Escolar Documentos
Profissional Documentos
Cultura Documentos
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/2634653?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management
Science
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
Externalities, Tangible Externalities, and
Queue Disciplines
E xternalities are the (marginal) costs that a user of a common resource imposes on others.
We introduce the efficient measure of tangible externalities that are the costs that a user
imposes on others while being served. Then, for a single server queueing system under various
service disciplines, we compute the expected externalities and the expected tangible externalities
conditioning on the length of the service requirement.
(Externalities; Cognestion Costs; Queue Discipline)
Users of a common facility may impose costs on each is that various ways of managing a common facility
other, costs that are known as negative externalities or may result in different amounts of extemalities imposed
congestion costs. In many cases, these externalities are in by the same user. Hence, the resulting pricing mecha-
the form of delays that are imposed on users prior to nism can serve as an additional criterion for deciding
the completion of their tasks. This is, for example, the which management scheme to adopt. We next elaborate
case when airplanes take off from a runway, when com- on these issues while considering arrivals to an M / M / 1
muters cross a bridge, when jobs share a common CPU, queue under various service disciplines. In many cases
and when messages are routed through the same data our results are applicable for M / G / 1 queues, too.
network. These models and many others are examples Recall that M / M/ 1 stands for a single server queue
of queueing systems. There is hence a need to quantify with a Poisson arrival process and exponential service
this notion. This is in particular the case due to the re- requirement. The waiting room is unlimited, all random
cent growth in the use of the Internet. For this context variables involved are independent, and the system is
see Walrand and Varaiya (1996). considered to be in steady state. The entrance to service
The amount of externalities that a user imposes on discipline is not specified yet. Let X be the rate of arrival.
others is typically a function of his own requirement. Each arrival requires an exponential amount of service
Specifically, in the context of service systems, the longer with an expected value of ,p-1. Denote X /,u by p, which
his service requirement is (or its expected value, re- is referred to as the traffic intensity. Of course, 0 c p
spectively), the larger are the negative externalities (or < 1. Finally, M/G/1 queues are defined as M/M/1
their expected value, respectively) he imposes. Defining queues but without the assumption of exponential ser-
quantitatively the externalities can serve two purposes. vice time. In this case, too, p stands for the traffic inten-
First, a zero profit operator who charges users for the sity, namely AK1 where Kl is the expected service re-
quirement.
use of a common facility usually likes to do so in accor-
dance with the congestion costs that they impose on The standard definition for externalities that an in-
dividual imposes on others is the difference between the
others. This is, of course, on top of other charges such
as access charges, user charges, or service quality actual total queueing time of the others and the corre-
sponding value in a (simulated) realization in which
charges. See Walrand and Varaiya (1996) for details.
everything stays the same but without the presence of
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
the user in question. By saying that everything stays the time is now associated with one and only one customer.
same we mean that all other arrivals occur at the same In particular, the sum of tangible externalities across
instants and with the same workload requirements. For- customers equals the total queueing time. Hence, tan-
mally, we consider a system, as described above, that is gible externalities can be used as an efficient pricing
in a steady state at time 0, and at that time a customer mechanism for payment to be made by the customers.
with a fixed service requirement of x is added to the The above suggested mechanism, used as is, might
system. We analyze the externalities imposed by this seem a bit unfair for some customers. For example, in a
customer. See ?2 for details. For more on this measure FCFS system, a customer who is asked to pay a bundle
of externalities and its applications see Carlin and Park might claim that it was not his fault that many custom-
(1970), Dewan and Mendelson (1990), Dolan (1978), ers arrived during his service period: he was served
Hassin (1983), and Hassin and Haviv (1994), or Loch during a particular period in which the number of ar-
(1991), Mendelson and Whang (1990), and Newbery rivals was high by a mere chance or just because he was
(1990). delayed service by other customers who have been
We like to mention that another approach for the han-served earlier. Of course, other customers might be
dling of externalities appears in the literature which is more lucky. A better measure is the expected tangible
the marginal added waiting cost per unit of time to the externalities that a customer imposes as a function of
society due to the infinitesimal last arrival. Specifically, his (actual) service requirement. This measure has the
if W(A) is the expected waiting time when the arrival same expectation (over all customers) as the tangible
rate is X (and hence the total waiting cost per unit of externalities, but its variance is smaller.
time is XW(A)), one looks at Besides variance reduction, there is an additional ad-
vantage to the expected tangible externalities over tan-
dA (AW(A)) = W(A) + X d W(A). gible externalities. This is the simplicity of using them
as a pricing mechanism: there is no need to monitor how
For this approach see Mendelson and Whang (1990), many customers are in the system upon each service
Dewan and Mendelson (1990) or Walrand and Varaiya commencement, how many arrive, and at exactly which
(1996, p. 386). instants during this service period. All that is required
Externalities as defined above suffer from the lack of is to monitor the length of the service requirements
efficiency, namely, their sum does not coincide with the themselves. Actually, the queue operator can place a
total queueing time. Actually, this sum overestimates sign that tells each customer how much to pay as a func-
the total queueing time. For example, consider a single tion of his service requirement.
server, first come first served (FCFS) queue where three Before describing the content of the rest of the paper,
customers A, B, and C with identical service require- we need the following terminology and results.
ments are the first to arrive, and in this order, at an A queueing discipline is said to be strong if the order
empty queue. Also, assume that the arrival times of cus- in which customers receive service is not a function of
tomers A and B (almost) coincide, while customer C their actual service requirements (i.e., future values
arrives when customer A is ready to clear the system. from the operator's point of view). Put differently, the
Hence, the externalities of both customers A and B in- queue operator is not informed of these values while
clude the waiting time of customer C. We then conclude deciding who gets service when. Also, a queueing sys-
that externalities cannot be used as a pricing mechanism tem is said to be work-conserving if the resulting process
for charging customers for their share of the total of total (residual) service requirement due to customers
queueing cost. currently in the system (also known as the virtual wait-
We suggest here an alternative measure. For each ar- ing time) coincides with that of a FCFS system having
rival of a customer we associate what we call a tangible the same characteristics. Consider now an M / M /1
externality. This is defined as the total queueing time queueing system with any entrance discipline so long
added to the others while this customer is in service. This as it is strong and work conserving. There, for p as the
measure is efficient since any part of the total queueing traffic intensity and for ,-1 as the expected service
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
requirement, the expected externalities per arrival (in a for the variability of the prices E(C I x). Here again, the
steady state) equals p / (,(1 - p)2). (See Proposition4.1.) values for E(C I x) (and hence, of course, for
This value is larger than its expected queueing time, p / Var(E(T I x))), for the last two queueing disciplines co
(,(1 - p)). In other words, the sum of expected exter- incide. Finally, for an M / M /1 queue under all three
nalities overestimates the total cost by a ratio of (1 cases we considered,
-p)-1. E(T x) = 1
The rest of the paper contains our detailed analysis
E(CIx) 1 -p' ?<o
of expected externalities and of tangible externalities
and it is organized as follows. In ?2 we derive the ex- However, this ratio does not hold for all strong work-
pected externalities that a customer imposes on others conserving disciplines. A counterexample and some
given his own service requirement. This is done sepa- other remarks are given in our final section, ?4.
rately for the following three cases:
1. Any strong and work-conserving discipline which
2. The Expected Externalities
does not allow preemption. Among them, FCFS, ran-
dom order queues and last come first served without 2.1. General Results
preemption (LCFS-HOL).1 In this section we compute the expected externalities
2. Last come first served with preemptive resume imposed by an x-customer on the others, E(T I x), and
(LCFS-PR). the resulting variance, Var(E(T l x)), in an M / G/1
3. Processor sharing (PS). queue under the first and the second disciplines and in
Detailed description of the last two disciplines are an M / M / 1 queue under the third discipline. However,
given in the next section. some notation and preliminary results are needed first.
Call a customer whose service requirement is x an x- Our analysis is conditioned on an arrival of a cus-
customer and let T I x be the (random) externalities that tomer with a service requirement of x at time 0. To be
he imposes on others. Then, let E(T I x) be the corre- more exact, we condition on an arrival of a customer
sponding conditional expected value and let Var(E(T I x))with service requirement in the interval (x - c, x + e)
be the variance of the conditional expected values. Note during the interval (- e, e) and consider the limit as
that Var(E(T I x)) (as oppose to Var(T)), for example) is e -- 0. Note that since the arrivals during disjoint inter-
the right measure for the variability for the prices paid vals are independent, the x-customer observes a Poisson
by customers in the case that the price mechanism of arrival process with independent service requirements
externalities is adopted. In the next section we compute distributed according to some distribution function G.
these values for an M / G/1 queue for the first two of We next refer to a simulated realization in which all
the above mentioned cases and for an M / M / 1 queue random variables get the same values but without the
for the last. As it turns out, the values E(T I x) (and presence of the x-customer. The case with him will be
hence, of course for Var(E(T I x))), for the last two cases referred to as the original case. Let M denote the number
coincide. of customers (other than the x-customer) who exit the
Section 3 considers tangible externalities. Specifically, original system from the (first) service commencement
of the x-customer and until the first time the server is
let C I x be the (random) tangible externalities incurred
by an x-customer. Our first interest is in E(C I x), whichidle. For 1 c i c M, let ri be the total amount of service
is the conditional expected tangible externalities im- received by the x-customer until the i-th departure time
posed by an x-customer. In ?3 we compute E(C I x) and
(in the original system). Finally, let Ii be the total idle
Var(E(C I x)) for the three abovementioned cases for M /time (in the simulated system) of the server between the
arrival of the x-customer and the arrival of the ith cus-
G / 1 queues. As said above with regard to Var(E(T I x)),
we can say here that Var(E(C I x)) is the right measure
tomer in the simulated system.
We next exemplify the notation just introduced. Sup-
pose that x = 5 and that the tagged x-customer arrives
1 HOL stands for head of the line. at an empty FCFS queue. Suppose a customer arrives at
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
units. At time 12, after the service requirements of these served before the tagged customer, or x, for those who
4 customers are met, the queue is empty again for the are served after his departure. This completes our proof
first time. Then the externalities imposed by the tagged for (1), case (i). For case (ii), Equation (4) does not hold
customer are (7 - 3) + (10 - 6) + (12 - 11) = 9. Clearly, as is. However, we can assume that all service between
M = 3. Since the other three customers are served after two consecutive exits from the simulated system was
given to the next to depart customer. This statistically
the tagged customer is served completely, ri = 5 for all
changes nothing due to the memoryless property of the
of them. Finally, I, = 1, I2 = 1, (the length of the first
time period before the arrival of the first customer) and service requirement, but now (4) holds for case (ii) as
I3 = 4 (the above time plus the idle time between the well.
departure of the second customer in the simulated sys- Next we prove (2) (for both cases (i) and (ii)). Let Ji
tem and the arrival of the third). Hence, we obtain that be the length of the ith idle period in the simulated sys-
the externalities are tem occurring during the time interval between the first
service commencement of the x-customer and the first
M M
time in which the server is idle in the original system,
Externalities = T i Ii.
i=l i=l for i, 1 ? i ? L, where L denotes the number of such
idle periods. Note that L = inf{j: ELji i 2 x}, that L
This is an example of the following proposition.
2 1 and that IM = EL-1 Ji. Clearly, Jl, . . ., JL are inde-
PROPOSITION 2.1. For (i) any strong and work- pendent and exponentially distributed random vari-
conserving M / G / 1 queue without preemption, or ables with expected value of A-1. Thus, L - 1 follows a
(ii) any
strong and work-conserving M / M / 1 queue, Poisson distribution with parameter Xx. In particular,
E(L) = Xx + 1. Note that the time interval between two
E(Tlx) =E , Ti IiF (1) idle periods in the simulated system corresponds to a
i=l i=l
__ 2
M L
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
? j c L, are exponentially distributed, L is independentNote that the time between his departure and the end
of 4=1 Ji. Hence, of the busy period is statistically identical to the sum of
M busy periods, each of which corresponds to one of the
E , Ii = (1 - p)-1[X-1(Xx + 1)2 customers present in the system at that instant. Since
the expected number of customers to be served in a busy
period is 1 / (1 - p), we conclude that
_-1 2 E{L(L + 1)}] (5)
E(TIx)= X,
E(TIx) x+ ( 2 x (7)
2 ( - p) 2(1l-p)2
and
and
2 2 p22.
Var(E(T I x)) = ) (K2 - K).
Var(E(T x)) = 4(1 - (K4 - K2
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
of nA/l(n + 1) on the others). Note that in order to that this is so also for r E (0, x]. Our proof follows Ross'
receive A units of service, a customer has to spend (n proof exactly with the understanding of the increment
dr as the service increment for the x-customer and not
+ 1 )A units of time in the system. The processor sharing
model corresponds to the case where A goes to zero. A as the real time increment. Specifically, denote the dis-
way to interpret this model is by assuming that a cus- tribution of the service requirement of a customer by F
tomer has his own watch which runs only while he is and the corresponding probability density function by
in service. In particular, an x-customer leaves when his f. If there are n additional customers in the system at
watch says x. In comparison with the natural clock, this service time T for the x-customer, let y = (yi, ..., yn), y
watch progresses at a variable pace: when n additional = y(r), be the vector of their completed workloads (yi
customers are in the system, his watch moves (n + 1) _ . yn), let y-i = (yl, . . ., yi-1, Yi+l, *., yn) and
times slower. finally, let yo = (0, y). Clearly, y(r), r E (0, x), is a Mar-
Before stating our main result for this subsection we kov process. In particular, the transition rates q(, )
need the following lemma. It is stated and proved for which correspond to a change in the number of custom-
ers in the system are
the M / G / 1 case (rather than the simpler M / M / 1 case
dealt with) since the general result is needed in the next
section. q(y, y-j) - lF(y )
1 - F(yi)
LEMMA 2.1. For an MI G / 1-PS queueing system (under
and
stationary conditions), let N(r) be the number of customers
(not including the x-customer) in queue as soon as the x- q(y, yo) = An.
customer has been servedfor T units, 0 ? T ? x. Then N(T)
(The factor n, the number of customers in the system,
follows a geometric distribution with parameter p. Moreover,
on the RHS, is because dr = n-ldt, where t is the real
the exit process of the other customers, with time measured
time.) We claim that p(y), the density of state y is,
by the amount of service the x-customer has already received,
is Markovian with intensity (K&)WN(T). n
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
p22.
Var(E(T Ix)) = P), (K2 - K1). (13) E(C I X) =2(1 K2 )X 2 ' (17)
and
In particular, both E(T I x) and Var(E(T I x)) have the same
values in LCFS-PR and PS M / M / 1 queueing systems.
x4 2
PROOF. We investigate the three terms in (2). Let M* Var(E(C x)) = (K2 - K1)
4(1 - p)2
be the number of customers who receive service before
the tagged customer clears the system. Any customer + - (K4 - K2)
4
that leaves the system as soon as the x-customer receives
r units of service, adds r to the first term on the RHS of A3K2
(2). It follows from Lemma 2.1 that + 2 ) (K3 - K2Kj). (18)
2(1 - p)
M* rx
E Ti =E TjILN(T)dT (14) PROOF. It is well-known that under these disciplines
i=l1 the expected number of customers in the queue when
the x-customer commences service is A2K2 / 2(1 - p). An
- X x~2. (15) expected number of AT are added until he is served for
2(1 p) x
r units of time, 0 - r T x. Hence,
On the other hand, 2MM*+1 Ti is clearly x times the
expected number of customers who leave from a ran- V(T) = X2K2/2(1-p) + XT, O 5 T 5 X.
dom time to the end of the busy period, that is,
The theorem easily follows now by Lemma 3.1. [l
and
LEMMA 3.1. For any T, T C (0, x), let v(r) be the ex-
pected number of customers in the queue, excluding the x-
customer, at the time in which the x-customer is servedfor Var(E(C Ix))=p22.
2 2 (K2 Ki) (20)
(1 - p)
his r unit of service. Then,
3.3. The Case of M / G / 1 PS Queue
E(C I x) = f v(T)dT. The following theorem is an immediate conclusion from
Lemma 2.1 and Lemma 3.1.
3.1. The Case of M / G / 1 Queues Without THEOREM 3.3. For a processor sharing MI G / 1 queue,
Preemption
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms
HAVIV AND RITOV
Externalities, Tangible Externalities, and Queue Disciplines
expected waiting time in the two systems conditioned 2Helpful discussions with I. Adan, G. Mills, and J. van der Wal im-
proved the content of this paper. An associate editor and a referee
on the initial state, where in one system there is an
helped in the much needed improvement of the presentation.
extra tag customer.
We compare the total waiting time of all customers References
Carlin, A. and R. R. Park, "Marginal Cost Pricing of Airport Runway
in the systems until the first time that both of them are
Capacity," American Econ. Rev., 60 (1970).
empty. Until this time, there is exactly one more cus-
Dewan, S. and H. Mendelson, "User Delay Costs and Internal Pricing
tomer in the original queue than in the simulated one. for Service Facility," Management Sci., 36 (1990), 1502-1517.
Thus, the total waiting time of all the customers in the Dolan, R. J., "Incentive Mechanism for Priority Queues," Bell J. Eco-
original system, is larger than the total waiting time in nomics, 9 (1978), 421-436.
Hassin, R., "On the Optimality of the First Come Last Served Queues,"
the other system, exactly by the length of time until the
Econometrica, 53 (1983), 201-202.
coupling. The expected value of the length of time until
and M. Haviv, "Equilibrium Strategies and the Value of Infor-
the coupling is 1/ (,u(l - p)2). This is the product of the mation in a Two Line Queueing System with Threshold Jockey-
expected number of customer in the original system just ing," Comm. in Statistics -Stochastic Models, 10 (1994), 415-436.
after the arrival of the tagged customer, 1/ (1 - p), and Kleinrock, S., Queueing Systems, Vol 2: Computer Applications, Wiley-
Interscience, New York, 1976.
the the expected time it takes to reduce the queue size
Loch, C., "Pricing in Markets Sensitive to Delay," Ph.D. Thesis, Stan-
by one, 1/ (,u(l - p)).
ford University, Stanford, CA, 1991.
However, the externalities imposed by the tagged Mendelson, H. and S. Whang, "Optimal Incentive Compatible Priority
customer equal the extra time except the time he is in Pricing for the M/M/ 1 Queue," Oper. Res., 38 (1990), 870-883.
the system. Therefore we have to subtract the expected Newbery, D. M., "Pricing and Congestion: Economic Principles Rele-
vant to Pricing Roads," Oxford Rev. of Econ. Policy, 6 (1990), 22-
value of the time the tagged customer is in the system,
38.
1/ (y.(l - p)), from the expected value of the total extra Ross, S. M., Stochastic Processes, John Wiley, New York, 1983.
waiting time. Thus, the expected externalities are 1/ Walrand, J. and P. Varaiya, High-Performance Communication Networks,
( -(l - p)2) - 1/(A(l - p)). D2 Morgan Kaufmann Publishers, San Francisco, CA, 1996.
Accepted by Linda V. Green; received in March 1996. This paper has been with the authors 12 months for 2 revisions.
This content downloaded from 201.234.181.53 on Mon, 19 Feb 2018 22:23:54 UTC
All use subject to http://about.jstor.org/terms