Stability using fluid limits: Illustration through an example "Push

Finite Buffer Fluid Networks with Overflows
Yoni Nazarathy,
Swinburne University of Technology, Melbourne.
Stijn Fleuren and Erjen Lefeber,
Eindhoven University of Technology, the Netherlands.
Talk Outline
• Background: Open Jackson networks
• Introducing finite buffers and overflows
–Interlude: How I got to this problem
• Fluid networks as limiting approximations
• Traffic equations and their solution
• Almost discrete sojourn times
Open Jackson Networks
Jackson 1957, Goodman & Massey 1984, Chen & Mandelbaum 1991
Problem Data:
, , P
1
Assume: open, no “dead” nodes
Traffic Equations (Stable Case):
M
i
i
pi j
M
pi  1   pi j
j 1
i   i    j p j i
j 1
    P '
  ( I  P ') 1
Traffic Equations (General Case):
i   i     j   j  p j i
M
M
j 1
    P '   
LCP   ( I  P ')  , ( I  P ')

Open Jackson Networks
Jackson 1957, Goodman & Massey 1984, Chen & Mandelbaum 1991
Problem Data:
, , P
1
Assume: open, no “dead” nodes
Traffic Equations (Stable Case):
M
i
i
pi j
M
pi  1   pi j
j 1
i   i    j p j i
j 1
    P '
  ( I  P ') 1
Product Form “Miracle”:
M
 j
lim P  X 1 (t )  k1 ,..., X M (t )  k M    1 
 
t 
j 1 
j
M
  j

 
 j



kj
Modification: Finite Buffers and Overflows
Problem Data:
K1
 ,  , P, K , Q
1
Assume: open, no “dead” nodes,
no “jam” (open overflows)
qi j
i
Ki
i
pi j
Explicit Solutions:
M
pi  1   pi j
j 1
M
qi  1   qi j
j 1
K M M
Generally No
Exact Traffic Equations:
Generally No
A Practical (Important) Model:
Yes
Our Contribution
(in progress)
Limiting Traffic Equations:
K1
1
qi j
i
Ki
    P '       Q '(   )
i
Efficient Algorithm for Unique Solution:
pi j
M
pi  1   pi j
j 1
M
qi  1   qi j
j 1
K M M
Limiting Deterministic Trajectories
(N )
 X (t )

lim supt 
 x(t )   0
N 
 N

Limiting Sojourn Time Distribution
S (N )  S
P( S  k )  1    T k 1
Interlude: How I got to this problem
Control of queueing networks:
Server 1
Server 2
PUSH

PULL
1
1
PULL
2
PUSH
2

Output process, D(t), asymptotic variance:
E  D (t ) 
lim
t 
t
vs.
Var  D (t ) 
lim
t 
t
2
3
BRAVO
effect for
M/M/1/K
load
When K is Big, Things are “Simpler”
for K big,
out rate    
overflow rate        (   ) 
Scaling Yields a Fluid System
A sequence of systems: N  1, 2,...
 (N )   
 (N )  N 
 (N )  N K
Make the jobs fast and the buffers big by taking N  
The proposed limiting model
is a deterministic fluid system:
Fluid Trajectories as an Approximation
 X ( N ) (t )

lim supt 
 x(t )   0
N 
 N

Traffic Equations (at equib. point)
out rate    
overflow rate  (   )
i  i     j   j  p ji     j   j  q ji
M
M
j 1
j 1

or

    P '       Q '(   )
or
LCP ( I  Q ') (  ( I  P ')  ) , ( I  Q ') ( I  P ')
1
1

LCP (Linear Complementarity Problem)
a
M
,G 
M M
LCP (a, G ) :Find z , w 
M
such that,
w  Gz  a,
w  0, z  0,
w ' z  0.
The last (complemenatrity) condition reads:
wi  0  zi  0 and zi  0  wi  0.
Min-Linear Equations as LCP
Find  :     B (   )
    B
0 
0  
(   ) '(    )  0
w  ( I  B) z    ( I  B) 
 
w    , z   
z  0, w  0
w' z  0
LCP(  ( I  B)  , I  B)
Existence, Uniqueness and Solution
Definition: A matrix, G 
M M
is a "P"-matrix if the
determinants of all (2n  1) principal submatrices are positive.
Theorem (1958): LCP(a, G) has a unique solution
for all a 
M
if and only if G is a "P"-matrix.
"P"-matrix means that the complementary cones "parition"
Immediate naive algorithm
with 2M steps
1 0  w1   g11
0 1   w    g

  2   21
g12   z1   a1 

g22   z2   a2 
 0
 
1
We essentially assume that our
1
matrix ( G  ( I  Q ') ( I  P ') )
is a “P”-Matrix
We have an algorithm
(for our type of G)
taking M2 steps
n
C{1,2}
C{2}
1 
 
0
 g 
 11 
  g 21 
C
 g 
 12 
 g22 
a 
 1
a2 
C{1}
Sojourn Time  Time in system of customer arriving
to steady state FCFS system
S ( N )  Sojourn time of customer in N'th scaled system
We want to find the limiting distribution of S
(N )
Sojourn Times Scale to a Discrete Distribution!!!
P  S (N )  x
x
“Molecule” Sojourn Times
F  {1,..., s}
i  i for i  F
F  {s  1,..., M }
i  i for i  F
Observe,
time through i  F 
For job at entrance of buffer
iF
w. p. 
:
Ki
i
time through i  F  0
i
enters buffer i
i
 
w. p.  1  i
 i

 qij routed to entrace of buffer j

 
w. p.  1  i
 i

 q i leaves the system

A job at entrance of buffer
iF:
A “fast” chain and “slow” chain…
routed almost immediately according to
P
The “Fast” Chain and “Slow” Chain
Example: M
 4
,
4

 1,  i 1,
1
2
i 1
K1
K2
F  {1, 2}, F  {3, 4}
4
 a
4
p
start
j 1
4
 a
j 1
j
1
1
ji
:
1’
1
1j
a j1
 j 1 1 j j 0
1   q1
 4 1 
p1   p1 j a j 0
j 1
4
p
“Fast” chain
on {0, 1, 2, 1’, 2’, 3’, 4’}:
ai j  Absorbtion probability
in j {0,1, 2} starting in i'
j 1
2’
 1 
1   q1i
 1 
1j
a j2
0
2
3’
“Slow” chain on {0, 1, 2}
DPH distribution (hitting time of 0)
transitions based on “Fast” chain
p4 i
p4
4’
E.g: Moshe Haviv (soon) book: Queues, Section on “Shortcutting states”
The DPH Parameters (Details)
F  {1,..., s}, F  {s  1,..., M }
“Fast” chain
BM s
 1
0 
 1







s 
 0
s 


0M s s


CM  M
 1
1  
1







0 M  s s



0



0 M  M  s   Q   0 M  s s M  s   P

I M s 


1 s

s


AM s  ( I  C ) 1  B
“Slow” chain
Tss
  I s
0s  M  s   P  A
 s1 
1
M

j 1
S ~ DPH (Tss ,1s )
T  A
j
P( S  k )  1    T k 1s1
Sojourn Times Scale to a Discrete Distribution!!!
“Almost Discrete” Sojourn Time Phenomena
Taken from seminar of Avi Mandelbaum, MSOM 2010 (slide 82).
Summary
– Trend in queueing networks in past 20 years:
“When don’t have product-form…. don’t give up: try asymptotics”
– Limiting traffic equations and trajectories
– Molecule sojourn times (asymptotic) – Discrete!!!
– Future work on the limits.