Upload
ngongoc
View
216
Download
2
Embed Size (px)
Citation preview
Queues with Multi-Type Jobs and Multi-Type Servers
under
First Come First Served Discipline
Gideon Weiss
University of Haifa
Newton Institute, SCS program, 2010
joint work and work by:
Ivo Adan, Rene Caldentey, Cor Hurkens,
Ed Kaplan, Rishi Talreja, Jeremy Visschers, Ward Whitt
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 1
Multi-type queues:
Customer types i=a,b,c, . . . S(i) servers of i
Server types j=1,2,. . .,K C(j) customers of j
Bipartite system graph
• Loss systems: Servers m1, m2, . . ., mK (Adan Hurkens W, ‘10)
Intenet traffic, switches, wireless,
• Stable FCFS queues: (Visschers ‘00, Visschers Adan W, ‘10)
Manufacturing
• Overloaded system with Reneging: : Fa, Fb, Fc, . . . (Talreja Whitt ‘07)
Call centers with skill based routing
• Infinite matching (Kaplan ‘88, Caldentey Kaplan ‘02, Caldentey Kaplan W ‘09),
Public housing, adoptions, kidney transplants
C ab c
1 3 2
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 2
Multi-type queues:
• Loss systems: Servers m1, m2, . . ., mK
• Stable FCFS queues:
• Overloaded system with Reneging: : Fa, Fb, Fc, . . .
• Infinite matching
m1 m3 m2
a c
m1 m2
ccb/c b/cb/c a/ca/b
m3
aa/b/ca/b/ca/b/c
b c a a b
3 1 2 2
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 3
Overloaded system with reneging (Talreja Whitt, ManSci’08)
Overloaded systems, arrival rates ! "i > ! µj
general arrival streams, general service times,
reneging patience distributions Fi
Uniform acceleration regime:
servers are busy almost all the time
queues force almost all customers to wait until almost reneging
on a fluid scale GLOBAL FCFS is achieved
- everyone waits approximately W.
Question: How many type i are served by type j rij
equivalently: what is the rate at which of i by j service #ij
!1 !2 !3
µ1µ2 µ3
!iFi (W )" = µ j"
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 4
Service rates in overloaded system
!iFi (W )" = µ j"
Question: How many type i are served by type j rij
equivalently: what is the rate at which of i by j service #ij
Solve for W from:
Then solve
!iji"C( j)# = µ j , j = 1,…, J
!ijj"S(i)# = $iFi (W ), i = 1,…, I
These may have no positive solution, unique solution of many solutions
Solved for trees, complete graphs, hybrids
1 2 3
1 2 3
1 2 3
1 2 3
4
4
Non-
Unique
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 5
Infinite matching (Kaplan ‘84, Caldentey Kaplan ‘02,Caldentey, Kaplan, W ‘09)
“. . . Households applying for public housing are allowed to specify those housing
projects in which they are willing to live; when a public housing unit becomes newly
available, of those households willing to live in the associated housing project, the
one that has been waiting the longest is o!ered the unit. . . .”
This is not a customer server situation ---
Housing units leave with the households
Other examples:
Adoptions,
Kidney Transplants,
Auctions ?
In assigning housing to entitled families several housing projects are available
(Kaplan, ‘84.’88).
b c a a b
3 1 2 2
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 6
FCFS Infinite matching
Customer types i=1,2,. . .,I S(i) servers of i
Server types j=1,2,. . .,J C(j) customers of j
System graph G
Customer types and server types are i.i.d from $, %
Matching is FCFS
1 2 3
1 2
C
S
!1 !2 !3
"1 "2
c1
c2
c3
cn!3
cn!2
cn!1
cn
s1
s2
s3
sn!3
sn!2
sn!1
sn
Fraction of i->j matches in first n :
Question: when does
what is
ri, jn
ri, jn a.s.! "! ri, j
ri, j
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 7
Existence of rates:
Balance equations: Matching rates must satisfy:
ri, jj!S(i)
" =#i , ri, ji!C( j)
" = $ j ,
Necessary condition: Nonnegative solution exists iff
for every subset C and every subset S
!(C) " #(S(C)), #(S) "!(C(S)), *
!
!
C S
a b
i
j
!"i
# j
Proof
max-flow / min-cut
But: solution may be non-unique
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 8
Markov chain
Consider the matching for s1 :
he will take the first cm in c1 ,c2,. . . that matches him,
leave a geometric number of unmatched c from C(s1)
Xn
i = 1,…, I
Define: the Markov chain of ordered unmatched customers
state space is words in alphabet of
Consider then sn+1 he will first look at those c left by s1 . . . sn
If he finds a match he will take the first, and leave one less unmatched behind.
Else he will look for a match among those not considered previously,
until he finds a match, adding a geometric number of c from C(sn+1) .
!(C) < "(S(C)), "(S) <!(C(S)), **
Theorem: If the Markov chains are ergodic then ri, jn
n!" a.s.# !### ri, j
Conjecture: The Markov chain X is ergodic iff for all non-trivial S,C:
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 9
Almost complete system graph
Customer types i=0,1, . . ., I
Server types j=0,1, . . .,I
Type 0 match everything,
Type i matches all except i
!X(ekx) = "k
#k(1$ "k
%&'
()*
x
1+#i"i
1$#i $ "ii=1
I
+%
&'(
)*
1 2 3
1 2
C
S
!1 !2 !3
"1 "2
3
"3
Z!I
The unmatched of the X process
will always be customers of a single type
X process moves on the integer axes in I dimensions
Ergodic iff
ek the k unit vector:
ri, j =!i" j (1#!i )(1# " j )#! j"i$% &'(1#!i # "i )(1#! j # " j )
1+!i"i
1#!i # "ii=1
I
()
*+,
-.
1!"i ! #i
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 10
`NN’ model states and transitions
!3 <"1, "3 < !1, "1 + !1 <1Necessary condition for ergodicity:
111
111
11
11
1
10
3
3
33
33
333
333
3333
3333
33333
33333
3
2
33
32
333
332
3333
3332
33333
33332
33
2x
333
32x
3333
332x33333
3332x
333
2xx
3333
32xx
33333
332xx
2
3
32
33
332
333
3332
3333
33332
33333
2x
33
32x
333
332x
3333
3332x
33333
333333
3332xx
3333333
33332xx
33333333
333332xx
333333
33332x
3333333
333332x
333333
333332
33332x
333333
333332x
3333333
333332
333333
!3"3
!1"1 !1"1
!1"1
!3"3
!1"1
(1#!1)(1# "1 )
(1#! 3)(1 # "1 )
!3"2
!2"3
!3 (1 #"1)
(1#! 3)"1 !2"1
!1 (1 #"1)
!3 (1 #"1)
!2"1!2"3
!1"2!3"2
!2 (1 # "1 )
!"2
1# !"2
333333
333333
!2"1 1# !"21
!"2 ="2
"2 + "3
1 2
23
C
S
!1
3
1
!2 !3
"3 "2 "1
State space and transitions:
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 11
A Lyapunov function of Fayolle, Malyshev and Menshkov.
To show positive recurrence we need to show that
E(!f (x, y)) " #h < 0
0 5 10 15 20
-10
-5
0
5
10
We were able to constructthe right f (choice of u,v,w)
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 12
1st attempt to solve: Caldentey-Kaplan ‘02 Algorithm
Let indicator that customer type i matches server type j,
is the incidence matrix of the bipartite system graph
Let
AijA I ! J
Li, j =!i Ai, j
!(C( j)), Mi, j =
" jAi, j
"(S(i))
For the complete graph matching rates are Fi, jall
=!i" j
Start with Fi, j0= Gi, j
0= Fi, j
allAi, j
It does not satisfy the balance equations,
fi0=!i " Fi, j
0
j
# row shortfalls are:
g j0= ! j " Fi, j
0
i
#column shortfalls are:
f!1
= 0,
g!1
= 0
fk+1
= gkLT
gk+1
= fkM
Fi, jk+1
= Fi, jk+ (!1)
k!1fik!1
Mi, j + (!1)kg jkLi, j
Converges to a solution of balance equations: Fi, jk
k!"# !## Fi, j
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 13
2nd attempt to solve: Quasi independent cells
Kaplan considered a hydraulic analogue: pipes connect customers and server types,
with as pressures. Using a limiting case non-Newtonian `power law’ fluid,
reduced to the model of `quasi independence’ of contingency tables.
!,"
In two dimensional contingency tables one often considers the hypothesis of independence of the 2 factors
If and are the row and column marginal frequencies, cell (i,j) should have frequency !i ! j !i" j
Often tables have missing cells, because of missing observations, forbidden or impossible combinations
Aijxiy j =!i ,j
" Aijxiy j = # j ,i
"
One postulates the existence of abstract full table marginals, and obtains quasi-independent model:
Then: solves the balance equations hi, j = xiy j
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 14
Attempts failed - examplesExample 1: the almost complete 3 type graph - Caldentey Kaplan works, Quasi-independent fails
1 2 3
1 2
C
S
!1 !2 !3
"1 "2
3
"3
Example 2: one more customer type - Quasi-independent seems to work, Caldentey Kaplan fails
1 2 3
1 2
C
S
!0 !2 !3
"1 "2
3
"3
!1
0
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 15
Back to Queueing - Exact asymptotics for the `N’ model
Adan, Foley, and Macdonnald, ‘08 analyze the queueing model for the `N’ system graph:
Arrivals and processing are memoryless,
Service is FCFS.
If system is empty, and type a arrives, machine 1 will take him w.p
m2
jobs
machines
!a
m1
!b
µ2µ1
a b
"1#"
!
States:
(0,0) -empty
(1,0) -one job, machine 2 working on it
(0,y) - machine 1 working, machine 2 idle y-1 type b jobs waiting
(x,y) - both machines working, y-1 type b waiting head of line,
x-1 additional unspecified waiting
Results: steady state exact asymptotics, for large (x,y)
with explicit values of the constants for various
combinations of parameter values and various directions
!(x, y) ~ c(x / y)"a + "bµ1 + µ2
#$%
&'(
x"bµ1
#$%
&'(
y1 2
12
C
S
! 1"!
1" # #
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 16
Procduct form solution for the `N’ queueing system
m2
jobs
machines
!a
m1
!b
µ2µ1
a b
"1#"
When this has a product form solution:
µ1
n
m
(n,m2 ,m,m1)!a + !b
(m,m1)
µ1
!a
!b
µ2p0
µ2p1
µ2p2
µ2p3
µ2"4
µ1
(n,m1,0,m2 )
!a + !b
µ2p0
µ2p1
µ2p2
µ2p3
(0,m1)
(0,m2 )
(0)
µ1
(n,m2 ,0,m1)
!a + !b
µ2p0
µ2p1
µ2p2
µ2p3
µ2"4
(0,m1,0,m2 ) (0,m2 ,0,m1)
µ2
µ1
!a# + !b!a (1$#)
!a + !b µ2µ1µ1
!b!a
µ2
! =!* ="a
2"a + "b
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 17
A reversible multi-type loss system (Adan, Hurkens, W ‘10)
Servers j=1, . . .,J working at rates µj exponential
Customers of types i=1,. . . I arriving at rates "i Poisson
Graph G : C(j) customer types served by server j
This is a loss system:
State: set of busy machines {M1, M2,. . ., Mk}
Completions: {M1, M2,. . ., Mk}--> {M1, M2,. . ., Mk-1} rate µMk
Arrivals: {M1, M2,. . ., Mk-1} --> {M1, M2,. . ., Mk} rate "Mk{M1, M2,. . ., Mk-1}
Assume MC reversible. Detailed balance:
!Mk{M1,…,Mk"1} = !iP(i
i#C(Mk )
$ chooses server Mk | {M1,…,Mk"1} busy)
!{M1,…,Mk}µMk
= !{M1,…,Mk"1}#Mk{M1,…,Mk"1}
Solution:
!{M1,…,Mk}µMk
= !{"}
!M1{"}!M2
{M1}!!Mk{M1,…,Mk#1}
µM1 µM2 ! µMk
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 18
Assignment condition
The solution:
!{M1,…,Mk}µMk
= !{"}
!M1{"}!M2
{M1}!!Mk{M1,…,Mk#1}
µM1 µM2 ! µMk
Makes sense only if for all permutations:
!M1
{"}!M2{M1}!!Mk
{M1,…,Mk#1} = !M1{"}!
M2{M1}!!
Mk{M1,…,Mk#1}
!M (S) = !ii"C(S)# 1+
! j (S$M )
!M (S$ j)j%S#
&
'()
*+
This uniquely determines
Question: Can you find
!M (S) = P(ii"C(M )
# , M | S)
Answer: Yes, always, maxflow of does it
This maxflow exists by monotonicity
!M (R) " !M (S), M #R $ S
a b
i
M
SC(S)
!i
!M (S)
"
P(i,M | S)
!ii"C(S)
#
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 19
Product form for multi-type FCFS (Visschers ‘02,Visschers, Adan, W ‘10)
Servers j=1, . . .,J working at rates µj exponential
Customers of types i=1,. . . I arriving at rates "i Poisson
Graph G : C(j) customer types served by server j
FCFS service, random assignment to idle servers.
Thm: if assignment is as in loss system, then MC obeys partial balance, and has
product form stationary distribution
! s( ) = ! 0( )"M1
{#}"M2{M1}!"Mk
{M1,…,Mk$1}
µM1 (µM1 +µM2 ) !(µM1 +!+ µMk)%1n1%2
n2!%
k
nk
%l ="U (M1,…,Ml )
µM1 +!+ µMl
M1M2Mk
n1
! ! !
nk
U (M1)U (M1,M2) U (M1,…,Mk )
n2
Gideon Weiss, University of Haifa, Infinite Matching, ©2008 20
Conclusion and future work
(1) Loss system may be insensitive
(2) In the product form model, when r-->1, we approach the !"matching model
(3) r-->1 in product form solution may solve the !"matching model
(4) In the overloaded with reneging, when r-->1, we approach the !"matching
model
(5) Solution of !"matching model may solve the overload reneging model
(6) The !"matching model is independent of processing and arrival distributions