/ department of mathematics and computer science 1/40
LNMB Course
Asymptotic Methodsin Queueing Theory
Lecture 3, February 25, 2013
Rudesindo Núñez-Queija (CWI/UvA), Sem Borst (TU/e)
http://www.win.tue.nl/˜sem/AsQT/
/ department of mathematics and computer science 2/40
Course overview
Four main topics
• Large deviations and tail asymptotics
– Introduction, large deviations, large-buffer asymptotics for light-tailed queues
– Many-sources asymptotics, large-buffer asymptotics for heavy-tailedqueues
– Large-buffer asymptotics for heavy-tailed queues, impact of servicediscipline, Processor Sharing and its variants
• Fluid and diffusion limits
• Perturbation analysis and time scale separation
• Heavy-traffic approximations
/ department of mathematics and computer science 3/40
Heavy-tailed vs. light-tailed distributions
Variability and decreasing failure rate are common characteristics of so-called heavy-tailed distributions, which exhibit (slow) polynomial decay:
P{B > x}eθx→∞ as x →∞
for all θ > 0, e.g. Pareto distribution:
P{B > x} ∼ γ x−ν as x →∞
In contrast, tails of light-tailed distributions exhibit (fast) exponential decay:
P{B > x} ∼ αe−βx as x →∞
While light-tailed distributions typically used to be assumed in (queueing)applications, in recent years empirical findings have pointed to widespreadoccurrence of heavy-tailed distributions
/ department of mathematics and computer science 4/40
Traffic characteristics
Internet traffic exhibits burstiness on wide range of time scalesManifests itself in long-range dependence & self-similarity
/ department of mathematics and computer science 5/40
Traffic characteristics (cont’d)
Contrasts with traditional traffic assumptions(Poisson arrivals, finite-variance service requirements)
/ department of mathematics and computer science 6/40
Heavy-tailed distributions
Burstiness is caused by extreme variability in traffic processes(activity periods, file sizes)
Measurements suggest that file sizes follow Pareto distribution
P{B > x} ∼ γ x−ν as x →∞,
with 1 < ν < 2 (ν ≈ 1.7)
• E{B} <∞: finite mean
• E{B2} = ∞: infinite variance!!
Queueing models provide fundamental insight into performance impact ofheavy-tailed traffic characteristics and potential role of service discipline inlimiting detrimental effect
/ department of mathematics and computer science 7/40
Heavy-tailed distributions (cont’d)
• Function F(·) is long-tailed if
limx→∞
1− F(x − y)1− F(x)
= 1, for all y ∈ R
• Function F(·) is subexponential if
limx→∞
1− F2∗(x)1− F(x)
= 2,
where F2∗(·) is two-fold convolution of F(·) with itself, implying
P{X1 + · · · + Xn > x} ∼ nP{X1 > x} as x →∞
/ department of mathematics and computer science 8/40
Heavy-tailed distributions (cont’d)
• Function F(·) is regularly varying of index −ν if
F(x) = 1−L(x)xν
, ν ≥ 0,
where L : R+→ R+ is a function of slow variation,i.e., limx→∞ L(ηx)/L(x) = 1, η > 1
• F(·) is intermediately regularly varying if
limη↓1
lim infx→∞
1− F(ηx)1− F(x)
= 1
/ department of mathematics and computer science 9/40
Heavy-tailed distributions (cont’d)
Focus on ‘exact’ large-buffer asymptotics of aggregate workload V in steadystate
Other issues
• ‘logarithmic’ many-sources asymptotics
• other performance measures (individual workloads, delays, queuelengths)
• impact of service discipline (priority mechanisms, PS, LPS, DPS, GPS,SRPT, LAS)
/ department of mathematics and computer science 10/40
Instantaneous input
Source generates instantaneous traffic burstsaccording to renewal process (G/G/1 queue)
time
Workload
Interarrival times generally distributed with mean 1/λBurst size distribution B(x) = P{B < x} with mean β <∞Traffic intensity ρ := λβResidual burst size distribution Br(x) = 1
β
∫ xy=0(1− B(y))dy
/ department of mathematics and computer science 11/40
Instantaneous input (cont’d)
Theorem [Cohen, Pakes]If Br(·) is subexponential, and ρ < 1, then
P{V > x} ∼ρ
1− ρP{Br > x} as x →∞
time
Workload
Disaster scenario:Due to SINGLE extremely large burst
/ department of mathematics and computer science 12/40
Heuristic derivation
Suppose that burst of size x + y(1− ρ) or largerarrives at time −y < 0Then workload is x or larger at time 0
P{V > x} ≈ λ
∫∞
y=0P{B > x + y(1− ρ)}dy
=λ
1− ρ
∫∞
y=0P{B > x + y}dy
=ρ
1− ρP{Br > x}
/ department of mathematics and computer science 13/40
Three proofs for M/G/1 queue
• Proof via sample-path upper and lower bounds (based on heuristicderivation)
• Direct proof
• Proof via Laplace-Stieltjes Transform and Bingham-Doney
/ department of mathematics and computer science 14/40
1. Proof via sample-path upper and lowerbounds
Lower bound: ‘easy’
P{V > x} ≥ρ
1− ρ + δ
∫∞
x(1+ε)
P{B > z}E{B} dz
Use Law of Large Numbers to show that single big jump yields this lowerbound
Upper bound: ‘hard’
P{V > x} ≤ρ
1− ρ − δ
∫∞
x(1−ε)
P{B > z}E{B} dz + o(x1−ν)
Include all other scenarios (like two big jumps) and show that they canasymptotically be neglected.
/ department of mathematics and computer science 15/40
2. Direct proof
Pollaczek-Khinchine formula for LST of amount of work
E{e−sV} =
1− ρ1− ρE{e−sBr
}=
∞∑n=0
(1− ρ)ρn(E{e−sBr
}
)n,
implies that V may be represented as
V =st Br1 + · · · + Br
N,
with N ∼ Geo(ρ), and hence
P{V > x} = (1− ρ)∞∑
n=0
ρnP{Br1 + · · · + Br
n > x}
∼ (1− ρ)∞∑
n=0
ρnnP{Br > x}
=ρ
1− ρP{Br > x} as x →∞
/ department of mathematics and computer science 16/40
Some observations
Observation 1
P{Br1 + · · · + Br
n > x} ∼ nP{Br > x} as x →∞
holds for subexponential distributions
Crucial property: if sum is large, it is most likely due to single big term
Observation 2If P{B > x} ∼ x−νL(x) as x →∞, then
P{Br > x} =∫∞
x
P{B > y}E{B} dy
∼1
(ν − 1)E{B}x1−νL(x) as x →∞
“one degree worse”
/ department of mathematics and computer science 17/40
3. Proof via Laplace-Stieltjes Transform and Bingham-Doney
Key lemma for regularly varying distributions [Bingham & Doney]:If k < ν < k + 1, k ∈ N, then next two properties are equivalent:
P{X > x} ∼ x−νL(x) as x →∞
and
E{e−sX} −
k∑j=0
E{X j}(−s) j
j !∼ −0(1− ν)sνL(
1s) as s ↓ 0
/ department of mathematics and computer science 18/40
3. Proof via Laplace-Stieltjes Transform and Bingham-Doney (cont’d)
If P{B > x} ∼ x−νL(x) as x →∞, 1 < ν < 2, then (Bingham-Doney):
1− E{e−sBr} = 1−
1− E{e−sB}
sE{B}
∼ −0(1− ν)
E{B} sν−1L(1/s) as s ↓ 0
Combine with Pollaczek-Khinchine formula
E{e−sV} =
1− ρ1− ρE{e−sBr
}=
1− ρ1− ρ + ρ(1− E{e−sBr
}),
to obtain, for s ↓ 0:
1− E{e−sV} ∼ −
ρ
1− ρ0(1− ν)
E{B} sν−1L(1/s)
Then apply Bingham-Doney once again
/ department of mathematics and computer science 19/40
Instantaneous input (cont’d)
In contrast, for light-tailed burst size distribution,
time
Workload
Conspiracy scenario:Combination of MANY relatively large burstsand MANY relatively short interarrival times
/ department of mathematics and computer science 20/40
Fluid input
Source generates fluid traffic according to On-Off process
time
Workload
Off-periods generally distributed with mean 1/λOn-period distribution A(x) = P{A < x} with mean α <∞Fraction On-time p = α/(α + 1/λ)While On, flow produces traffic at constant rate rTraffic intensity ρ := prResidual On-period distribution Ar(x) = 1
α
∫ xy=0(1− A(y))dy
/ department of mathematics and computer science 21/40
Fluid input (cont’d)
Theorem [Jelenkovic & Lazar]If Ar(·) is subexponential, and ρ < 1 < r , then
P{V > x} ∼ (1− p)ρ
1− ρP{Ar > x/(r − 1)} as x →∞
time
Workload
Due to SINGLE extremely long On-period
/ department of mathematics and computer science 22/40
Heuristic derivation
Suppose that On-period of length xr−1 + y 1−ρ
r−ρ or largerstarts at time −y − x
r−1 < 0Then workload is x or larger at time 0
P{V > x} ≈1
α + 1/λ
∫∞
y=0P{A > x
r − 1+ y
1− ρr − ρ
}dy
=1
α + 1/λr − ρ1− ρ
∫∞
y=0P{A > x
r − 1+ y}dy
= (1− p)ρ
1− ρP{Ar >
xr − 1
}
/ department of mathematics and computer science 23/40
Fluid input (cont’d)
In contrast, for light-tailed On-period distribution,
time
Workload
Combination of MANY relatively long On-periodsand MANY relatively short Off-periods
/ department of mathematics and computer science 24/40
Several On-Off sources
Now suppose there are N statistically identical, independent On-Off sources,with Nρ < 1 for stability
Previous results suggest that large workload typically occurs due to singleextremely long On-periodIf r + (N − 1)ρ > 1, then long On-period of just a single source is sufficientto cause persistent positive drift
Large workload is typically caused by extremely long On-period of just asingle source, while other N − 1 sources show roughly normal behavior
/ department of mathematics and computer science 25/40
Several On-Off sources (cont’d)
TheoremIf A(·) is regularly varying, and Nρ < 1 < r + (N − 1)ρ, then
P{V > x} ∼ NP{V1−(N−1)ρ > x} as x →∞
Here V1−(N−1)ρ represents workload in system with just a single source andservice rate c = 1 − (N − 1)ρ, i.e., original rate reduced by average rate ofother N − 1 sources
P{Vc > x} ∼ (1− p)ρ
c − ρP{Ar > x/(r − c)} as x →∞
/ department of mathematics and computer science 26/40
Several On-Off sources (cont’d)
What if r + (N − 1)ρ < 1?Multiple sources must experience long simultaneous On-periods in orderfor workload to build up
Let M be such that (M − 1)r + (N − M + 1)ρ < 1 < Mr + (N − M)ρThus, M is minimum number of sources that must be On in order to causepersistent positive drift
Large workload is typically caused by extremely long simultaneous On-periods of exactly M sources, while other N − M sources show roughlynormal behavior
/ department of mathematics and computer science 27/40
Several On-Off sources (cont’d)
TheoremIf A(·) is regularly varying, and (M−1)r + (N −M+1)ρ < 1 < Mr + (N −M)ρ, then
P{V > x} ∼(
NM
)P{V1−(N−M)ρ
{1,...,M} > x} as x →∞
Here Vc{1,...,M} represents workload in critical system with only M sources
and service rate c = 1− (N −M)ρ, i.e., original rate reduced by average rateof other N − M sources
P{Vc{1,...,M} > x} ∼ G M
(P{Ar > x/(Mr − c)}
)M as x →∞
G M is some constant determined by an integral expression, which capturesgeometric probabilistic structure of overlap of M simultaneous On-periods
/ department of mathematics and computer science 28/40
Heterogeneous On-Off sources
Previous results suggest that large workload typically occurs due to simulta-neous long On-periods of "minimum combination" of sources
In case of heterogeneous source characteristics, "minimum combination"is no longer determined by just number of sources, but also by rates ri ,traffic intensities ρi , and tail exponents νi
In order for set of sources S ⊆ {1, . . . , N } to cause positive drift, we musthave ∑
j∈S
r j +∑j 6∈S
ρ j > 1
In order to reach workload x , simultaneous On-periods must last for periodof order x , which happens with probability
x−∑j∈S(ν j−1)
/ department of mathematics and computer science 29/40
Heterogeneous On-Off sources (cont’d)
In order for S to be "minimum combination", it should minimize tail expo-nent
∑j∈S(ν j − 1), subject to drift condition:
minS⊆{1,...,N }
∑j∈S
(ν j − 1)
sub∑j∈S
r j +∑j 6∈S
ρ j > 1
Above optimization problem amounts to knapsack problem
/ department of mathematics and computer science 30/40
Heterogeneous On-Off sources (cont’d)
TheoremIf S∗ is unique solution to above optimization problem, then
P{V > x} ∼ P{VcS∗S∗ > x} as x →∞
Here VcS∗S∗ represents workload in critical system with only sources in S∗ and
service rate cS∗ := 1 −∑
j 6∈S∗ρ j , i.e., original rate reduced by average rate of
other sources
P{VcS∗S∗ > x} ∼ GS∗
∏j∈S∗
P{Arj > x/(rS∗ − cS∗)} as x →∞,
with rS∗ :=∑
j∈S∗r j
GS∗ is some constant determined by an integral expression, which capturesgeometric probabilistic structure of overlap of simultaneous On-periods
/ department of mathematics and computer science 31/40
Combination with light-tailed sources
• Set I1 of general light-tailed sources
• Set I2 of heavy-tailed On-Off sources
Dichotomy in qualitative behavior of P{V > x}If ρI1 + rI2 > 1, then P{V > x} has heavy-tailed characteristicsas considered before
P{V > x} ∼ P{V1−ρI1I2
> x} as x →∞
If ρI1 + rI2 < 1, then P{V > x} has light-tailed characteristics
P{V > x} ≤ P{V1−rI2I1
> x}
Thus light-tailed sources must show abnormal behavior too in order forworkload to build up
/ department of mathematics and computer science 32/40
Combination with light-tailed sources (cont’d)
Consider workload VcI1
in isolated system with only light-tailed sources andservice rate c = 1 − rI2, i.e., original rate reduced by aggregate peak rate ofheavy-tailed sources
Large workload x is most likely caused by ‘increase in traffic intensity’ fromρI1 to some value ρI1 > c for period of time x
ρI1−c
Let’s now return to combined system with heavy-tailed On-Off sources
Suppose that all heavy-tailed sources happen to be constantly On while light-tailed sources show abnormal behavior
P{V > x} ∼ P{V1−rI2I1
> x}∏j∈I2
p jP{Arj >
xρI1 + rI2 − 1
}
/ department of mathematics and computer science 33/40
Delay and impact of service discipline
So far we have focused on workload asymptotics
Workload asymptotics hold for any service discipline
For FCFS, waiting time has same distribution as workload so
P{WFC F S > x} ∼ρ
1− ρP{Br > x} as x →∞
and
P{SFC F S > x} = P{WFC F S + B > x} ∼ P{WFC F S > x} ∼ρ
1− ρP{Br > x}
However, for non-FCFS service disciplines there is no simple relationshipin general between workload and waiting or sojourn time
/ department of mathematics and computer science 34/40
Delay and impact of service discipline (cont’d)
Some service disciplines of interest
• Processor Sharing (PS)
– total service capacity is shared fairly among all customers
– when there are n customers in system, each receives fraction 1/n oftotal service capacity
– provides idealization of Round-Robin (RR) scheduling
• Limited Processor Sharing (LPS)
– total service capacity shared among up to M customers
– when there are n customers in system, oldest min{M, n} of them eachreceive fraction 1/min{M, n} of total service capacity
– remaining n −min{M, n} = max{n − M, 0} customers are waiting
/ department of mathematics and computer science 35/40
Delay and impact of service discipline (cont’d)
Some service disciplines of interest
• Discriminatory Processor Sharing (DPS)
– there are K customer classes with weights w1, w2, . . . , wK
– when there are nk class-k customers, k = 1, . . . , K , each class-l cus-tomer receives fraction wl
w1n1+w2n2+···+wK nKof total service capacity
– provides abstraction for service differentiation mechanisms, e.g.,Weighted Round-Robin (WRR) scheduling
• Generalized Processor Sharing (GPS):
– there are K customer classes with weights w1, w2, . . . , wK
– when subset of non-empty classes is L ⊆ {1, 2, . . . , K }, each classl ∈ L receives fraction wl∑
k∈Lwknkof total service capacity (which may
then be shared among the class-l customers in various ways)
– provides abstraction for service differentiation mechanisms such asWRR, but allocation on per-class rather than per-customer basis
/ department of mathematics and computer science 36/40
Delay and impact of service discipline (cont’d)
Some service disciplines of interest
• Shortest Remaining Processing Time First (SRPT)
– assigns priority to customer with shortest remaining service time
– minimizes number of customers in system sample-path wise(and hence expected sojourn time because of Little’s law)
– requires advance knowledge of service times
• Least Attained Service First (LAS)
– assigns priority to customer with least amount of service received sofar
– does not require advance knowledge of service times
– minimizes number of customers in system in distribution when ser-vice time distribution has decreasing failure rate(and hence expected sojourn time because of Little’s law)
/ department of mathematics and computer science 37/40
Delay and impact of service discipline (cont’d)
For broad class of ‘preemptive’ service disciplines (PS, DPS, SRPT, LAS), ifB(·) is regularly varying, then
P{S > x} ∼ P{B > (1− ρ)x} as x →∞
In particular, for regularly varying distributions, sojourn time has same tailindex ν as service time, rather than one degree worse ν − 1 as for FCFS
This agrees with fact that E{S} <∞ even when E{B2} = ∞ for these service
disciplines, and reflects some notion of tail optimality
Interpretation
• During sojourn time of customer with large service time, other cus-tomers will take away fraction ρ of server capacity
• Thus large service time (1− ρ)x will result in sojourn time x
/ department of mathematics and computer science 38/40
References
S. Aalto, U. Ayesta, S.C. Borst, V. Misra, R. Núñez-Queija (2007). BeyondProcessor Sharing. Perf. Eval. Rev. 34 (4), 36–43.R. Agrawal, A.M. Makowski, Ph. Nain (1999). On a reduced load equiva-lence for fluid queues under subexponentiality. Queueing Systems 33, 5–41.V. Anantharam (1999). Scheduling and long-range dependence. QueueingSystems 33, 73–89.N.H. Bingham, R.A. Doney (1975). Asymptotic properties of supercriticalbranching processes I: The Galton-Watson process. Adv. Appl. Prob. 6, 711–731.N.H. Bingham, C. Goldie, J. Teugels (1987). Regular Variation. CambridgeUniversity Press.S.C. Borst, O.J. Boxma, R. Núñez-Queija, A.P. Zwart (2003). The impact ofthe service discipline on delay asymptotics. Perf. Eval. 54 (2), 175–206.S.C. Borst, R. Núñez-Queija, A.P. Zwart (2006). Sojourn time asymptoticsin processor-sharing queues. Queueing Systems 53, 31–51.
/ department of mathematics and computer science 39/40
References (cont’d)
O.J. Boxma, A.P. Zwart (2007). Tails in scheduling. Perf. Eval. Rev. 34 (4),13–20.J.W. Cohen (1973). Some results on regular variation for distributions inqueueing and fluctuation theory. J. Appl. Prob. 10, 343–353.P. Embrechts, C. Klüppelberg, T. Mikosch (1997). Modelling Extremal Events.Springer, Berlin.P. Embrechts, N. Veraverbeke (1982). Estimates for the probability of ruinwith special emphasis on the possibility of large claims. Insurance: Math.Econ. 1, 55–72.F. Guillemin, Ph. Robert, A.P. Zwart (2003). Tail asymptotics for processor-sharing queues. Adv. Appl. Prob. 36 (2), 525–543.P.R. Jelenkovic, A.A. Lazar (1999). Asymptotic results for multiplexingsubexponential on-off processes. Adv. Appl. Prob. 31, 394–421.J.K. Nair, A. Wierman, A.P. Zwart (2010). Tail-robust scheduling via limitedProcessor Sharing. Perf. Eval. 67, 978–995.
/ department of mathematics and computer science 40/40
References (cont’d)
R. Núñez-Queija (2000). Processor-Sharing Models for Integrated-ServicesNetworks. PhD Thesis Eindhoven University of Technology.A.G. Pakes (1975). On the tails of waiting-time distributions. J. Appl. Prob.12, 555–564.N. Veraverbeke (1977). Asymptotic behaviour of Wiener-Hopf factors of arandom walk. Stoch. Proc. Appl. 5, 27–37.A.P. Zwart (2001). Queueing Systems with Heavy Tails. PhD Thesis Eind-hoven University of Technology.A.P. Zwart, O.J. Boxma (2000). Sojourn time asymptotics in the M/G/1processor-sharing queue. Queueing Systems 35, 141–166.
Top Related