may also assume that A, ,, .., A, and M are all N1l(k+r)U( 1). The
obvious 2A, - 1 > A, for-j = I,..., k, together with the first constraint, imply
w > 2 c ,4, + M N.
( l<j<h. )
From this it is clear that A, ,..., A, and M must all be O(N1l(h+l)), else we should not
have W = O(N1+l/(k+l)). These conditions, together with the first constraint, imply that
441 >..., A, and M must all be Q(Nr/(“+l)). Thus they must all be Nll(;+r) U(1).
In view of the rate of growth of A, , . . . , A, , the lower bound for W can be rewritten as
For purposes of comparison, we give the first few values of these sequences.
1 5.656 6
2 15.119 16
3 32.000 36
4 60.628 76
5 107.756 156
6 166.43 I 316
A similar result can be obtained when the number of stages is not constrained. The cost
is 2Vlf”(l), but N(log N)w(l). We shall not do this, since there are better ways of exploiting
Criterion 4.2 when the number of stages is large. The best way currently available is a
recursive method due to Cantor (51, which yields non-blocking networks with cost
O(N(log N)o), where /3 = 2.269... is the unique positive root of the transcendental
SWITCHING NETWORKS 155
LEMMA 4.4. If F is an H-access network with I inputs and J outputs and 8’ is a non-
blocking network with I’ inputs and J’ outputs, then F x 5’ is an HI-access network, where
H’ =: HJ’ - I(/’ - 1).
The proof can be found in Cantor [5] (L emma 4.1). (An H-access network is, in
Cantor’s terminology, a network of type T( 1, H).)
It is easy to show, using Lemma 4.4 inductively, that Vfi is an Hh-access network with
inputs and
outputs, where
The criterion thus asserts that H, is more than half Jk , that is, that Vk is a majority-access
network. u
THEOREM 4.6. The minimum possible cost of a (2k + I)-stage sandwich with at least N
Z&es that satisfies Criterion 4.5 is asymptotic to -rkN1+lVk+ll, where
7k = 4(k + 1)2(+)l/(k+l).
fl A, M 2 NT
i lQ<k >
W> 2 C Aj+M N.
l&$ >
As in the proof of Theorem 4.3, this implies that A, ,. . ., A, and M must all be N1l(k+l) U( 1).
The second constraint can be rewritten as
ABl> 21;<k(l~<j%)(
1 “j8,l)*
In view of the rate of growth of A, ,..., A, this implies
$22 c
lQQ lrJ<j 2 u(N-l’(k+l))-
1 ( 1
w3 2(l~<k(lrJj~))(21F,.(l~j~)A~
\ +(l~k~)M)Nu(N-l/(k+l)).
This can be minimized by elementary calculus; the calculations are simplified by elimin-
ating B, ,..., B, in favor of the new variables 8, ,..., iSkdefined by
aji= n Bi.
l<&gj Ai
SWITCHING NETWORKS 157
The result is
w > 4(k + I)2 (~)l”k+l) U(N-‘I’k+“),
Q = [(N/S)lIfk+l)].
1 5.656 5.656
2 18.000 15.119
3 38.054 32.000
4 65.975 60.628
5 101.823 107.756
6 145.627 166.43 1
We see that TVis no better than uk until k = 5, which corresponds to 11 stages, a value
just beyond those used in current practice. Once the crossover occurs, however, 7k
quickly asserts its advantage over ulc, since the former grows quadratically, and the
latter exponentially. This ultimate advantage reveals itself in
THEOREM 4.7. The minimum possible cost of a sandwich (of any number of stages) with
at least IV lines that satisjies Criterion 4.5 is asymptotic to l(iN(log, N)2.
we obtain
we have
SWITCHING NETWORKS 159
A!i! = [N/5”],
THEOREM 5.1. There are non-blocking networks with at least N lines and cost asymptotic
to 90N log, N.
The first step of the proof is a lemma concerning “sparse crossbars”. (This felicitous
terminology is due to G. M. Masson.) A sparse crossbar is like a crossbar, but has only
a small fraction of the contacts. The lemma we need is
LEMMA 5.2. For every integer k > 0 there exists a sparse crossbar L&. with 3 * 3L inputs,
9 . 3” outputs, and cost at most 45 . 3k, such that any set of 3” inputs is joined by contacts
to at least 6 . 3” different outputs.
Proof. Let K = 3k. (The proof will not use the fact that K is a power of 3.) Let p be a
permutation on the set X = (0, l,..., 45K - l}. Fromp we obtain a sparse crossbar Y(p)
160 NICHOLAS PIPPENGER
having (0, I,..., 3K - I} as inputs and {0, I ,..., 9K - I> as outputs, and having a contact
joining the input (X mod 3K) to the output (p(x) mod 9K) for every x in ~6.
We shall say that a sparse crossbar Y’(p) is “good” if there do not exist a set J&’ of K
inputs and a set L%of 6K outputs such that every contact that joins an input in ~2 also
joins an output in L@; we shall say that it is “bad” otherwise. A good sparse crossbar
clearly satisfies the requirements of the lemma. We shall show that one exists by obtaining
an upper bound less than unity on the fraction of all permutationsp for which Y’(p) is bad.
Any set of JS?of K inputs corresponds to a set J of 15K elements of X, and any set 9
of 6K outputs corresponds to a set% of 30K elements of X. Every contact of Y(p) that
joins an input in .RYwill also join an output in 9 if and only if p sends every element of .Y
into $. Of the (45K)! permutations of S, there are [30KJI,,(30K)! that satisfy this
condition, where [NJs = N(N - 1) ... (N - R + 1). There are
possible choices for LL#‘.Thus an upper bound on the fraction of all permutations p for
which Y(p) is bad is
We observe that
since the number of ways of choosing 15K out of 45K objects is not less than the number
of ways of choosing K out of the first 3K, 3K out the next 9K, and 11K out of the last
33K. Thus the fraction of bad permutations is at most
This is easily shown to be less than unity for all K by use of Stirling’s formula. 1
The next step of the proof is to combine sparse crossbars to obtain majority-access
networks. For technical reasons, a property stronger than majority-access will be used.
We shall prove
LEMMA 5.3. For every integer k 2 0 there exists a 5 * 3k-accessnetwork & with 3k
inputs, 9 . 3” outputs, and cost at Most (45k + 5) 3”.
SWITCHING NETWORKS 161
Proof. There is certainly a 5-access network &, with 1 input, 9 outputs, and cost at
most 5. We shall assume that Yk exists and build &+r , We do this by taking 3 copies
of Y,, and identifying their outputs with the inputs of the sparse crossbar Y;,+i , as shown
in Fig. 4.
3k
k+l
3 3k
3k
The 3 copies of Y,, each have cost at most (45k + 5) 3”, and Yti+.r has cost at most
45 . 3L+1.Thus Yk+l has cost at most 3(45k + 5) 3k + 45 * 3”+l = (45(K + 1) + 5) 3”l.
It remains to show that an idle input of Yti+l has accessto at least 5 . 3kl-1 outputs.
Consider first only those calls in the same copy of & . Despite these calls, the idle input
has access to at least 5 . 3k inputs of P$+r , by definition of Yk. Consider now the calls
in the other 2 copies of Yk . There are at most 2 +3” such calls, since each copy has only
3” inputs. Thus, despite these calls, the idle input has access to at least 3 3” = 3”‘+l
inputs of &r . By definition of Yk+, , these 3’i+1 inputs of S$+r are joined by contacts
to at least 6 . 3’i+1 different outputs of Yk+l . At most 3”f1 of these outputs can be busy,
since Yt+l has only 3”+1 inputs. Thus an idle input has accessto at least 5 . 3”+l outputs. 1
We are now ready for
Proof of Theorem 5.1. Consider the network
4Yk(M) = Yk x x g(M).
We set
k = llog, N - 4 log, log, N + log, $1,
M = [N/3’“].
Then the number of lines is
3”M3N
and
U < 90Nlog, N + O(N(log N)‘12).
This completes the proof. 1
That non-blocking networks can be built with cost O(N log N) was first proved by
Bassalygo and Pinsker [2]. Their method, with (Y= l/3, /I = 2/3, and K = 4, gives the
asymptotic formula 66N log, N. The preceding theorem, which is proved by a modifica-
tion of their method, yields a coefficient of 90 rather than 66 log, 3 = 104.607... . This
result is definitly capable of still further improvement by consideration of sparse crossbars
which are “irregular.” The proof of the analog of Lemma 5.2 then becomes very com-
plicated, however, and the improvement is very slight.
A lower bound asymptotic to 3N log, N, applicable to both rearrangeable and non-
blocking networks, is attributed to R. L. Dobrushin by Bassalygo and Pinsker [2]. A
proof can be found by specializing the results in Pippenger [II] to E = 0; in fact, it
follows from Lemma 3 and the obvious inequality H(S) > log, n! = n log, n + O(n).
REFERENCES