control for stochastic models via diffusion approximations amy ward, ans lecture series 2008...
TRANSCRIPT
Control for Stochastic Modelsvia Diffusion Approximations
Amy Ward, ANS Lecture Series 2008
1.Control of High Volume Assemble-to-Order Systems• Queue-length control via tracking policies
2.Control of a Many-Server Queueing Model• Workload division
Fair Dynamic Routing in Large-Scale Heterogeneous-Server
Systems
Amy Ward
Joint work with Mor Armony
ANS Lecture II 2008
The Inverted-V Model
NK
K
K 21
• Calls arrive at rate (Poisson process).
• K server pools.
• Service times in pool k are exponential with rate k
N1
1
¹ 2 > ¹ 1
¹ > ¹
Experienced employees on averageprocess requests faster than new hires.Gans and Shen (2007)
…
The ProblemRouting: When an incoming call arrives to an empty queue, which agent pool should take the call? The objective is to minimize steady-state wait time.
¹ 2 > ¹ 1
¹ > ¹
x = y
NK
K
K 21
N1
1
…
No. The Slow-server Problem.2 servers.It is sometimes necessary to keepcustomers waiting even when the slower server is idle in order not to starve the faster server.Threshold control is optimal.Lin and Kumar (1984).
Is an exact analysis possible?
General multi-heterogeneous server case is still open.(Vericourt and Zhou, 2006)
The ProblemRouting: When an incoming call arrives to an empty queue, which agent pool should take the call?
¹ 2 > ¹ 1
¹ > ¹
x = y
Armony (2005) shows routing to the fastest server first (FSF) asymptotically minimizes the steady-state wait time.BUT … asymptotically only the slow servers have any idle time.Is this fair? Do we care?
NK
K
K 21
N1
1
…
Fairness
¹ 2 > ¹ 1
¹ > ¹
x = y
Call centers care!
Employee burnout and turnover.
Increased employee turnover leads to worse performance (Whitt 2006).
Call Centers address fairness byrouting to the server that has idled the longest (LISF).
How does LISF perform?
Do any other fair policies perform better?
NK
K
K 21
N1
1
…
The Fairness Problem
¹ 2 > ¹ 1
¹ > ¹
x = y
Minimize E[Waiting Time]
Subject to:
E[# of idle servers of pool k]= fk
E[Total # of idle servers]
*All in steady-state
NK
K
K 21
N1
1
…
How to determine f?
)1()( then ,)(/)( If
1211
221111 ff
NNfEIfEIEI kk
21
1121 implies
/)(1
NN
Nf
NEI kkk
¹ 2 > ¹ 1
¹ > ¹
x = y
2 classes.Expected waiting time is decreasing in the pool 1 idleness proportion f1. So we would like to choose high f1.
Should we ensure all servers have the same utilization?Expected utilizationof a pool k server
Any fairness criterion that involves individual serverutilization translates into a choice for f1
The same effective processing rate for all servers: 2211
The Fairness Problem
¹ 2 > ¹ 1
¹ > ¹
x = y
Minimize E[Waiting Time]
Subject to:
E[# of idle servers of pool k] = fk
E[Total # of idle servers]
Solution Approach:
1. Solve approximating diffusion control problem.2. Translate solution to original system.
NK
K
K 21
N1
1
…
Literature Review• Conventional Heavy Traffic, Parallel Server
Systems– Harrison (1998), Bell and Williams (2001) (2005)
• The Limit Regime– Halfin and Whitt (1981)
• The Inverted V Model– Armony (2005), Tezcan (2006), Atar (2007) Gurvich and Whitt (2007)
• Fairness literature in EE and CS– Deals with fairness towards flows/customers– Avi-Itzhak et al (2006), Weirman (2007)
• Fairness literature in human resources– Deals with the effect fairness on employee
performancs– Cohen-Charash et al (2001), Colquitt et al (2001)
The Asymptotic Regime
¹ 2 > ¹ 1
¹ > ¹
x = y
0,0 ,1
, As
1
1
k
K
k k
kkk
K
k kk
aa
oaN
oN
X̂ ¸ = X ¸ ¡ N ¸p
N ¸hX̂ ¸
i += scaled queue length
hX̂ ¸
i ¡= scaled # of idle servers
(under the assumption of work conservation)
NK
K
K 21
N1
1
…
Fairness
¹ 2 > ¹ 1
¹ > ¹
x = y
Call centers care!
Employee burnout and turnover.
Increased employee turnover leads to worse performance (Whitt 2006).
Call Centers address fairness byrouting to the server that has idled the longest (LISF).
How does LISF perform?
Do any other fair policies perform better?
NK
K
K 21
N1
1
…
The Longest-Weighted-Idle Server First (LWISF) Policy
K
m mm
kk
N
N
1
• LISF might not obtain the desired idleness constraint
• fk(LISF)=
• To fix this, we propose LWISF: LWISF routes to pool k if wkik> wmim, where ik is the idle time of the server that has been idle the longest in pool k
fk=
• Proposition: LWISF asymptotically satisfies the idleness constraint
BUT: Does it minimize E[Waiting time]? Can we do better?
K
m mmm
kkk
wN
wN
1/
/
The Asymptotic Behavior of LWISF
0 ,
0 ,)( 1
x
xfxxm
K
k kk
LWISF is asymptotically equivalent to a preemptive policythat, at all times, balances the workload between server poolsby requiring the fraction of idle servers in pool k is fk,where f1+f2+ +fK=1.
For the preemptive policy . as ~ XX
The proposition proof is due to Stone’s theorem.The asymptotic equivalence is due to Atar (2007).
X has infinitesimal mean
X has infinitesimal variance 2.
The Asymptotic Equivalence
The diffusion limit is identical. There is state-space collapse.
.at poolin servers idle ofnumber theesapproximat )(
. at time servers idle ofnumber theesapproximat )(
at time waitingcallers ofnumber theesapproximat )(
tktXNf
ttXN
t.tXN
k
Note that under the preemptive policy, there is state-space collapse for each . The state is the number in system.
The LWISF policy maintains fixed ratios betweenthe number of idle servers in each pool.
Fairness
¹ 2 > ¹ 1
¹ > ¹
x = y
Call centers care!
Employee burnout and turnover.
Increased employee turnover leads to worse performance (Whitt 2006).
Call Centers address fairness byrouting to the server that has idled the longest (LISF).
How does LISF perform?
Do any other fair policies perform better?
NK
K
K 21
N1
1
…
The Fairness Problem
¹ 2 > ¹ 1
¹ > ¹
x = y
Minimize E[Waiting Time]
Subject to:
E[# of idle servers of pool k]= fk
E[Total # of idle servers]
*All in steady-state
NK
K
K 21
N1
1
…
The Diffusion Control Problem
1
),(
)(2)(),(ˆ)0(ˆ)(ˆ
1
1
0
K
k k
K
k kk
t
u
xuuxm
tBdssusXmXtX
x = y2
},...,1{ ,)(ˆ)(ˆ)( subject to
)(ˆmin
KkXEfXuE
XE
kk
Can you guess the solution?
Threshold Control
)(2)(),(ˆ)0(ˆ)(ˆ
0if }{1
0 if ),(
thatso
})(ˆ{1)(
has ) and0(with
levelsat control dA threshol
0
1 1
1
0
11
tBdssusXmXtX
xLxLx
xuxm
LtXLtu
L L
,...,LL
t
K
k kkk
kkk
K
K
x = y2
Solving the Diffusion Control Problem
K
k kk fddXE
1
for )(ˆ
},...,1{for )(ˆ)( subject to
)(ˆ minimize
KkdfXuE
XE
kk
X̂
Step 1: Observe that
Step 2: Hence an equivalent DCP is
Step 3: We can now formulate the Lagrangian and solve.
K
k kkk dfXuEXE1
)(ˆ)()(ˆmin
Solving the Diffusion Control Problem Cont.
K
k kkk dfXuEXE1
)(ˆ)()(ˆmin
What are the correct penalty parameters?
},...,1{for )(ˆ1)(ˆ
ly,equivalent or,
,)(ˆ)(
1 KkdfLXLXE
dfXuE
kkk
kk
The ones under which
.)(ˆ)(ˆ
, control admissiblean under
)(2)(),(ˆ)0(ˆ)(ˆ
satisfies and
},,...,1{for )(ˆ)(
has that ˆany for Then,
}.,...,1{for )(ˆ1)(ˆ
such that be let and , control
thresholdunder thediffusion thebe ˆLet
:
*
0
1
*
XEXE
u
tBdssusXmXtX
KkdfXuE
X
KkdfLXLXE
LL
X
t
kk
kkk
Theorem
Browne and Whitt (1995).
How do we find the threshold levels?
Policy Translation: (2 classes)
Lx
NxL
xN
Use FSF.
No servers idle.
Number in system is x.
Use SSF.
Threshold level: L ¸ = N ¸ ¡ Lp
N ¸
xN
NxL
LxL
Lx
K
21
10
.constant positive somefor level Threshold kkk LNLNL
Use FSF excluding pool K (the fastest).
Use FSF excluding pool K-1.
Use FSF.
No servers idle.
. Assume 21 KLLL
Number in system is x.
Policy Translation: (K classes)
Asymptotic Optimality (a.o.)kK
k k
k fEI
EI
1);(
);(lim
Asymptotic Feasibility:
Asymptotic Optimality: 1. ¼ is asymptotically feasible, and2. If ¼’ is asymptotically feasible then
Conjecture 1: The preemptive Threshold Policy is a.o.(ext. of Stone’s theorem; Atar, Budhjiraja and Ramanan (2007))
But what about the non-preemptive threshold policy?Showing a.o. would require a non-continuous form of s.s.c..
);(ˆinflim);(ˆsuplim
WEWE
²-Asymptotic Optimality (²-a.o.) ²-Asymptotic Optimality:
1. ¼ is asymptotically feasible, and2. If ¼’ is asymptotically feasible
then
The ²–Threshold Policy
);(ˆinflim);(ˆsuplim WEWE
X
Death rate
slope ¹2
slope ¹1
L N
²-Threshold Policy: Construction
)(ˆ)(ˆ)(ˆ * ul XXX
)(,ˆ
X
)(ˆ)(ˆlim *
)(,0
XEXE
)(ˆ uX)(ˆ lX
)(ˆ)(ˆ *
)(, XEXE
1. Construct upper and lower bound diffusion processes.
2. Construct a process whose drift is a convex combination of and
3. Choose such that
4. The diffusion is -optimal because
)(ˆ)(ˆ *
)(, XEXE
5. Hence for any >0, there exists () such that
Policy Translation: -Threshold Policy
Lx
LxL
NxL
xN
Use FSF.
No servers idle.
Number in system is x.
Use SSF.
Threshold level: L ¸ = N ¸ ¡ Lp
N ¸
Depends
Adjustment: NN
Thm: The ²-Threshold Policy is ²-a.o.
);(ˆ);(ˆ
;ˆ and );(ˆ);(ˆ
1
21
1
fTHIETHIE
THIETHXETHQE
)(ˆ);(ˆinflim *XEQE
as ;ˆ;ˆ THXTHX
Proof:
)(ˆ);(ˆ and )(ˆ);(ˆ ** XETHXEXETHXE
1. Diffusion construction
2. Weak convergence (Stone)
3. State-space collapse (G&W) and Tightness and UI
4. Asymptotic lower bound (DCP solution)
5. Little’s Law
Asymptotic Performance (Predicted)
)(XE
1 = 1, 2 = 2, = 1, = 1.5, 2 = 2 = 3
f(L)0.000
0.100
0.200
0.300
0.400
0.500
0.0000 0.2000 0.4000 0.6000 0.8000 1.0000
Slow Server Idleness Proportion
Threshold
LWISF
FSF
f1
E [X̂ (1 )]+
Asymptotic Performance (Predicted)
0.00%
1.00%
2.00%
3.00%
4.00%
5.00%
6.00%
7.00%
8.00%
0.0 0.2 0.4 0.6 0.8 1.0
1 = 1, 2 = 2, = 1, = 1.5, 2 = 2 = 3
)(ˆ)(ˆ)(ˆ
LWISFTPLWISF XEXEXE
f(L)
Asymptotic Performance (Simulation)
1 = 1, 2 = 2, = 1, = 1.5, 2 = 2 = 3, N1=300, N2=200, ¸=674
A Simulation Comparison of the Threshold and LWISF Policy
0.0000002.000000
4.0000006.000000
8.00000010.000000
12.000000
0 0.5 1
Slow Server Idleness Proportion
E[N
um
be
r o
f W
ait
ing
C
us
tom
ers
]
Threshold
LWISF
Accuracy of Idleness Constraint
1 = 1, 2 = 2, = 1, = 1.5, 2 = 2 = 3, N1=300, N2=200, ¸=674
Slow Server Idleness Proportion
0.000000
0.200000
0.400000
0.600000
0.800000
1.000000
0.000000 0.500000 1.000000
Predicted
Sim
ula
ted
Threshold
LWISF
Predicted
Summary
• Formulation of the server fairness problem.
• Solution of the approximating diffusion control problem.
• Construction of threshold policy for the original system.
• ²-Threshold Policy (²TP) is ²–asymptotically
optimal.
• TH outperforms LWISF.
Further Research
• Non-ldling assumption• Multi-skill environment• Server Compensation Schemes
Acknowledgement: Rami Atar, Itay Gurvich, Tolga Tezcan & Assaf Zeevi