Thheeoorryy ooff...
Transcript of Thheeoorryy ooff...
TT hh ee oo rr yy oo ff SS cc hh ee dd uu ll ii nn gg
Scheduling problems have been classified, in the following 3 field notaiton. Acomprehensive survey of the theory is contained in Lawler, Lenstra, Rinnooy-Kan and Shmoys (1993):
1st Field describes the machine environment:Single machine: 1/ /Parallel machines Pm/ /Uniform machines Qm/ /Unrelated machines Rm/ /Open shop Om/ /Flowshop Fm/ /Jobshop Jm/ /2nd Field describes the Job EnvironmentUnit processing times /Xj=1/Release dates /rj,/Deadlines /Dj/Precedence constraints, of various types /prec/ /chain/ /intree/ /serpar/Preemptions /pmtn/Stochastic /~stoc/
3nd Field describes the Optimality CriteriaGeneral Minmax / /fmaxGeneral Minsum / /ΣfjMakespan / /CmaxFlowtime / /ΣCjWeighted flowtime / /ΣwjCjMax lateness / /LmaxMax tardiness / /TmaxSum of tardiness / /ΣTjWeighted tardiness / /ΣwjTjNumber of late jobs / /ΣUjExamples (1) A 3 machine flowshop, minimize sum of completiontimes (flowtime), allowing preemptions: F3/pmtn/ΣΣΣΣCj(2) Schedule jobs with precedence constraints and release dates, on parallelmachines (we are interested in solving the problem for sny number ofmachines) so as to minimize the weighted sum of late jobs:
MS&E324, Stanford University, Spring 2002 2-1 Gideon Weiss© manufacturing & control
P/rj,prec/ΣΣΣΣwjUj
CC oo mm pp uu tt aa tt ii oo nn ll CC oo mm pp ll ee xx ii tt yy
The world of NP
NP = P
P
NP-Complete
NP
The World of NP
Using the above 3 field classification 4536 have beenconsidered.Polynomial time algorithms were devised for 416 ofthem.3582 of the problems were shown to be NP-complete
MS&E324, Stanford University, Spring 2002 2-2 Gideon Weiss© manufacturing & control
(This left us in ~1993 with room for 538 PhD theses).
MM ii nn ii mm ii zz ii nn gg FF ll oo ww tt ii mm ee oo nn aa SS ii nn gg ll ee MM aa cc hh ii nn ee
1/ /ΣCj
Jobs numbered j N= 1, ,KCompletion times denoted C CN1, ,KIf jobs are scheduled in the order 1, ,K N the schedule is
X X X . . . X 1 2 3 n
A Gantt Chart of a Single Machine, No inserted idle time
Theorem: SPT minimizes flowtime on a single machineProof: 3 proofs
C X
N j X
X X
k j
jj
N
kk
j
j
N
jj
N
jj
N
k j kk jj
N
k j
= ==
=
= ≠=
∑ ∑∑
∑
∑ ∑∑
=
= − +
= +
=
1 11
1
1 1
1
1
0
( )
,
,
δ
δprecedes
else
• Pairwise interchange,• Hardy Littlewood Polya• Long wait for short only, simultaneously
For stochastic processing times:
MS&E324, Stanford University, Spring 2002 2-3 Gideon Weiss© manufacturing & control
Theorem: SEPT minimizes flowtime on a single machine
HH oo ww mm uu cc hh ww ii ll ll SS (( EE )) PP TT ss aa vv ee yy oo uu ??
We do a probabilistic analysis: Problem instance is chosen from apopulation of problems with X Fj ~ , and you schedule them.
Without information, E C list E XN N
mjj
N
kk
j
j
N( | ) ( )
( )
= ==∑ ∑∑= =
+
1 111
12
With full information,
E C SPT E X X NmN N
mjj
N
jj
N
k j kk jj
N( | ) ( )
( ), :
= = ≠=∑ ∑ ∑∑= + = +
−
1 1 11 1 2
12
δ
Theorem Assume X Fj ~ i.i.d, and you have information Tj i.i.d, anduse Y E X T Gj j j= ( | ) ~ to predict X j . Then
E C SEPT NmN N
m rm
djj
NF
G( | )( )
( )=∑ = +
−−
11 1
1
12
1σ
where:r X T X Yj j j j= =correlation correlation( , ) ( , )
and for any distribution: dm m
=−1 1 2:σ
. Typically, d d dF G≈ ≈ ≈ 12
MS&E324, Stanford University, Spring 2002 2-4 Gideon Weiss© manufacturing & control
Performance ratio:
E C List
E C SEPTr
md
jj
N
jj
NF
G
( | )
( | )( )=
=
∑
∑≈ −1
1
11
σ
WW ee ii gg hh tt ee dd FF ll oo ww tt ii mm ee
1/ / /ΣWjCj
Theorem: Smith's rule (cµ rule), schedule by decreasing EW
EXj
j
minimizes weighted flowtime E W Cj jj
N
=∑
1 on a single machine
MS&E324, Stanford University, Spring 2002 2-5 Gideon Weiss© manufacturing & control
time
weight
DD uu ee dd aa tt ee cc ii rr tt ee rr ii aa ..
1/prec/hmax
Lawler's algorithm for h h C h CN Nmax max( ( ), , ( ))= 1 1 K
Step 1: Classify job intor unscheduled jobs, scheduled jobs, and jobswith no successors:
J J n Jc= ∅ = =, { , , }, ©1 K jobs with no unscheduled successors
Step 2: calculate
j h X h Xj j
j J j Jj j
j Jc c
*
© ©©: ( ) min ( )*
∈ ∈ ∈∑ ∑=
move j*from Jc to J and update J©
Step 3: Stop when J J nc = ∅ =, { , , }1 K , schedule jobs in oppositeorder to their entry to J
Example: Jackson's rule, EDD (earliest due date) minimizes
L C dj jmax max= −
EDD for all on-time jobs minimizes U j∑
NN PP -- hh aa rr dd pp rr oo bb ll ee mm ss ..
1/rj/Lmax 1/rj/ΣΣΣΣCj are NP-hard scheduling problems
MS&E324, Stanford University, Spring 2002 2-6 Gideon Weiss© manufacturing & control
PP aa rr aa ll ll ee ll mm aa cc hh ii nn ee ss cc hh ee dd uu ll ii nn gg -- MM aa kk ee ss pp aa nn
P/ /Cmax
Let X XN1, ,K be integers. Can the be partitioned into two sets with
equal sums? All you need to do is check all subsets, but that is 2N
calculations.
In fact this problem "two partition" is NP-complete (binary sense), and"3-partition" etc. are (unary) NP-complete.
So scheduling jobs on two parallel machines is NP-hard. But inparctice this problem is easy:
(1) Simple heuristics:
Cmax |ListCmax |Opt
< 2 −1m
Cmax |LPTCmax |Opt
<43
−1
3m
(2) Karp and Karmarkar devised a fully polynomial approximationscheme.
(3) Probabilistically, for jobs from some reasonable distribuiton,
ECmax | LEPCmax |Opt
Population of Jobs
n→∞
→ 1
For stochastic jobs: Theorem: when processing times are exponential,
MS&E324, Stanford University, Spring 2002 2-7 Gideon Weiss© manufacturing & control
LEPT minimizes makespan on two machines.
PP aa rr aa ll ll ee ll mm aa cc hh ii nn ee ss cc hh ee dd uu ll ii nn gg -- FF ll oo ww tt ii mm ee
P/ /ΣΣΣΣCj
Theorem: SPT minimizes flowtime on parallel machines.
L L-1 12
A1A2AL-1
AL
Schedule looks like:
1 5 9
2 6 10
3 7
4 8
For stochastic jobs proof does not work. In fact theorem is notgenerally true.
But: If jobs are stochastically comparableX Y P X u P Y uST≥ ⇔ ≥ ≥ ≥( ) ( ),
then SEPT minimizes flowtime stochastically (Weber, Varaiya,Walrand)
MS&E324, Stanford University, Spring 2002 2-8 Gideon Weiss© manufacturing & control
PP aa rr aa ll ll ee ll mm aa cc hh ii nn ee ss -- WW ee ii gg hh tt ee dd FF ll oo ww tt ii mm ee
P/ /ΣΣΣΣWjCj is NP-hard. However, Smith's rule is a good heuristic:
. . . . . .
. . .. . .
1
2
3
4n
n-1
n-2
smith'srule OK
inefficiency
Worst case performance of Smith's rule is
RE w C SR
E w C Optj j
j j= = + ≈
∑
∑
|
|.1
212
1 2
This is for very many short jobs and a few long ones, all with almost thesame Smith's ratio.
However, for any set of jobs, on two parallel machines, if one usesSmith's rule
E w C SR E w Cm
m
w
EXDj j j j
j n
j
j∑ ∑− ≤
−
≤ ≤| |
( )maxπ
12
2
1
2
where:
D E X t X tt j j2 2= − >sup (( ) | )all jobs and all
MS&E324, Stanford University, Spring 2002 2-9 Gideon Weiss© manufacturing & control
SS hh oo pp mm oo dd ee ll ss aa nn dd mm uu ll tt iioo pp ee rr aa tt ii oo nn jj oo bb ss
Single machine
(ARRIVALS)(BATCH)
MACHINEDEPARTURES
Parallel machine
(ARRIVALS)
(BATCH)
MACHINE 2
DEPARTURESMACHINE 1
MACHINE 3
Machines in series flow-shop
(ARRIVALS)(BATCH)
MACHINE 1
BUFFER
MACHINE 2DEPARTURES
BUFFER
MACHINE 3
job-shop / queueing network
MS&E324, Stanford University, Spring 2002 2-10 Gideon Weiss© manufacturing & control
Job 1
Job 2Job 3
Job 4
MACHINE 3MACHINE 1
MACHINE 2 MACHINE 4
FF ll oo ww SS hh oo pp ss cc hh ee dd uu ll ii nn gg
For 2 machine, minimal makespan, F a b Cj j2 | , | max,
Johmson's rule is optimal:
Partition jobs into:
a bj j< Schedule these first, according to SPT on machine 1
a bj j≥ Schedule these next, according to LPT on machine 2
Proof:
makespan == =∑ ∑max
kj
j
k
jj k
Na b
1
This rule is actually really bad: In order to prevent starving on machine 2you create work for it as fast as you can and then keep the buffer betweenthe two machines as full as you can for as long as possible.
t i m e
buffer size
In fact (probabilistic analysis): For balanced machines:
E C N N
E C N N
max
max
| .
| .
Johnson
RAND
≈ +
≈ +
µ σ
µ σ
0 56
1 12If machines not balanced, the difference is even smaller:
E C E C O Nmax max| | (log )Johnson RAND− ≈
MS&E324, Stanford University, Spring 2002 2-11 Gideon Weiss© manufacturing & control
All other flowshop scheduling problems are NP-hard
JJ oo bb SS hh oo pp ss cc hh ee dd uu ll ii nn gg tthhee 1100xx1100 pprroobb ll eemmFirst Appeared in 1963 book of Muth and Thompson.
Solution found by Carlier and Pinson, 1988,Solution is 930 Lower bound 631.
1
2
3
4
5
6
7
8
9
10
job 1 job 2 job 3 job 4 job 5 Step#
1 29 1 43 2 91 2 81 3 14
2 78 3 90 1 85 3 95 1 6
3 9 5 75 4 39 1 71 2 22
4 36 10 11 3 74 5 99 6 61
5 49 4 69 9 90 7 9 4 26
6 11 2 28 6 10 9 52 5 69
7 62 7 46 8 12 8 85 9 21
8 56 6 46 7 89 4 98 8 49
9 44 8 72 10 45 10 22 10 72
10 21 9 30 5 33 6 43 7 53
1
2
3
4
5
6
7
8
9
10
job 6 job 7 job 8 job 9 job 10 Step#
3 84 2 46 3 31 1 76 2 85
2 2 1 37 1 86 2 69 1 13
6 52 4 61 2 46 4 76 3 81
4 95 3 13 6 74 6 51 7 7
9 48 7 32 5 32 3 85 9 64
10 72 6 21 7 88 10 11 10 76
1 47 10 32 9 19 7 40 6 47
7 65 9 89 10 48 8 89 4 52
5 6 8 30 8 36 5 26 5 90
8 25 5 55 4 79 9 74 8 45
Job shop scheduling is not just NP-hard: Even small problems are veryhard. Minimizing makespan for the above took 10 years to solve. Suchproblems will now be solved in half an hours run. 20x20x20 is still
MS&E324, Stanford University, Spring 2002 2-12 Gideon Weiss© manufacturing & control
intractable.
TT hh ee oo rr yy oo ff tt hh ee SS ii nn gg ll ee SS ee rr vv ee rr QQ uu ee uu ee
Arrivals: 0 1 2≤ < <A A L
Cumulative arrivals: A( )t
Processing X X1 2, K
Departures D D1 2, K
Cumulative arrivals: D( )t
Queue length Q t t t( ) ( ) ( )= −A D
Delayed Q tD( )
In service Q tS ( )
Sojourn W D Aj j j= −
Delay V W Xj j j= −
Virtual workload (system, delayed jobs, in service) V W( ), ( ), ( )t t tS
Policy: FIFO (FCFS),
LIFO (LCFS),
Priority (preemptive of non-preemptive),
EDD, SERPT, SEPT
Rates: arrival rate: λ =
→∞t
tt
lim( )A
MS&E324, Stanford University, Spring 2002 2-13 Gideon Weiss© manufacturing & control
service rate: µ = =→∞
=∑1 1
mm
X
nn
jjn
, lim
QQ uu ee uu ee aa rr ee tt hh ee rr ee ss uu ll tt ss oo ff vv aa rr ii aa bb ii ll ii tt yy
Consider a single server station:
Assume arrival every hour on the hour, and service lasts precisely 54minutes
8:00 9:00 10:00 11:00 12:00 13:00
λµ
ρλµ
λ= = = = = = <11
0 9 0 9 1, . , .m m
With the same rates, if interarrivals and or sevice are variable, we see:
8:00 9:00 10:00 11:00 12:00 13:00
For exponential interarrivals and services: Average queue length is
MS&E324, Stanford University, Spring 2002 2-14 Gideon Weiss© manufacturing & control
ρρ1
9−
=
LL ii tt tt ll ee '' ss ff oo rr mm uu ll aa
Let the arrival rate be λ , and assume that service is regular enough to have
WW
nn
jjn
=→∞
=∑lim
1. Then the long term average number of customers in
the system LT
Q t dtT
T=
→∞∫lim ( )
1
0 exists and is
L W= λ
. .
.
W1
W2
W3
W4
Wn
A(t)
D(t)
0 T
DD
AA
D A( )( )
( )( )
( )
( ) ( )tT t
WT
Q t dtt
T tWj
j
t T
jj
t1 1 1
1 0 1= =∑ ∫ ∑≤ ≤
FF rr aa cc tt ii oo nn bb uu ss yy tt ii mm eeConsider the server as a system, by Little's formula:
Average number in system ρ λ= m
Sysem has either 1 or is empty: ρ Fraction busy
MS&E324, Stanford University, Spring 2002 2-15 Gideon Weiss© manufacturing & control
1− ρ Fraction idle
PP AA SS TT AA
Let Z t( ) be a stochastic process, Λ( )t a Poisson process with events at
0 1< < < <T TnL L, and assume for all t that Λ Λ( ) ( ) :s t s t− >{ } isindependent of Z s s s t( ), ( ) :Λ ≤{ }. Then the following (if they exist) areequal:
Tj
j
t
T
T
tZ T
TZ t dt
→∞ = →∞∑ ∫=lim lim
( )( ) ( )
( )1 1
1 0Λ
Λ
WW oo rr kk cc oo nn ss ee rr vv aa vv tt ii oo nnVirtual work load increases at arrivals, and otherwise decreases at rate 1.
X1
X2 X3
X4X5
W(t)
t
MS&E324, Stanford University, Spring 2002 2-16 Gideon Weiss© manufacturing & control
This is entirely invariant for all policies.
SS ee rr vv ee rr ww oo rr kk ll oo aa dd ::The server has one customer at a time, with triangular workload
X1
X2
X3
X4 X5
W (t)
t
S
Again for all policies:
W
AA
A
sT
s
T
T
jj
T
Tt dt
TT
X
TE X= = =
→∞ →∞
=∫
∑
lim ( ) lim( )
/
( )( ) /
( )
12
20
2
1 2λ
DD ee ll aa yy ee dd ww oo rr kk ll oo aa dd ::Each customer contributes his work requirement times his delay.
X1
X2 X3
X4X5
W(t)
MS&E324, Stanford University, Spring 2002 2-17 Gideon Weiss© manufacturing & control
V
AA
A
= = = =→∞ →∞
=∫
∑
lim ( ) lim( )
( )( ) ( )
( )
T
T
T
j jj
T
Tt dt
T
T
V X
TE V E X V
1
0
1 λ ρ
MM // MM // 11 qq uu ee uu ee ::
Interarrivals are i.i.d. ~ exp( )λ . Service is i.i.d. ~ exp( )µ . ρλµ
= < 1.
This is a birth and death process,
µµµµ
0000 1111
µµµµ
n-1 n
µµµµ
. . . M/M/1
λλλλ λλλλ λλλλ
Steady state queue length distribution:
P Q j jj( ( ) ) ( ) , ,∞ = = − =1 0 1ρ ρ K
Steady state sojourn time Wj ~ exp( )µ λ− ,
Long term average sojourn time Wm
=−1 ρ
, delay V m=−ρ
ρ1MM // GG // 11 qq uu ee uu ee ::
Interarrivals are i.i.d. ~ exp( )λ . Service i.i.d general. ρλµ
= < 1.
V V E Xs= = = +PASTA
= W V + W ρ λ ( ) /2 2,
Hence (Pollatshek-Khinchine)
VE X
mcS=
−=
−+λ
ρρ
ρ1 2 11
2
2 2( )
where cX
Xs
j
j
22=
Variance(
Mean
)
( ) squared coefficient of variation of service time.
MS&E324, Stanford University, Spring 2002 2-18 Gideon Weiss© manufacturing & control
GG II // GG // 11 qq uu ee uu ee ::
Interarrivals are i.i.d. general. Service i.i.d general. ρλµ
= < 1.
cA A
A Aa
j j
j j
2 1
12=
−
−
−
−
Variance(
Mean
)
( ), c
X
Xs
j
j
22=
Variance(
Mean
)
( ),
squared coefficients of variation of interarrivals and service.
Kingman's inequality
V mc cA S≤
−+ρ
ρ1 2
2 2
and for heavy traffic:
ρρ
ρ≤≈ ≈
−+
11 2
2 2, V m
c cA S
Congestion as function of traffic intensity
ρ
ρ1 − ρ
Note: c cA S2 2, are 0 for no variability, are 1 for ~ exp, and are
typically ≈ .17 .
MS&E324, Stanford University, Spring 2002 2-19 Gideon Weiss© manufacturing & control
MM // GG // 11 bb uu ss yy pp ee rr ii oo dd ::Fraction of time busy is ρ , fraction idle is 1− ρ , busy and idle periodsaltenate, idle period is (M/G/1) ~ exp( )λ with expected length 1/ λ .
BPm
=−
=−
11 1λ
ρρ ρ
If first service is x : EFSBPx
=−1 ρ
Note
Average delay of a customer:m
1− ρ
Average busy periodm
1− ρ
A typical customer arrives to find ρ
ρ1− in queue, and when he leaves
queue will change by ~1. This is called the flashlight principle: Queuechange time scale is much slower than the customer time scale.
Note also that busy periods are very variable. An overwhelming majorityare very short, but a few are very long. The typical customer encounters a
MS&E324, Stanford University, Spring 2002 2-20 Gideon Weiss© manufacturing & control
very long one.
SS cc hh ee dd uu ll ii nn gg ww ii tt hh aa rr rr ii vv aa ll ss ,,SS cc hh ee dd uu ll ii nn gg ss ii nn gg ll ee ss ee rr vv ee rr qq uu ee uu ee ::
The problem 1/ rj /ΣCj is NP-hard.
We now consider scheduling a stream of jobs,
Arrivals: 0 1 2≤ < <A A L
Processing X X1 2, K
Departures D D1 2, K
We want to minimize:1
1ND Aj j
j
N( )−
=∑
Observe: This is equivalent to some of completion times, but for stream ofarrivals it is O( )1 , while for batch (all N jobs present at time 0) it isO N( ).
We can say:
(1) With preemptions: SRPT is optimal
This requires you know the processing time when a job arrives, and areallowed to preempt the job that is running.
Beautiful proof, won't give it here.
(2) M/G/1: SEPT minimizes expected flowtime, Smith's "cµ" ruleminimizws expected weighted flowtime.
V k =λE(X2 ) / 2
(1 − ρii=1
k −1∑ )(1− ρi
i=1
k
∑ )
(3) Heavy traffic: Lowest priority customers do all the waiting,
State space collapse:
MS&E324, Stanford University, Spring 2002 2-21 Gideon Weiss© manufacturing & control
W
W
E X j||
( )no prioritiesSEPT
of class of longest jobs
m=
FF ll uu ii dd aa nn dd DD ii ff ff uu ss ii oo nn AA pp pp rr oo xx ii mm aa tt ii oo nn oo ffGG II // GG // 11 QQ uu ee uu ee ::• FSLLN, FCLT, FSAT for renewal processes.
Service times: X X1 2, K
Mean is m, service rate is µ =1m
, squared coefficient of variation is cS2
Moments of order r > 2 exist.
Service completions count: S t n X X tn( ) max{ : }= + + ≤1 L
FSLLN:1n
S nt ta s( ) . . → µ
FCLT: nn
S nt t c BM ta s
S( ( ) ) ( ). .1 2− →µ µ
FSAT: S t t c BM t TSr( ) ( ) ( )/= + +µ µ 2 10
• Dynamics of GI/G/1 queue:
Q t Q A t S B t
B t dsQ s
t
( ) ( ) ( ) ( ( ))
( ) ( )
= + −
= >∫
0
1 00
• Reflection mapping: For x t( ) RCLL ( x t D( ) ∈ ), x( )0 0>
(i) z t x t y t( ) ( ) ( )= + ≥ 0
(ii) y y t( ) , ( )0 0= nondecreasing.
(iii) z t d y t( ) ( ( )) =∫ 0.
Then x t( ) determines z t y t( ), ( ) uniquely, (iii') y t( ) is minimal is
MS&E324, Stanford University, Spring 2002 2-22 Gideon Weiss© manufacturing & control
equivalent to (iii), and the mapping is Lipschitz continuous in D.
• Centering the queue balance:
Q t Q t A t t S B t B t
t B t t t
( ) [ ( ) ( ) ( ( ) ) ( ( ( )) ( ))]
[ ( )] ( ) ( )
= + − + − − −
+ − = +
0 λ µ λ µ
µ µX Y
• Fluid Approximation
Fluid scaling: z tn
z nt( ) ( )=1
Sequence of systems, 1
0 0n
Q Qn ( ) ( )→ ,
Xn t Q t( ) ( ) ( )→ + −0 λ µ
and the limits of the queue length and the busy time are obtained throughthe reflection mapping:
Q t Q t Q t
B t B tt t
t t
t Q t
n
n
( ) ( ) ( ) ( )
( ) ( )( )
min{ : ( ) }
→ = + −( )
→ =≤
+ − ≥
= =
+0
0
λ µ
τ
τ ρ τ τ
τ
• Diffusion Approximation
diffusion scaling: √( )( ) ( )
z tz nt z nt
n=
−
Sequence of systems,1
0 0n
Q Q nn n n n n( ) √( ), lim lim , lim ( )→ = = = −λ µ λ θ λ µ ,
MS&E324, Stanford University, Spring 2002 2-23 Gideon Weiss© manufacturing & control
√ ( ) √( ) √( ) ( ) ( )
√ ( ) √( ), √ ( ) ( ) √( )
X X
X X
nA S
n n n
t t Q t c c BM t
Q t t B tn
B nt t t
→ = + + +
→ = − → −
0
1 1
2 2θ λ
λΨ Φ
BB rr oo ww nn ii aa nn MM oo tt ii oo nn aa nn dd RR ee ff ll ee cc tt ee dd BB MM
A standard Brownian motion BM t( ) (Wiener process) is a stochasticprocess which
(i) BM t( ) has independendt stationary increments
(ii) BM BM t N t( ) , ( ) ~ ( , )0 0 0=
(iii) BM t( ) has continuous paths (almost surely)
Let BM t m x mt BM t N x mt tx ( ; , ) ( ) ~ ( , )σ σ σ2 2= + + +
Reflected Brownian motion is defined as the reflection of a Brownianmotion, for x ≥ 0:
RBM t m BM t m Y tx x( , ; , ) ( , ; , ) ( , )ω σ ω σ ω2 2= +
Here, Y t BM s m s tx( , ) sup{ ( , ; , ) : }ω ω σ= < <−2 0
and:
P Y t yy x mt
te
y x mtt
P RBM t m zz x mt
te
z x mtt
my
xmz
( ( ) ) ( ) ( )
( ( ; , ) ) ( ) ( )
/
/
≤ =+ +
−− − +
≤ =− −
−− − −
−Φ Φ
Φ Φ
σ σ
σσ σ
σ
σ
2
2 2
2
2
For m < 0, RBM t mx ( , ; , )ω σ 2 is positive Harris recurrent, with
MS&E324, Stanford University, Spring 2002 2-24 Gideon Weiss© manufacturing & control
P RBM t m z exmz( ( ; , ) ) /→ ∞ ≤ = −σ σ2 21
2
MM aa xx ii mm aa ll qq uu ee uu ee ll ee nn gg tt hh ii nn GG II // GG // 11We study the maximal queue length for the 1st n arrivals.
Lemma (Dai & W, 2001) Let interarrivals and services be u vi i, . Assume
E u E v f E ei iu vi i( ) ( ), ( ) ( ) ,( )> = < ∞ >+θ θθ
for some 0. Let τn bethe time of the nth arrival, and Q t( ) the queue length process, Q( )0 0= .There exists c such that for all n
P Q t c nnt n
(max ( ) log )01
< < > ≤τ
Proof By assumption, f ( )θ θ0 01 0< >for some . We have for everym l k≥ ≥ ≥0 0 1, ,
P u u v v f E em m l k m m ll u k( ) ( ) ( )+ + + + +
−+ + < + + ≤1 1 00 1L L θ θ
Hence
P Q t c n
P u u v v
n E e
c n n E e
t
m l nm m l c n m m l
u c n
u
n(max ( ) ( ))
( )
( ) ,
( ) log / log ( ) .
,( )
( )
0
01 1
2 0 1
0 13
< <
≤ ≤+ + + + +
−
−
> ≤
≤ + + < + +
≤
= −
τ
θ
θ
U L L
MS&E324, Stanford University, Spring 2002 2-25 Gideon Weiss© manufacturing & control