Upload
andreas-brandt
View
214
Download
0
Embed Size (px)
Citation preview
Queueing Syst (2006) 52: 281–286
DOI 10.1007/s11134-006-6264-1
A sample path relation for the sojourn times in G/G/1 − P Ssystems and its applicationsAndreas Brandt · Manfred Brandt
Received: 27 June 2003 / Revised: 18 October 2005C© Springer Science + Business Media, LLC 2006
Abstract For the single server system under processor shar-
ing (PS) a sample path result for the sojourn times in a busy
period is proved, which yields a sample path relation between
the sojourn times under PS and FCFS discipline. This relation
provides a corresponding one between the mean stationary
sojourn times in G/G/1 under PS and FCFS. In particular,
the mean stationary sojourn time in G/D/1 under PS is given
in terms of the mean stationary sojourn time under FCFS,
generalizing known results for G I/M/1 and M/G I/1. Ex-
tensions of these results suggest an approximation of the
mean stationary sojourn time in G/G I/1 under PS in terms
of the mean stationary sojourn time under FCFS.
Keywords G/G/1 . G/D/1 . G/G I/1 . Processor
sharing . Sojourn time . Waiting time . Busy period .
Sample path . First come first served . Comparative
analysis . Mean sojourn time approximation
Mathematics Subject Classification (MSC 2000)60K25· 68M20· 60G17· 60G10
1. Introduction
In this paper we consider a single server system where the re-
quests are served under processor sharing (PS) discipline, i.e.,
This work was supported by a grant from the Siemens AG.
A. Brandt (�)Wirtschaftswissenschaftliche Fakultat, Humboldt-Universitat zuBerlin, Spandauer Str. 1, D-10178 Berlin, Germanye-mail: [email protected]
M. BrandtKonrad-Zuse-Zentrum fur Informationstechnik Berlin (ZIB),Takustr. 7, D-14195 Berlin, Germany
if n (>0) requests are in the single server then each request
receives 1/n of the service capacity.1 Besides the PS disci-
pline, we consider the corresponding single server system
with infinite waiting room under the first come first served
(FCFS) discipline.
Since the PS and FCFS discipline are work conserving dis-
ciplines, their corresponding work load processes and hence
busy periods are identical. However, this is clearly not the
case for the sojourn times of requests, which are very sensi-
tive with respect to the service discipline.
Processor sharing systems with stochastic arrival and ser-
vice processes have been studied extensively in many papers,
see e.g. [18] and the references therein. Some recent papers
are [4, 5, 12, 16, 19]. Since the purpose of these studies mainly
was to determine sojourn time characteristics of a request,
e.g. its mean and variance (given its service requirement),
independence as well as distributional assumption are sup-
posed. However, in the general case, few structural properties
seem to be known. For an analysis of the general processor
sharing system there seems to be no other tool than sample
path analysis. For corresponding results see [2, 6].
The aim of this paper is to prove in Section 2 a sample path
result (Lemma 2.1) for the sojourn times of a busy period in
a single server system under PS, and as an application there
is given a relation (Theorem 2.1) between the sojourn times
under PS and FCFS discipline. In Section 3 we consider the
stable G/G/1 system under PS and FCFS. From Theorem 2.1
we obtain a relation between the mean stationary sojourn
times in G/G/1 under PS and FCFS (Theorem 3.1), and in
particular a formula for the mean stationary sojourn time in
G/D/1 under PS in terms of the mean stationary sojourn time
under FCFS, generalizing corresponding known results for
1 This type of PS discipline is often called the egalitarian processorsharing discipline, cf. [18] p. 102.
Springer
282 Queueing Syst (2006) 52: 281–286
G I/M/1 and M/G I/1. Then, extensions of the formula are
given for a large subset of G/G I/1 systems (Theorem 3.2),
suggesting an approximation of the mean stationary sojourn
time in G/G I/1 under PS in terms of the mean stationary
sojourn time under FCFS (Remark 3.2).
2. Sample path relations for the sojourn times in abusy period under PS and FCFS
For the single server system under PS discipline let us con-
sider a sample path of a finite busy period2 during which
n requests arrive and are served. Denote by vk and sk ,
k = 1, . . . , n, the sojourn time and service requirement of
the k-th arriving request ordered according to their arrivals,
respectively. Note that zero service requirements and equal
arrival instants are allowed. Further, let v∗k the sojourn time
of the k-th request if the arrival process is stopped after the
arrival of the k-th request.
Lemma 2.1. It holds
n∑k=1
(vk − sk) = 2n∑
k=1
(v∗k − sk). (2.1)
Proof: The proof will be given by induction on n. Since for
n = 1 it holds v1 = s1 = v∗1 , obviously (2.1) is valid. Assume
now that (2.1) is true for busy periods for which at most
n requests arrive. Consider an arbitrary busy period during
which n + 1 requests are served and which arrive at the time
instants
0 = t1 ≤ t2 ≤ · · · ≤ tn ≤ tn+1
and have service times sk , k = 1, . . . , n + 1. The departure
instant τk of the k-th request is given by
τk = tk + vk, k = 1, . . . , n+1, (2.2)
cf. Fig. 1. Note that the busy period has length∑n+1
k=1 sk in
view of the work conservation law. The induction step is
divided into four steps.
1. Consider the modification of the busy period where the
(n + 1)–st arriving request is not considered. Hence the mod-
ified busy period consists of n requests. Denote the corre-
sponding sojourn times of the modified busy period by vk ,
k = 1, . . . , n. Then the modified departure times τk are given
by
τk = tk + vk, k = 1, . . . , n. (2.3)
2 See e.g. [9] p. 11.
Fig. 1 Sample path of a busy period with four requests
Applying the induction assumption to the modified busy pe-
riod and taking into account that v∗k , k = 1, . . . , n, does not
depend on the requests arriving after the k-th request, we
obtain
n+1∑k=1
(vk −sk)
=n∑
k=1
(vk − sk) +n∑
k=1
(vk −vk) + (vn+1 − sn+1)
= 2n∑
k=1
(v∗k − sk) +
n∑k=1
(vk − vk) + (vn+1 − sn+1).
(2.4)
2. Let π be a permutation of {1, . . . , n} such that
τπ (1) ≥ τπ (2) ≥ · · · ≥ τπ (n).
In the following we will exploit several times the fact that
under the PS discipline the service capacity is divided equally
among the requests being in the system, which implies that
the residual service times of the requests in the system are
reduced with equal speed. Therefore it holds
τπ (1) ≥ τπ (2) ≥ · · · ≥ τπ (n), (2.5)
too. In view of (2.5), there are indices i, j ∈ {0, . . . , n} such
that
τπ (1) ≥ · · · ≥ τπ ( j) ≥ tn+1 > τπ ( j+1) ≥ · · · ≥ τπ (n),
τπ (1) ≥ · · · ≥ τπ (i) ≥ τn+1 > τπ (i+1) ≥ · · · ≥ τπ (n),
Springer
Queueing Syst (2006) 52: 281–286 283
respectively. Obviously, 1 ≤ j ≤ n and 0 ≤ i ≤ j . Further,
it holds
τπ (k) = τπ (k), k > j, (2.6)
τπ (k) − τπ (k+1) = τπ (k) − τπ (k+1), k < i, (2.7)
in view of the PS discipline.
For the second summand on the r.h.s. of (2.4) from (2.2),
(2.3), (2.6) and (2.7) we obtain
n∑k=1
(vk − vk)
=n∑
k=1
(τk − τk) =j∑
k=1
(τπ (k) − τπ (k)
)
=j−1∑k=1
(k + 1)((
τπ (k) − τπ (k)
) − (τπ (k+1) − τπ (k+1)
))+ ( j + 1)
(τπ ( j) − τπ ( j)
) − (τπ (1) − τπ (1)
)=
j−1∑k=max(i,1)
(k + 1)((
τπ (k) − τπ (k+1)
) − (τπ (k) − τπ (k+1)
))+ ( j + 1)(τπ ( j) − τπ ( j)) − τπ (1) + τπ (1). (2.8)
3. Next, by expressing the summands on the r.h.s. of (2.8) in
terms of departure times of the original busy period we prove
that
n∑k=1
(vk −vk) = τn+1 − tn+1 −n+1∑k=1
sk + τπ (1). (2.9)
3.1 Let 0 < i = j ≤ n. Then (τπ ( j) − tn+1) − (τπ ( j) −tn+1) is just the reduction of vπ ( j) since tn+1 in the modi-
fied busy period compared to the original one as the portion
of the service capacity assigned to each of the requests in the
system is raised from 1/( j + 1) to 1/j until τn+1. Hence
τπ ( j) − τπ ( j) = − 1
j +1(τn+1 − tn+1).
Thus (2.8) provides
n∑k=1
(vk −vk) = τn+1 − tn+1 − τπ (1) + τπ (1). (2.10)
Since i > 0 implies τπ (1) ≥ τn+1, it follows that the original
busy period, which has length∑n+1
k=1 sk , ends at τπ (1), and
thus (2.10) yields (2.9).
3.2 Let 0 ≤ i < j ≤ n. Then (τπ ( j) − tn+1) − (τπ ( j) −tn+1) is just the reduction of vπ ( j) since tn+1 in the modi-
fied busy period compared to the original one as the portion
of the service capacity assigned to each of the requests is
raised from 1/( j + 1) to 1/j . Hence
τπ ( j) − τπ ( j) = − 1
j +1
(τπ ( j) − tn+1
). (2.11)
If i < k < j , then (τπ (k) − τπ (k+1)) − (τπ (k) − τπ (k+1)) is the
reduction of vπ (k) since τπ (k+1) as the portion of the ser-
vice capacity assigned to each of the requests is raised from
1/(k + 1) to 1/k. Hence(τπ (k) − τπ (k+1)
) − (τπ (k) − τπ (k+1)
)= − 1
k+1
(τπ (k) − τπ (k+1)
). (2.12)
Finally, because (τπ (i) − τπ (i+1)) − (τπ (i) − τπ (i+1)) is the re-
duction of vπ (i) since τπ (i+1) as the portion of the service
capacity assigned to each of the requests is raised from
1/(i + 1) to 1/ i until τn+1, it holds(τπ (i) − τπ (i+1)
) − (τπ (i) − τπ (i+1)
)= − 1
i +1
(τn+1 − τπ (i+1)
). (2.13)
Thus from (2.8) and (2.11)–(2.13) it follows
n∑k=1
(vk −vk) = I{i > 0}(τn+1 − τπ (1)
) − tn+1 + τπ (1).
(2.14)
If i = 0 then τn+1 > τπ (1), and therefore the original busy pe-
riod ends at τn+1 = ∑n+1k=1 sk , and hence (2.14) implies (2.9).
If i > 0 then τn+1 ≤ τπ (1), and the original busy period ends
at τπ (1) = ∑n+1k=1 sk , and hence (2.14) implies (2.9), too.
4. Since the modified busy period ends at τπ (1), from the work
conservation law it follows τπ (1) = ∑nk=1 sk , and hence from
(2.4), (2.9) and vn+1 = τn+1 − tn+1 we obtain
n+1∑k=1
(vk − sk) = 2n∑
k=1
(v∗k − sk) + τn+1 − tn+1 − sn+1
+ (vn+1 − sn+1)
= 2n∑
k=1
(v∗k − sk) + 2(vn+1 − sn+1)
= 2n+1∑k=1
(v∗k − sk),
where the last equality is valid as obviously v∗n+1 = vn+1.
�
Springer
284 Queueing Syst (2006) 52: 281–286
From Lemma 2.1 we derive a relation between the sojourn
times of a finite busy period in a single server system under
PS and FCFS. As in Lemma 2.1 consider a sample path of
a finite busy period of n requests with service times sk , k =1, . . . , n. Additionally to the sojourn time vP
k (= vk) of the
k-th request under the PS discipline let vFk , k = 1, . . . , n, be
the corresponding sojourn time under the FCFS discipline.
Theorem 2.1. For the sojourn times vPk and vF
k of a finitebusy period of n requests in a single server system under PSand FCFS discipline, respectively, it holds
n∑k=1
(vP
k − sk) ≤ 2
n∑k=1
(vF
k − sk). (2.15)
In case of constant service times we have equality in (2.15).
Proof: Consider a sample path of a busy period of n re-
quests, and let 0 = t1 ≤ · · · ≤ tn be the arrival epochs of the
n requests. Analogously to Lemma 2.1 let v∗k be the sojourn
time of the k-th request in the PS system if the arrival process
is stopped after the arrival of the k-th request. Obviously, the
work conservation law provides
tk + v∗k ≤
k∑j=1
s j , k = 1, . . . , n, (2.16)
as the busy period induced by the first k arrivals
(t1, s1), . . . , (tk, sk) has length∑k
j=1 s j . If all service times
are equal, we have equality in (2.16) because then the re-
quests leave the system in the order of their arrivals. In case
of FCFS discipline it holds
tk + vFk =
k∑j=1
s j , k = 1, . . . , n. (2.17)
Combining (2.1) where vk = vPk , (2.16) and (2.17) we obtain
(2.15), where in (2.15) we have equality in case of constant
service times. �
Remark 2.1. The difference of the sojourn and service time
of a request can be interpreted as the waiting time of the
request also in case of a PS system, cf. [18] p. 107, namely
as the time which the request has to spend additionally to its
service time in the system in view of the presence of other
requests. Theorem 2.1 provides the sharp bound 2 in the worst
case comparative analysis for the sum of the waiting times
under PS and FCFS. Note that there is no finite bound for
the other direction because of the example (t1, s1) := (0, 2),
(t2, s2) := (1, 0).
3. Relations between the mean sojourn times forG/G/1 under PS and FCFS
In this section we derive some relations between the mean
stationary sojourn times for the single server system under
PS and FCFS with a stochastic arrival and service process.
By G/G/1 we denote the single server system where the
stochastic arrival and service process is a stationary and er-
godic sequence � = {[A�, S�]}∞�=−∞, where A� (≥0) denotes
the inter-arrival time between the �-th and (� + 1)–st request
and S� (≥0) the service requirement of the �-th request.3 If
additionally {S�}∞�=−∞ is a sequence of i.i.d. r.v.’s and inde-
pendent of the sequence {A�}∞�=−∞ then the system is de-
noted by G/G I/1. The arrival instants of the requests are
given by T� = − ∑−1k=� Ak , � < 0, T� = ∑�−1
k=0 Ak , � ≥ 0, cf.
e.g. [1, 7]. The marked point process � = {[T�, S�]}∞�=−∞(or equivalently the sequence �) is the arrival stationary
model of the input. By means of the inversion formula
for marked point processes, a corresponding stationary er-
godic marked point process � = {[T�, S�]}∞�=−∞ on R, with
· · · ≤ T−1 ≤ T0 ≤ 0 < T1 ≤ · · ·, can be given, describing the
time stationary version of the input, cf. e.g. [1, 7, 8]. (Note
that � has the Palm distribution of �.)
For the general G/G/1 system under any work conserving
discipline, in particular for the PS and FCFS discipline, the
stability condition reads
� := E S
E A< 1, (3.1)
where A and S denote generic r.v.’s of A� and S�, respectively,
cf. e.g. [1, 7]. Under (3.1) a uniquely determined stationary
workload process can be constructed (in the arrival as well as
time stationary model), having empty points and thus busy
periods of finite length. Hence for the general G/G/1 sys-
tem under PS a uniquely determined stationary process of
the number of requests and their residual service times can
be constructed along the lines of [7], but which will not be
outlined here.
By means of ergodicity arguments, cf. e.g. [1, 7, 8], the
sample path result of Theorem 2.1 yields a corresponding
one for the mean stationary sojourn times.
Theorem 3.1. Let � < 1. Then for the stationary sojourntimes V P of the G/G/1 − P S system and V F of the corre-sponding G/G/1 − FC F S system it holds
EV P −E S ≤ 2 (EV F −E S) (3.2)
3 Since A� and S� are assumed to be non-negative only, batch arrivalsas well as zero time requests are included.
Springer
Queueing Syst (2006) 52: 281–286 285
with equality in case of deterministic service times, i.e., forthe G/D/1 system it holds
EV P −E S = 2 (EV F −E S). (3.3)
We will generalize (3.3) to a broader class of G/G I/1 sys-
tems, covering in particular corresponding known results for
G I/M/1 and M/G I/1. Note that for a G/G I/1 system in
the time stationary version � = {[T�, S�]}∞�=−∞ of the input,
{S�}∞�=−∞ is independent of {T�}∞�=−∞ and an i.i.d. sequence
of r.v.’s with the same distribution as {S�}∞�=−∞, cf. e.g. [7].
Consider the service time distributions Bi (x) := P(S ≤ x),
i ∈ {1, 2}:
B1(x) := 1 − p I{x <s}, x ∈ R+, (3.4)
B2(x) := 1 − p exp(−x/s), x ∈ R+, (3.5)
where p ∈ (0, 1] and s ∈ R+ \ {0}, i.e., Bi (x) is a mixture of
a zero and a deterministic or exponential time with mean s,
respectively.
Theorem 3.2. Let � < 1. Then for the systems G/G I/1 withG I = Bi , i ∈ {1, 2}, and M/G I/1 it holds
EV P − E S = 2(E S)2
E S2(EV F − E S), (3.6)
where V P and V F denote the stationary sojourn times in thecorresponding systems under the PS and FCFS discipline,respectively.
Proof: 1. For the M/G I/1 system (3.6) follows from the
well-known formulae for EV P and EV F , cf. e.g. [18] p. 109
and [17] p. 278.
2. Consider the systems G/G I/1 with G I = Bi , i ∈ {1, 2}.2.1 First let p = 1, i.e., we deal with the G/D/1 and
G/M/1 system, respectively. For a G/D/1 system (3.6) fol-
lows from (3.3) as E S2 = (E S)2. In case of a G/M/1 sys-
tem (3.6) is equivalent to EV P = EV F because of E S2 =2(E S)2. In view of the exponential service times, under the
PS as well as under the FCFS discipline the departure pro-
cess is a Poisson process of intensity 1/E S as long as there
are requests in the system. Consequently, the numbers of re-
quests in the system are stochastically equivalent under both
disciplines, and by Little’s formula it follows EV P = EV F .
2.2 Now let p ∈ (0, 1), i.e., the service times are a proper
mixture of a zero and a positive (deterministic or exponential)
time. Consider the time stationary input � = {[T�, S�]}∞�=−∞and the modified thinned input � = {[T�, S�]}∞�=−∞ with
· · · ≤ T−1 ≤ T0 ≤ 0 < T1 ≤ · · · consisting of all those re-
quests of � having a positive service time.4 In the following
let us endow all quantities and variables which are related
to � with a tilde, thus S denotes the generic deterministic
or exponential service time with mean s, and V P , V F are
the stationary sojourn times in the single server system with
input � under the PS and FCFS discipline, respectively. The
following observation is crucial: the requests with a zero
service time have under the PS discipline a sojourn time of
length zero whereas under the FCFS discipline they do not
have any impact on the workload process and hence on the
waiting times of the requests with positive service times.
This, (3.4), (3.5), and since the service times {S�}∞�=−∞ are
i.i.d. r.v.’s independent of the arrival process, yield in case of
G I = Bi , i ∈ {1, 2}, the relations
EV P = pEV P , EV F − E S = EV F − E S, (3.7)
E S = pE S, E S2 = pE S2. (3.8)
Note that EV P (EV F − E S) is the mean stationary so-
journ time (waiting time) in the corresponding G/D/1 − P S(FC F S) or G/M/1 − P S (FC F S) system, respectively,
where G stands for the process {T�}∞�=−∞ of arrival instants
of �, D for the constant service times s in case of B1(x) and
M for the exponential service times with mean s in case of
B2(x). (A rigorous proof of (3.7) can be given by applying
Palm’s formula for marked point processes and using the in-
dependence assumptions of the service times.) From (3.7),
(3.8) and step 2.1 above we obtain
EV P − E S = p(EV P − E S) = p2(E S)2
E S2(EV F − E S)
= 2(E S)2
E S2(EV F − E S).
�
Remark 3.1. For the G I/M/1 system the fact EV P = EV F
is well-known, cf. e.g. [13] p. 441. Theorem 3.2 states that
(3.6) holds for a much broader class than for G I/M/1 and
M/G I/1 systems.
Although Theorem 3.2 shows that (3.6) holds for a broad
class of G/G I/1 systems, this relation is not true for
G/G I/1 systems in general. Consider the MG I /G I/1 sys-
tem, i.e., batches of positive size X arrive according to a
Poisson process of intensity λ. Note that the MG I batch ar-
rival process is a special G process, cf. e.g. [1, 7]. The mean
sojourn time EV F of an arriving request under FCFS is given
4 The arrival stationary model of � can be given via Palm’s inversionformula for marked point processes, cf. e.g. [1, 7].
Springer
286 Queueing Syst (2006) 52: 281–286
by, cf. e.g. [15] p. 277,
EV F = E S + E S
2(1 − �)
(�
E S2
(E S)2+ b
), (3.9)
where � = λ E X E S < 1, cf. (3.1), and b := E X2/E X −1 is the mean number of requests that arrive together with
the tagged arriving request other than this request. Since for
the MG I /G I/1 − P S system an explicit formula for EV P
seems not to be available, let us restrict to H2-distributed
service times with balanced means, i.e.,
H2(x) := p1(1 − exp(−x/s1))
+ p2(1 − exp(−x/s2)), x ∈ R+, (3.10)
where p1 = 1 − p2 ∈ (0, 1), s1, s2 ∈ R+ \ {0}, s1 �= s2 and
p1s1 = p2s2. (3.11)
Specializing the general results of [3, 10] for the
MG I /G I/1 − P S and MG I /G H/1 − P S system, where
G H stands for generalized hyperexponential distribution, re-
spectively, to the H2 distribution with balanced means, after
tedious algebra one obtains that for the MG I /H2/1 − P Ssystem with (3.11) it holds
EV P = E S + E S
2(1 − �)
(2� + b − ab
2 − �
), (3.12)
where a := (E S2 − 2(E S)2)/E S2 ∈ (0, 1). From (3.9),
(3.12) it follows
EV P − E S
= 2(E S)2
E S2
(1 + (1 − �)ab
(2 − �)(2� + (1 − a)b)
)(EV F − E S).
(3.13)
Therefore, for the MG I /H2/1 system with (3.11) it holds
EV P − E S ≥ 2(E S)2
E S2(EV F − E S) (3.14)
with equality iff b = 0, in which case we have a M/H2/1
system covered by Theorem 3.2.
Remark 3.2. As (3.6) holds for a broad class of G/G I/1
systems and implies (3.2), in case of the general G/G I/1
system (3.6) can be used as an approximation of EV P in
terms of EV F . The mean stationary sojourn time EV F un-
der FCFS can be approximated e.g. by a moment fitting of
the service time distribution B(x), preferentially by a two-
moment fitting to B1(x) as (3.6) is fulfilled for B1(x) and
B1(x) covers the full range of the first two moments. For the
G I/G I/1 system EV F can also be approximated by the for-
mula given in [11]. For a different approximation of EV P in
the G I/G I/1 system see [14].
References
1. F. Baccelli and P. Bremaud, Elements of Queueing Theory. Appli-cations of Mathematics 26, (Springer-Verlag, Berlin, 2003).
2. F. Baccelli and D. Towsley, The customer response times in theprocessor sharing queue are associated, Queueing Systems 7 (1990)269–282.
3. N. Bansal, Analysis of the M/G/1 processor-sharing queue withbulk arrivals, Oper. Res. Lett. 31 (2003) 401–405.
4. S. Borst, O. Boxma, and P. Jelenkovic, Reduced-load equivalenceand induced burstiness in GPS queues with long-tailed traffic flows,Queueing Systems 43 (2003) 273–306.
5. A. Brandt and M. Brandt, On the sojourn times for many-queuehead-of-the-line processor-sharing systems with permanent cus-tomers, Math. Methods Oper. Res. 47 (1998) 181–220.
6. A. Brandt and M. Brandt, A note on the stability of the many-queue head-of-the-line processor-sharing system with permanentcustomers, Queueing Systems 32 (1999) 363–381.
7. A. Brandt, P. Franken, and B. Lisek, Stationary Stochastic Models(Akademie-Verlag, Berlin; Wiley, Chichester, 1990).
8. P. Franken, D. Konig, U. Arndt, and V. Schmidt, Queues and PointProcesses (Akademie-Verlag, Berlin; Wiley, Chichester, 1982).
9. D. Gross and C.M. Harris, Fundamentals of Queueing Theory (Wi-ley, New York, 1998).
10. L. Kleinrock, R.R. Muntz and E. Rodemich, The processor-sharingqueueing model for time-shared systems with bulk arrivals, Net-works 1 (1971) 1–13.
11. W. Kramer and M. Langenbach-Belz, Approximate formulae for thedelay in the queueing system G I/G/1, in: Proc. ITC 8, Melbourne,Australia, (1976) pp. 235.
12. R. Nunez-Queija, Sojourn times in a processor sharing queue withservice interruptions, Queueing Systems 34 (2000) 351–386.
13. V. Ramaswami, The sojourn time in the G I/M/1 queue with pro-cessor sharing, J. Appl. Probab. 21 (1984) 437–442.
14. B. Sengupta, An approximation for the sojourn-time distribution forthe G I/G/1 processor-sharing queue, Commun. Statist. StochasticModels 8 (1992) 35–57.
15. H.C. Tijms, Stochastic Models: An Algorithmic Approach (Wiley,Chichester, 1994).
16. M. van Uitert and S.C. Borst, A reduced-load equivalence for gen-eralised processor sharing networks with long-tailed input flows,Queueing Systems 41 (2002) 123–163.
17. R.W. Wolff, Stochastic Modeling and the Theory of Queues(Prentice-Hall International, Englewood Cliffs, N.J., 1989).
18. S.F. Yashkov, Mathematical problems in the theory of shared-processor systems, J. Soviet Math. 58 (1992) 101–147. (OriginalRussian version in: Itogi Nauki i Tekhniki, Seriya Teoriya Veroyat-nostei, Matematicheskaya Statistika, Teoreticheskaya Kibernetika29 (1990) 3–82.)
19. A.P. Zwart and O.J. Boxma, Sojourn time asymptotics in theM/G/1 processor sharing queue, Queueing Systems 35 (2000) 141–166.
Springer