Systems & Control Letters 35 (1998) 291–299

Regularizability of complex switched server queueing networks

modelled as hybrid dynamical systems
Andrey V. Savkin ∗
Department of Electrical and Electronic Engineering, University of Western Australia, Nedlands, WA 6907, Australia

Received 11 March 1998; received in revised form 18 August 1998; accepted 25 August 1998

Abstract
The paper considers the problem of the qualitative analysis of complex switched server queueing networks. Such networks
can be used to model various exible manufacturing, communications, and computer systems. We introduce the concept of
regularizability for such systems and obtain a necessary and sucient condition for a switched server queueing network to
be regularizable. c 1998 Elsevier Science B.V. All rights reserved.

Keywords: Control of networks; Hybrid dynamical systems; Queueing networks; Flexible manufacturing systems;
Logical-di erential equations

1. Introduction furnace or air-conditioner is modelled as a continuous-

time system. Some other examples include vehicle
Hybrid dynamical systems (HDS) have attracted transmission systems and stepper motors, computer
considerable attention in recent years (see, e.g. disk drives, robotics systems, high-level exible man-
[7, 1, 4, 6, 2]). In general, HDS are those that com- ufacturing systems, intelligent vehicle=highway sys-
bine continuous and discrete behavior and involve, tems, sea=air trac management systems as well as
thereby, both continuous and discrete state variables. various systems with relays, switches, and hysteresis.
In many cases (but not always), such systems operate In fact many problems facing control engineers, com-
as follows. While the discrete state remains constant, puter scientists, and mathematicians as they seek to
the continuous one obeys a de nite dynamical law. use computers to control complex physical systems,
Transition to another discrete state implies a change naturally t into the HDS framework.
of this law. In its turn, the discrete state evolves as The class of HDS introduced in the paper, consists
soon as a certain event occurs with both the evolution of complex switched server queueing networks. Some
and the event depending on the continuous state. special classes of such networks were introduced in
A typical hybrid system is a logical discrete-event [5] to model exible manufacturing systems. More-
decision-making system interacting with a continuous over, such networks may be useful to model various
time process. A simple example is a home climate- computer and communications systems, especially
control system. Due to its on–o nature, the thermostat systems with time-sharing schemes. Unlike the papers
is modelled as a discrete-event system, whereas the [5, 3], we consider systems with time-varying arrival
rates. Such a model is much more realistic, es-
∗ Tel.: 00618 9380 1621; fax: 00618 9380 1065; e-mail: pecially in the case of computer or communication
[email protected]. networks.

0167-6911/98/$ – see front matter c 1998 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 7 - 6 9 1 1 ( 9 8 ) 0 0 0 6 9 - 3
292 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299

A network studied in this paper is controlled by is called a path on the graph G. The edge ( 1 ; 1 ) is
several servers. The locations of these servers are called the rst edge of this path, and the edge ( m ; m )
control variables. It is known that even very simple is called the last edge of this path.
switched single server systems can exhibit chaotic ir-
regular unpredictable behavior [3]. Such a behavior Deÿnition 2. An edge ( ; ) ∈ E of an oriented graph
is unacceptable for most real systems. Our aim is to G is said to be non-isolated if there exists a path con-
nd a feedback switching policy which guarantees the taining this edge with the rst edge of the form (∞; ·)K
boundedness and a regular predictable behavior of all and the last edge of the form (· ; ∞).
the trajectories of the network.
For the class of HDS under consideration, we intro-
duce the concept of regularizability. Intuitively, this In this paper, we will assume that the graph G sat-
means that there exists a feedback control policy such is es the following two assumptions.
that the following two conditions hold:
(i) All the trajectories of the closed-loop system are Assumption 1. For any vertix vi ∈ V there exists one
bounded. and only one edge ( ; ) ∈ E such that = vi and
(ii) In the case of constant average arrival rates, ∈ V̂ . In other words, this assumption means that any
the relevant dynamics of the closed-loop system is vertix has the only edge which departs from it.
determined by a periodic nite-state timed automaton,
and perforce must be eventually periodic.
Assumption 2. Any edge of the graph G is non-
Such systems exhibit a simple predictable behavior.
isolated.
We obtain a necessary and sucient condition of
regularizability. This algebraic condition is very sim-
ple and easy to check. For regularizable networks, we Assume that the set V is divided into l non-empty
nd an implementable in real time decentralized con- subsets S1 ; S2 ; : : : ; Sl where l6n. Also, we assume
trol policy which guarantees a regular behavior of the that for any edge ( ; ) ∈ E such that ∈ V̂ ; ∈ V ,
closed-loop system. there exists a bu er B[( ; )]. We refer to the contents
of a bu er as “work”, it will be convenient to think of
work as a uid, and a bu er as a tank. For any bu er
2. Network description B[(∞; vj )K ], work arrives in this bu er at an unknown
time-varying rate r[(∞; vj )K ](t)¿0. Furthermore,
Consider an oriented graph G which consists of for any set Si , there exists the only server si which
n vertices v1 ; v2 ; : : : ; vn and a nite number of edges. removes work at a given constant rate pi from any
Let V = {v1 ; v2 ; : : : ; vn } be then set of all the vertices, selected bu er B[( ; )] where ∈ Si , and sends work
and let E be the set of all the edges. The set E consists along the edge ( ; ˆ) which is unique according to
of edges of the following three kinds: Assumption 1. Then work arrives in the bu er
1. Edges (vi ; vj ) where 16i; j6n; i 6= j. Any edge of B[( ; ˆ)] if ˆ ∈ V , or leaves the system if ˆ = ∞.
this kind departs from one of the vertices and tends Note that one server may serve bu ers which corre-
to another. They are called interior edges. spond to di erent vertices. The notion of work will
2. Edges (∞; vj )K where 16j6n; 16K6Nj . Edges be formalized below.
of this kind come from outside the system. They are The location of any server is a control variable, and
called inputs. Here Nj is the number of edges which may be selected using a feedback policy.
come from outside the system to the vertix vj . Our description of the system has been phrased in
3. Edges (vj ; ∞) where 16j6n. Edges of this kind terms of work, uid, bu ers, and tanks. However, in
depart from vertices and go outside the system.They applications, work can represent a continuous approx-
are called outputs. We will consider networks with imation of the discrete ow of parts in a manufactur-
no more than one output for any vertix. ing system, or jobs in a computer system, etc.


We also introduce the set V̂ = {v1 ; v2 ; : : : ; vn ; ∞}. Let T ¿0 be a given time. Let r m [(∞; vj )K ]¿0 be
given known constants which are called the maximum
Deÿnition 1. Any sequence of edges ( 1 ; 1 ); ( 2 ; 2 ); arrival rates. In this paper, we consider the follow-
: : : ; ( m ; m ) such that m = 1; 2; : : : is an arbitrary num- ing class RT of arrival rates r[(∞; vj )K ](t). We say
ber; i ; i ∈ V̂ ; and i = i+1 for all i = 1; : : : ; m − 1 that r[(∞; vj )K ](·) ∈ RT if the following condition
 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299 293

holds: Also, it should be noted that the models for ex-

Z ible manufacturing systems introduced in [5] t in
1 NT
r[(∞; vj )K ](t) dt6r m [(∞; vj )K ] our framework. In [5], the network consists of single
T (N −1)T input–single output vertices and all the arrival rates
∀N = 1; 2; 3; : : : : (1) are constant.
Also, we will consider the following subclass of arrival Notation. Let (t) be any function of time. Then
rates RT0 ⊂ RT such that r[(∞; vj )K ]( · ) ∈ RT0 if and

(t + 0) = lim¿0; →0 (t + ).
only if r[(∞; vj )K ](·) ∈ RT and there exists a constant
For any 16i6l, introduce a symbolic variable qi (t)
r 0 [(∞; vj )K ] such that
to describe the state of the server si . Let Qi be the set of
Z the symbols corresponding to all the edges ( ; vj ) ∈ E
1 NT
r[(∞; vj )K ](t) dt = r 0 [(∞; vj )K ]

T (N −1)T such that vj ∈ Si , and Q̂i = {0} ∪ Qi . Then the function
∀N = 1; 2; 3; : : : : (2) qi (t) ∈ Q̂i is de ned for all times t as follows:
qi (t) := 0 if the server si does not work at time t,
0
Condition (1) obviously implies that r [(∞; vj )K ]6 and qi (t) := ( ; vj ) if the server si works with the bu er
r m [(∞; vj )K ] for any r[(∞; vj )K ](·) ∈ RT0 . Arrival B[( ; vj )] at time t.
rates from the class RT0 are called arrival rates with Furthermore, let x[( ; vj )](t) be the amount of work
a constant average. in the bu er B[( ; vj )] at time t. Then the network N
A critical feature of the network is that set-up times can be described by the following logical-di erential
are required whenever a server switches from one equations which formalize our network description.
bu er to another. Thus, if the server si has been remov- Let vj ∈ Si . Assume that there exists an edge
ing work from one bu er, and it is desired to switch of the form (∞; vj )K . Then the amount of work
instead to the removing of work from another bu er, x[(∞; vj )K ](t) in the bu er B[(∞; vj )K ] at time t is
then there is a minimum set-up time m i ¿0 during described by the following equation
which the server does not work. A set-up time can be
arti cially increased to achieve our control goal. In if qi (t) = (∞; vj )K then ẋ[(∞; vj )K ](t)
other words, set-up time i (t) of the server si is a con- = r[(∞; vj )K ](t) − pi ;
trol variable. However, condition
if qi (t) 6= (∞; vj )K then ẋ[(∞; vj )K ](t)
i (t)¿m
i ; (3)
= r[(∞; vj )K ](t): (5)
should be satis ed.
Let vj ∈ Si ; vs ∈ Sm and s 6= j, then the amount of work
We also assume that work incurs a xed transporta-
x[(vs ; vj )](t) in the bu er B[(vs ; vj )] at time t can be
tion delay [(vi ; vj )]¿0 when moving from one vertix
described as
to another.
if (qi (t) = (vs ; vj ) and qm (t − [(vs ; vj )]) 6= (vk ; vs )
Deÿnition 3. A collection for any k)

m
N = {G; T; r [(∞; vj )K ]; S1 ; : : : ; Sl ; p1 ; : : : ; pl ; then ẋ(vs ; vj )(t) = −pi ;
m m
1 ; : : : ; l ; [(vi ; vj )]} (4) if (qi (t) = (vs ; vj ) and qm (t − [(vs ; vj )]) = (vk ; vs )
is called a switched server queueing network. for some k)
then ẋ[(vs ; vj )](t) = pm − pi ;
In other words, any switched server queueing net-
work is de ned by an oriented graph, maximum ar- if (qi (t) 6= (vs ; vj ) and qm (t − [(vs ; vj )]) 6= (vk ; vs )
rival rates, server subsets, server set-up times, server for any k)
removal rates, and trasportation delays.
An example of a switched server queueing network then ẋ[(vs ; vj )](t) = 0;
is shown in Fig. 1. It is obvious that this network if (qi (t) 6= (vs ; vj ) and qm (t − [(vs ; vj )]) = (vk ; vs )
satis es Assumption 1 and 2. Here l = 3; S1 = {v1 ; v2 };
for some k)
S2 = {v3 ; v4 ; v5 } and S3 = {v6 ; v7 ; v8 }. This network has
6 inputs and 2 outputs. then ẋ[(vs ; vj )](t) = pm : (6)
294 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299

Fig. 1. A switched server queueing network.

The state of the network at time t can be described It should be pointed out, that the closed-loop system


by the vectors q(t) = [q1 (t); q2 (t); : : : ; ql (t)] and (5) – (7) is a system of logical-di erential equations


x(t) = {x[( ; vj )](t)} where ( ; vj ) ∈ E; vj ∈ V and with time delays. From a mathematical viewpoint, it
∈ V̂ . implies that the state space of such a system is in nite-
Our goal is to design sets Ti (B[( ; vj )] 7→ B[( ; vs )]) dimensional. We will consider the following physi-
and functions Fi : ([q(·); x(·)]|t0 ) 7→ [m cally natural initial conditions
i ; ∞) where
16i6l; ( ; vj ) ∈ Qi ; ( ; vs ) ∈ Qi . These sets and
x(0) = x0 ; q(t) = 0 ∀t¡0; (9)
functions de ne a feedback policy of the form
where x0 is a given vector with non-negative compo-
if (qi (t) = ( ; vj ) and [q(·); x(·)]|t0 ∈ Ti (B[( ; vj )] nents. If condition (9) holds, then it is obvious from
Eqs. (5), (6), that any solution of the closed-loop sys-
7→ B[( ; vs )])); tem (5) – (7) does not depend on the continuous state
 
i (t) := Fi ([q(·); x(·)]|t0 ); variable x(t) for t¡0. Moreover, x(t) is continuous
  and q(t) is piecewise-constant and left-continuous.
then  qi (tˆ) := 0 ∀tˆ ∈ (t; t + i (t)];  : (7)
qi (t + i (t) + 0) := ( ; vs )
Deÿnition 4. The closed-loop system (5) – (7) is said
to exhibit a regular behavior with the period T , if the
In other words, our feedback control policy is a rule following conditions hold:
for switching servers from one bu er to another. When (i) For any solution [q(t); x(t)] to the system with
such a “discrete event” occurs, we can choose a set- arrival rates from the class RT and initial condi-
up time i (t). However, the minimum set-up time re- tion (9), the vector function x(t) is bounded on
quirement (3) must be satis ed. Also, the following [0; ∞).
natural non-negativity requirements must be ful lled: (ii) For any solution [q(t); x(t)] to the system with ar-
rival rates from the class RT0 and initial condition
x[( ; vj )](t)¿0 ∀( ; vj ) ∀t¿0: (8) (9), the following two conditions are satis ed:
 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299 295

(ii-1) There exists a vector function q̂(t) such that which is called the total arrival rate of the server si .
q̂(t + T ) = q̂(t) ∀t¿0 and Consider the network shown in Fig. 1 and determine
the total arrival rate R1 of the server s1 . We have
lim mes{t ∈ [(N − 1)T; NT ]: q(t) 6= q̂(t)} = 0:
N →∞ R[v2 ] = R[(v3 ; v2 )] = r m [(∞; v8 )1 ] + r m [(∞; v8 )2 ]:
(ii-2) Furthermore,
lim [x((N − 1)T ) − x(NT )] = 0: R[v1 ] = r m [(∞; v1 )1 ] + R[(v6 ; v1 )]
N →∞
= r m [(∞; v1 )1 ] + r m [(∞; v6 )1 ] + r m [(∞; v8 )1 ]
Here mesH denotes the Lebesque measure of the
set H . + r m [(∞; v8 )2 ]:

Remark. Condition (i) of the above de nition is sim- Therefore,

ilar to the stability requirement from [5]. The second
R1 = R[v1 ] + R[v2 ] = r m [(∞; v1 )1 ] + r m [(∞; v6 )1 ]
condition implies that the system exhibits a regular
eventually periodic behavior if its arrival rates have + 2r m [(∞; v8 )1 ] + 2r m [(∞; v8 )2 ]:
constant average.
Also, let ni be the total number of bu ers B[( ; vj )]
Deÿnition 5. A switched server queueing network (4) where vj ∈ Si , or the total number of bu ers served
is said to be regularizable if there exists a feedback by the server si . For the network shown in Fig. 1,
policy of the form (7) such that the closed-loop system n1 = 3; n2 = 4 and n3 = 5.
(5) – (7) exhibits a regular behavior with the period T . Now we are in a position to present the main result
of this paper.

Theorem 6. Consider a switched server queueing net-

work
3. The main result
N = {G; T; r m [(∞; vj )K ]; S1 ; : : : ; Sl ; p1 ; : : : ; pl ;
Consider a switched server queueing network (4).
Let 16i6l, and let ( ; vj ) be any edge of the graph m m
1 ; : : : ; l ; (vi ; vj )};
G such that vj ∈ Si . Let W [( ; vj )] be the set of all
described by the logical-di erential Eqs. (5), (6).
paths w on the graph G with the rst edge of the form
Suppose that Assumptions 1 and 2 hold. Then, the
(∞; ·)K and with ( ; vj ) as the last edge. Assumption 1
following statements are equivalent:
implies that this set is not empty for any edge. Also,
(i) The switched server queueing network N is
let r m (w) be the maximum arrival rate r m [(∞; ·)K ]
regularizable.
of the rst edge (∞; ·)K of the path w. Furthermore,
(ii) For any i = 1; 2; : : : ; l, the following inequalities
introduce the following constants:
hold:
X
R[( ; vj )] =


r m (w); (10) pi ni m
i
Ri ¡pi ; 6T: (13)
w∈W [( ; vj )] pi − Ri

which is called the arrival rate of the bu er B[( ; vj )],

Proof. ((ii) ⇒ (i)) Suppose that condition (ii) holds.
and
Introduce the following feedback policy. The server


X si works with ni bu ers. Denote the corresponding
R[vj ] = R[( ; vj )]; (11)
edges E1 ; E2 ; : : : ; Eni in an arbitrary order. Form the
( ; vj )∈E]
following cyclic sequence of these edges
which is called the total arrival rate of the vertix vj . E1 7→ E2 7→ · · · Eni 7→ E1 : (14)
Moreover, introduce the constant
X Let ( ; vj ) be any edge such that vj ∈ Si or one of the


Ri = R[vj ]; (12) edges E1 ; : : : ; Eni . Then next[( ; vj )] is the edge which
vj ∈Si is the next to ( ; vj ) in the cyclic sequence (14) (or
296 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299

the symbol corresponding to the edge). Let [x(t); q(t)] and

be a trajectory of the system (5); (6). Introduce the
R[( ; vj )]T
functions i [x(·)|t0 ; q(·)|t0 ](t) as follows: Tk [( ; vj )] −tk [( ; vj )]6 ∀k = 1; 2; 3; : : : :
pi
i [x(·)|t0 ; q(·)|t0 ](t) = i [qi (·)|t0 ] (20)


= inf {t0 6t: qi (s) = qi (t) ∀s ∈ (t0 ; t]}: (15) Furthermore, from Eqs. (19) and (16) we obtain that
the server si always makes the switching cycle (14)
In other words, i (t) is the time when the server si had for the time
started to work with the current bu er. Introduce the
X
following constants: (tk [( ; vj )] − tk [next[( ; vj )])

(pi − Ri )T ( ; vj )∈Si
M
i = ∀i = 1; 2; : : : ; l: (16) X  R[( ; vj )]T 
pi ni Ri T
= + M
i = + ni M
i = T:
We will consider the following feedback policy: pi pi
( ; vj )∈Si

(21)
if qi (t) = ( ; vj ) and x[( ; vj )](t) = 0
Furthermore,
!!
R[( ; vj )]T x[( ; vj )](kT )
or t − i [qi (·)|t0 ] =
pi = x[( ; vj )]((k − 1)T ) + Aa ((k − 1)T; kT )
 
R[( ; vj )]T − As ((k − 1)T; kT ) ∀k = 1; 2; : : : : (22)
i (t):=M + −t + [q
i i (·)|t
];
 i
pi 0 
then  ˆ ˆ
 qi (t ) := 0 ∀t ∈ (t; t + i (t)];
:
 Here for any T1 ¡T2 ; Aa (T1 ; T2 ) is the amount of work
qi (t + i (t) + 0) := next[( ; vj )] which has arrived in the bu er B[( ; vj )] over the time
interval (T1 ; T2 ], and As (T1 ; T2 ) is the amount of work
(17)
which has been removed by the server si from the
Now we prove that the closed-loop system (5); (6), bu er B[( ; vj )] over the time interval (T1 ; T2 ]. More-
(17) exhibits a regular behavior with the period T . over,

Claim. For any trajectory [q(t); x(t)] of the system Aa ((k − 1)T; kT )
(5); (6); (17) with arrival rates from RT and initial 6r m [(∞; vj )K ]T = R[( ; vj )]T if = ∞:
condition (9), there exists a constant c¿0 such that
In the case ∈ V , it follows from Eq. (21) that
x[( ; vj )](t)6c ∀t¿0 ∀( ; vj ): (18)
Aa ((k − 1)T; kT )
X
Indeed, introduce for this trajectory the sequences of = (Tk [( ; )] − tk [( ; )])pq ;
times tk [( ; vj )] and Tk [( ; vj )] such that tk ¡Tk ; qi (t) ∈Vˆ ; ( ; )∈E
= ( ; vj ) for t ∈ (tk [( ; vj )]; Tk [( ; vj )] where k = 1; 2;
3; : : :, and qi (t) 6= ( ; vj ) for all other t. In other words, where q is the index such that ∈ Sq . From this and
(tk [( ; vj )]; Tk [( ; vj )]] is the sequence of the time Eq. (20), we obtain that
intervals during which the server si removes work
Aa ((k − 1)T; kT )6R[ ]T = R[( ; vj )]T:
from the bu er B[( ; vj )]. Furthermore, for the edge


( 1 ; vs ) = next[( ; vj )], introduce the corresponding Hence, we have proved that
sequences tk [( 1 ; vs )] and Tk [( 1 ; vs )]. Then, it follows
Aa ((k − 1)T; kT )6R[( ; vj )]T: (23)
immediately from Eqs. (15)–(17), that
R[( ; vj )]T Furthermore, Eq. (20) implies that if x[( ; vj )]((k −
tk [( 1 ; vs )] − tk [( ; vj )] = + M
i 1)T )¿R[( ; vj )]T , then the equation
pi
∀k = 1; 2; 3; : : : (19) x[( ; vj )](t) = 0
 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299 297

cannot hold for any t ∈ ((k −1)T; kT ], hence it follows has not yet arrived into the corresponding destination
from the switching rule (17), that bu ers because of the transportation delays. Then
R[( ; vj )]T Aa (0; NT ) = R0 [( ; vj )]NT;
Tk [( ; vj )] − tk [( ; vj )] = :
pi N
X
Therefore As (0; NT ) = pi (Tk [( ; vj )] − tk [( ; vj )]);
k=1
As ((k −1)T; kT ) = R[( ; vj )]T Ad (NT )6T0 lp; (27)
if x[( ; vj )]((k −1)T)¿R[( ; vj )]T: (24) where T0 is de ned as
The relationships (22)–(24) imply, that

T0 = max [(vs ; vj )];
vs ; vj ∈V; (vs ; vj )∈E
x[( ; vj )](kT )6x[( ; vj )]((k − 1)T )


if x[( ; vj )]((k − 1)T )¿R[( ; vj )]T: l is the number of servers, and p = max {p1 ; p2 ; : : : ;
pl }. Furthermore, substituting Eq. (27) into Eq. (26),
Hence, condition (18) holds. This completes the proof we obtain
of the claim. Condition (i) of De nition 4 follows N
X
immediately from Claim. pi (Tk [( ; vj )] − tk [( ; vj )])
Now we prove condition (ii) of De nition 4. For k=1
any arrival rate from class RT0 and any bu er B[( ; vj )], X
introduce a constant R0 [( ; vj )] by the rule (10) with ¿R0 [( ; vj )]NT − x[( 1 ; vs )](NT )
r m (w) replaced by r 0 (w) where r 0 (w) is de ned by ( 1 ; vs )∈E[( 1 ; vs )]
Eq. (2). Now we prove that for arrival rates from the X
class RT0 , the following condition holds: + x[( 1 ; vs )](0) − T0 lp
( 1 ; vs )∈E[( 1 ; vs )]

lim (Tk [( ; vj )] − tk [( ; vj )]) ∀N = 1; 2; 3; : : : :

k→∞

R0 [( ; vj )]T This and condition (18) imply that there exists a con-
= ∀i = 1; : : : ; l ∀( ; vj ): (25) stant c0 such that
pi
N
X
Indeed, let E[( ; vj )] ⊂ E be the set of all the edges of
pi (Tk [( ; vj )] − tk [( ; vj )])¿R0 [( ; vj )]NT − c0
the graph G of the form ( 1 ; vs ) where 1 ∈ V̂ ; vs ∈ V
k=1
such that G contains a path with ( 1 ; vs ) as the rst
∀N = 1; 2; 3; : : : : (28)
edge and with ( ; vj ) as the last edge. Then we have
X It is obvious, that Eqs. (28) and (20) imply Eq. (25).
x[( 1 ; vs )](NT ) We have proved that switching times of all the servers
( 1 ; vs )∈E[( 1 ; vs )] are asymptotically T -periodic. Hence condition (ii-1)
X of De nition 4 holds. Furthermore,
= x[( 1 ; vs )](0)
( 1 ; vs )∈E[( 1 ; vs )] x[( ; vj )](t + T )
a s
+A (0; NT )−A (0; NT )−A (NT ) d = x[( ; vj )](t) + Aa (t; t + T ) − As (t; t + T ) ∀t¿0;
∀N = 1; 2; 3; : : : ; (26) (29)

where Aa (0; NT ) is the amount of work which has where Aa (t; t + T ) is the amount of work which has
arrived into the subgraph E[( ; vj )] from outside the arrived in the bu er B[( ; vj )] over the time interval
network over the time interval (0; NT ]; As (0; NT ) is (t; t + T ], and As (t; t + T ) is the amount of work which
the amount of work which has been removed by the has been removed by the server si from the bu er
server si from the bu er B[( ; vj )] over the time inter- B[( ; vj )] over the time interval (t; t + T ]. Moreover,
val (0; NT ], and Ad (NT ) is the amount of work which
Aa ((k − 1)T; kT )
has been already removed by one of the servers from
one of the bu ers B[( 1 ; vs )]; ( 1 ; vs ) ∈ E[( ; vj )] but = r 0 [(∞; vj )K ]T = R0 [( ; vj )]T if = ∞:
298 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299

In the case ∈ V , it follows from Eq. (21) that De nition 4 imply that

Aa ((k − 1)T; kT ) lim [Aa ((k − 1)T; kT )[( ; vj )]

k→∞
X − As ((k − 1)T; kT )[( ; vj )]] = 0: (33)
= (Tk [( ; )] − tk [( ; )])pq ;
∈Vˆ ; ( ; )∈E Moreover, we have

where q is the index such that ∈ Sq . From this and Aa ((k − 1)T; kT )[( ; vj )]
Eq. (25), we obtain that = R[( ; vj )]T ∀k = 1; 2; : : : : (34)

lim Aa ((k − 1)T; kT ) = R0 [ ]T = R0 [( ; vj )]T: (30) On the other hand,

k→∞
Asi ((k − 1)T; kT ) = pi Tk [( ; vj )]; (35)
On the other hand, Eq. (25) implies that
where Tk [( ; vj )] is the time during which the server si
s
lim A ((k − 1)T; kT )R [( ; vj )]T: 0
(31) works with the bu er B[( ; vj )] over the time interval
k→∞ ((k − 1)T; kT ]. Since the server si must work with ni
bu ers, condition (ii-1) of De nition 4 implies that
Hence, it follows from Eqs. (29)–(31) that condition
the server si makes at least ni switches on the time
(ii-2) of De nition 4 holds. This completes the proof
interval ((k − 1)T; kT ] for large k. This and condition
of this part of the theorem.
(3) imply that
((i) ⇒ (ii)) Assume that condition (i) of the theo-
rem holds and there exists a feedback policy of the

X
Tk = Tk [( ; vj )]6T − ni m
i (36)
form Eq. (7) such that the closed-loop system (5) –
( ; vj )∈E; vj ∈Si
(7) exhibits a regular behavior with the period T . Let
E[( ; vj )] ⊂ E denotes the set of all the edges of the for large k. Finally, conditions (33) – (36) imply that
graph G of the form ( 1 ; vs ) where 1 ∈ V̂ ; vs ∈ V such X
that G contains a path with ( 1 ; vs ) as the rst edge Ri T = R[( ; vj )]T 6pi (T − ni m
i )
and with ( ; vj ) as the last edge. Consider the follow- ( ; vj )∈E; vj ∈Si
ing constant arrival rates from the class RT , de ned as
for large k. This immediately implies conditions (13)
r[(∞; vj )K ](·) ≡ r m [(∞; vj )K ]. Let [x(t); q(t)] be any
and completes the proof of the theorem.
trajectory of the closed-loop system with these arrival
rates and initial condition (9). Then condition (ii-1)
The following corollary can be immediately ob-
of De nition 4 implies that
tained from the proof of Theorem 6.

X Corollary. If condition (ii) of Theorem 6 holds, then
lim  x[( 1 ; vs )](kT )
k→∞ the switched server queueing network N with the
( 1 ; vs )∈E[( 1 ; vs )]
 control policy deÿned by the relationships (14) – (17)
X exhibits a regular behavior with the period T .
− x[( 1 ; vs )]((k − 1)T )
( 1 ; vs )∈E[( 1 ; vs )] Remark. It should be pointed out that the feedback
a
= A ((k − 1)T; kT )[( ; vj )] control policy (14) – (17) is decentralized: the location
of the server si is selected based only on a quantized
− As ((k − 1)T; kT )[( ; vj )]; (32) observation of the current bu er state. Therefore, this
control policy is very easy to implement.
where Aa ((k − 1)T; kT )[( ; vj )] is the amount of
work which has arrived into the subgraph E[( ; vj )]
from outside the network over the time interval
((k − 1)T; kT ]; As ((k − 1)T; kT )[( ; vj )] is the amount 4. Illustrative example
of work which has been removed by the server si
from the bu er B[( ; vj )] over the same time inter- To illustrate Theorem 6, consider the switched
val. Furthermore, Eq. (32) and condition (ii-2) of server queueing network shown in Fig. 1. Assume
 A.V. Savkin / Systems & Control Letters 35 (1998) 291–299 299

that all the arrival rates of this network are constant Acknowledgements
and equal to r.
ˆ Furthermore, assume that all the server
removal rates are also equal and This work was supported by the Australian Re-
search Council.
p1 = p2 = p3 = 7r:
ˆ

Finally, assume that the server set-up times are equal References
too
[1] P.J. Antsaklis, J.A. Stiver, M. Lemmon, Hybrid systems
m m m
1 = 2 = 3 = :
modeling and autonomous control systems, in: R.L. Grossman,
A. Nerode, A.P. Ravn, H. Rishel (Eds.), Hybrid Systems,
We consider the following problem: Springer, New York, 1993.
[2] E. Asarin, O. Maler, A. Pnueli, Reachibility analysis of
To determine the minimum time T such that this
dynamical systems having piecewise constant derivatives,
network is regularizable with the period T . Theoret. Comput. Sci. 138 (1995) 35 – 65.
We apply Theorem 6. It follows from Fig. 1 and [3] C. Chase, J. Serrano, P. Ramadge, Periodicity and chaos from
Eqs. (10) – (12) that switched ow systems: Contrasting examples of discretely
controlled continuous systems, IEEE Trans. Automat. Control
R1 = 6r;
ˆ R2 = 4r;
ˆ R3 = 5r:
ˆ 38 (1) (1993) 70 – 83.
[4] A. Nerode, W. Kohn, Models for hybrid systems: automata,
Hence the rst of conditions (13) of Theorem 6 is topologies, controllability, observability, in: R.L. Grossman,
A. Nerode, A.P. Ravn, H. Rishel (Eds.), Hybrid Systems,
satis ed for any i = 1; 2; 3. Furthermore, we have Springer, New York, 1993.
[5] J.R. Perkins, P.R. Kumar, Stable, distributed, real-time
n1 = 3; n2 = 4; n3 = 5: scheduling of exible manufacturing=assembly=disassembly
systems, IEEE Trans. Automat. Control 34 (2) (1989)
The second of conditions (13) now can be re-written 139 –148.
as the following three inequalities: [6] A.V. Savkin, I.R. Petersen, E. Ska das, R.J. Evans, Hybrid
dynamical systems: robust control synthesis problems, Systems
28 35 Control Lett. 29 (2) (1996) 81– 90.
216T; 6T; 6T: [7] L. Tavernini, Di erential automata and their discrete
3 2
simulators, Nonlinear Anal. Theory Meth. Appl. 11 (6) (1987)
Therefore, the minimum time is T = 21. 665 – 683.