Preemptive Coordination Mechanisms for Unrelated Machines By Fidaa Abed Max-Planck-Institut für Informatik Chien-Chung Huang Humboldt-Universität zu Berlin
Jan 16, 2016
Preemptive Coordination Mechanisms for Unrelated
Machines
By
Fidaa AbedMax-Planck-Institut für Informatik
Chien-Chung HuangHumboldt-Universität zu Berlin
2
Unrelated Machine Scheduling
Classical problem
m unrelated machines
n jobs
pij – processing time of job i on
machine j
Each job is owned by a selfish
user
User goal: minimize his
completion time
System goal: minimize the
worst completion time
(Makespan)
Job1
Job2
Job3
M1 M2 M3
Job 2 Job 1
Job 3
1
1
1
3 3
3
3
3
3
3
1
3
Scheduling Policies
3
5
25
3
2Longest first
2
3
5Shortest first
4
The Game
Nash equilibrium (NE): no user wants to change his machine
NE can be far from optimal
Cost (NE) = 3Cost (OPT) = 1
PoA = Cost (Worst NE)/Cost (OPT)PoA = 3/1
PoA can be unbounded
Job 1Job 3
Job1
Job2
Job3
M1 M2 M3
1
1
1
3 3
3
3
Job 1
3
3
3
Job 3Job 21
Job 2
Job 1
Longest first
5
Coordination Mechanisms
Clever scheduling policies
Examples:Longest firstShortest first
Can be preemptiveJob1
Job2
Job1
Job2
t1
t2
Job2
Job1
t2
t1
Job2
Job1
Job1t1
t2
6
Goal: Minimize PoA
Shortest firstJob 3Job 2
Job1
Job2
Job3
M1 M2 M3
1
1
1
3 3
3
3
3
3
Job 1
Job 2Job 1 Job 3
Job 1
7
History of Coordination Mechanisms
Introduced by Christodoulou et. al in 2004
Mechanism PoA Preemptive?
Longest first Unbounded No
Shortest first m No
Inefficiency based
logm No
BCOORDlogm
loglogmYes
8
Open Problem
Can we achieve Constant PoA using preemption or randomization?
Azar, Jain, and Mirrokni (SODA 2008)
Caragiannis (SODA 2009)
9
Our Results(1)
• All deterministic mechanisms, even with preemption, if they are
– symmetric– satisfy Independence of Irrelevant Alternatives
(IIA) property
have the PoA Ω ( ) .log mloglog
m
10
Our Results(2)
• All randomized mechanisms, even with preemption, if they are
– symmetric– unbiased
have the PoA Ω ( ) .log mloglog
m
11
Symmetry
t_z
x
yy
t_z
t_y t_y
a b
x
zz
t_xt_xAll known mechanisms are symmetric.
12
Independence of Irrelevant Alternatives (IIA) Property
• If job z is “preferred” over job y by machine a, then this “preference” should not change because of the availability of some other job x.
x
y
t1(z)
t2(y)
z
t3(x)
• appears as axiom in voting theory and logic
• was assumed by Azar et. al. [SODA 2008] in their lower bound
• All known mechanisms have this property
13
IIA Property
Lemma: IIA each machine has order over the jobs
- Proof is omitted.
The order based on:
1. Jobs IDs (non-anonymous case)
2. Machines IDs (anonymous case)
14
Anonymous Case
1 1
1 1
1 1
M4
M2 M3 M4M1
Job
Job
Job
M3 M1 M2
15
Lower Bound for Non-Anonymous Case
node = machine
Edge = job that can go to two machines
Processing time of all jobs = 1
For k =3
k
k-1
m = 1+k + k(k-1) + …… + k!
k = Θ( )
log mloglog
m
16
Lower Bound for Non-Anonymous Case
Cost (OPT) = 1
Cost (NE) >= k
PoA >= k
k = Ω ( )
PoA = Ω ( )log mloglog
m
log mloglog
m
17
Lower Bound for Non-Anonymous Case
x
x
y
z
w
y
z
wxy
yxz
zxyw
afmlkqponjwqgex
18
Lower Bound for Non-Anonymous Case
ab
t_z
x
zy
t_y
t_x t_x x
y
z
a b
19
Lower Bound for Anonymous Case
m1
m2
m3
m1m3 m2
20
Lower Bound for Anonymous Case
21
Lower Bound for Anonymous Case
ab
t_z
x
zy
t_y
t_x t_x x
y
z
a b
22
Positive Result
• The previous lower bound was because of the unbounded inefficiency.
• If the inefficiency is bounded by a constant C then we can achieve constant PoA by known mechanisms.
• Ex: Inefficiency-based mechanism achieves PoA <= C + 2 log C + 2 = O(C).
23
Open Problem
• BCOORD is optimal but it is not known whether it guarantees Pure Nash Equilibrium.
• Open Problem: Design a mechanisms that achieves Θ( ) and guarantees the convergence to Pure Nash Equilibrium.
log mloglog
m
24
Conclusion
• Achieving constant PoA using preemption or randomization is impossible.
• If the inefficiency is bounded then we can achieve constant PoA.