optimal scheduling among intermittently unavailable servers
DESCRIPTION
Optimal Scheduling Among Intermittently Unavailable Servers. Simon Martin & Isi Mitrani University of Newcastle upon Tyne. Model. m 1 , x 1, h 1. Arrivals l. policy. m 2 , x 2, h 2. Parameters. Arrival rate = l At queue i (i=1,2) Average service time = 1/ m i - PowerPoint PPT PresentationTRANSCRIPT
Optimal Scheduling Among Intermittently Unavailable
ServersSimon Martin & Isi Mitrani
University of Newcastle upon Tyne
Model
Arrivals
1, 1
policy
2, 2
Parameters
• Arrival rate = At queue i (i=1,2)
• Average service time = 1/i
• Average operative period = 1/i
• Average repair time = 1/i
• Average holding cost = ci
Problem
A scheduling policy specifies, for every possible system state, whether an incoming job which finds that state is sent to queue 1 or to queue 2.
Find a policy that minimizes average holding costs.
Solution
• The problem is tackled using the tools of Markov decision theory.
• The optimal policy can be computed numerically by – uniformizing the continuous time Markov
process,– replacing it with an equivalent discrete time
Markov chain; and – truncating the state space to make it finite.
• The instantaneous transition rates are modified, so that the transition rate out of any state is 1, with the addition of transitions which do not change the current state.
Uniformization
222111 ,max,max
1;
max
ijii
ijij
jij
i
rq
rq
r
State of discrete time Markov chainS = (i, j, b1, b2, a)
• Number of jobs in server 1: i• Number of jobs in server 2: j
• Availability of server 1: b1
• Availability of server 2: b2
• Arrival event: a
Stationary policy for minimizing total average discounted costs over
an infinite horizon
S
dSVSSqdcScSV ,min
This equation can be solved iteratively.
The interesting case is →∞
Minimize total average cost over a finite horizon of n steps
S
nd
n SVSSqdcScSV 1,min
Solve recurrences in n steps, starting from
00 SV
Steady-state average cost per step, independent of the starting state
n
SVV n
n lim
Obtained by simulation
Policies examined
• Heuristic (smallest expected conditional holding cost per job)
• Random• Selective (send only to operative servers;
N.Thomas)• Shortest Queue• Optimal (minimal steady-state cost per step)
Varying
0
10
20
30
40
50
60
70
80
90
100
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Lambda
Ave
rag
e C
ost
Random
Selective(Nigel)
ShortestQueue
Heuristic
Optimal
Varying
0
50
100
150
200
250
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Mu1
Ave
rag
e C
ost
Random
Selective(Nigel)
ShortestQueue
Heuristic
Optimal
Varying
0
20
40
60
80
100
120
140
160
180
200
0 0.2 0.4 0.6 0.8 1 1.2
Xi1
Av
era
ge
Co
st
Random
Selective(Nigel)
ShortestQueue
Heuristic
Optimal
Varying Holding Cost
0
20
40
60
80
100
120
140
0 0.5 1 1.5 2 2.5 3 3.5 4
C1
Ave
rag
e co
st
Random
Selective(Nigel)
ShortestQueue
Heuristic
Optimal