abdolhamid ghodselahi: serving online requests with mobile servers
TRANSCRIPT
Serving Online Requests with Mobile Servers
Abdolhamid Ghodselahi University of Freiburg, Germany
joint work withFabian Kuhn (University of Freiburg)
Presented at ISAAC 2015, Nagoya, JapanDecember 9-11, 2015
Our online problem:
๐ points are given
๐ mobile servers
Online requests
service cost = ๐
service cost = ๐
service cost = ๐service cost = ๐
service cost = ๐service cost = ๐
Goal: Minimize #movements
service cost = ๐
#movements = 1#movements = 2
๐ฎ = 1๐ฎ = 2๐ฎ = 4๐ฎ = 1๐ฎ = 3๐ฎ = 1
What if some algorithm moves no server at all?!
Feasible Configuration:
Any algorithm that solves the problem must satisfy the following condition at all time steps :
Problem condition: ๐ฎ < ๐ผ โ ๐ฎโ + ๐ฝ
๐ฎ โ Current service cost of any algorithmโข Service cost is not cumulative over time
๐ฎโ โ Optimal current service cost =Minimum service cost among all configurations
๐ผ โฅ 1 and ๐ฝ โฅ 0 are two given parameters
Recap:
๐ points are given ๐ mobile servers Online requests
Requests need to be servedโข At the requested pointโข By a remote server
A request has to be served atall time steps after it is issuedโข Reassignment is allowed
Problem condition must be satisfied at all time steps
Goal: Minimize #movements
service cost = ๐
service cost = ๐
service cost = ๐
service cost = ๐
Our Model VS. ๐-Server/Paging
Our Model
Requests are served
to serve However, some on the current service cost
๐-Server/Paging
Requests are served when they are issued
Servers serve No service cost
Known Results for ๐-Server & Paging Deterministic
โข ๐-Server conjecture: Competitive factor is ๐[Manasse, McGeoch, & Sleator 1990]
โข Competitive factor of 2๐ โ 1 for ๐-Server[Koutsoupias & Papadimitriou 1995]
โข Any deterministic algorithm is ฮฉ ๐ -competitive[Sleator & Tarjan 1985]
โข Least recently used (LRU) algorithm is ๐-competitive[Sleator & Tarjan 1985]
Randomized
โข Competitive factor of ฮ log ๐ 2 log ๐ 3
[Bansal, Buchbinder, Madry, & Naor 2011]
Outline
1. Motivation & Model
2. Minimizing #Movements
a. Lower-Bound
3. Minimizing #Movements + Service Cost
a. Upper-Bound
b. Lower-Bound
4. Future Work
Any deterministic online algorithm is
ฮฉ ๐ -competitive
2. Minimizing #movements
Proof Sketch ๐ โถ Any deterministic online algorithm (ALG)
๐ช โถ Any optimal offline algorithm (OPT)
Two cases:
๐ > ๐/2 :
โข Competitive factor is โฅ ๐
๐ โค ๐/2 :
โข Competitive factor is โฅ ๐ โ ๐
โฅ max ๐, ๐ โ ๐
โฅ ๐ 2 โ ฮฉ(๐)
๐ โค ๐/2 โถ Main Idea
large enough #requests
๐ฎ๐ โฎ ๐ผ โ ๐ฎโ + ๐ฝ
ALG must move some server(s)
OPT moves to a point where#requests is large at all time steps
points without servers
๐ = 3 , ๐ = 1
Assume ๐ผ = 1
Problem condition:โ๐ก โถ ๐ฎ๐(๐ก) < ๐ฎโ(๐ก) + ๐ฝ
Repeat for ๐ , ๐ = 1 โโฅ ๐ โ 1 #movements
๐ฎโ = ๐ฝ๐ฎ๐ = 2๐ฝ
๐ฎ๐ = ๐ฎโ + ๐ฝ
๐ โค ๐/2 : Simple Example#Movements by ALG
๐ฝ
๐ฝ
๐ฎโ = ๐ฝ๐ฎ๐ = ๐ฝ
๐ฎ๐ < ๐ฎโ + ๐ฝ
2๐ฝ
2๐ฝ
๐ฎโ = 3๐ฝ๐ฎ๐ = 4๐ฝ
๐ฎ๐ = ๐ฎโ + ๐ฝ
๐ฎโ = 3๐ฝ๐ฎ๐ = 3๐ฝ
๐ฎ๐ < ๐ฎโ + ๐ฝ
๐ = 3 , ๐ = 1
Assume ๐ผ = 1
OPT knows the sequence in advance
Problem condition:โ๐ก โถ ๐ฎ๐ช(๐ก) < ๐ฎโ(๐ก) + ๐ฝ
Repeat for ๐ , ๐ = 1 โโค 1 #movements
๐ฎโ = ๐ฝ๐ฎ๐ช = 2๐ฝ
๐ฎ๐ = ๐ฎโ + ๐ฝ
๐ โค ๐/2 : Simple Example#Movements by OPT
๐ฝ
๐ฝ
๐ฎโ = ๐ฝ๐ฎ๐ช = ๐ฝ
๐ฎ๐ < ๐ฎโ + ๐ฝ
2๐ฝ
2๐ฝ
๐ฎโ = 3๐ฝ๐ฎ๐ช = 3๐ฝ
๐ฎ๐ช < ๐ฎโ + ๐ฝ
๐ โค ๐ 2 โถ Reduction to any ๐, ๐
โฏ
๐
โฏ
๐ โ 1โฏ โฏ
๐ โ 1
๐
โ
All algorithms can only move this server
โฅ ๐ โ ๐ #movements by ALG and โค 1 by OPT
3. Minimizing Combined CostThe objective is to minimize the
Current service cost+ #Movements
This modification in the objective helps us to be more competitive against OPT
A natural greedy algorithm (denoted by ๐) is introduced which provides an almost tight bound
Minimizing combined cost is closer to an online variant of
mobile facility location problem [Friggstad & Salavatipour FOCSโ08]
The algorithm does nothing as long as ๐ฎ๐ < ๐ผ โ ๐ฎโ + ๐ฝ
It greedily moves some server(s) as soon as ๐ฎ๐ โฎ ๐ผ โ ๐ฎโ + ๐ฝ
Greedy Approach:
Decrease current service cost as much as possible
Maximal improvement = 6๐ฝ
Greedy Algorithm
5๐ฝ
2๐ฝ
6๐ฝ
8๐ฝ
๐ฎโ = 7๐ฝ๐ฎ๐ = 14๐ฝ
๐ฎ๐ < 2๐ฎโ + ๐ฝ
๐ฎโ = 7๐ฝ๐ฎ๐ = 15๐ฝ
๐ฎ๐ = 2๐ฎโ + ๐ฝ
7๐ฝ๐ฎโ = 7๐ฝ๐ฎ๐ = 9๐ฝ
๐ฎ๐ < 2๐ฎโ + ๐ฝ
Our online algorithm is 1 + ๐ -competitive
for every constant ๐ > 0,
at the cost of an additional additive term
Results
Any deterministic online algorithm cannot get
a better competitive factor than almost similar
above upper-bound
Upper-Bound: Proof Sketch
Goal: Minimize the combined cost
๐ฎ๐ < ๐ผ โ ๐ฎโ + ๐ฝ ๐๐ โค ?
๐ฎ๐ช +๐๐ช โฅ ๐ฎโ
๐๐ โค ๐ โ ๐ฎโ + ฮ(๐ log ๐)
General Service Cost Function Recall:
๐๐ฃ ๐ฆ โ 0, ๐ ๐๐๐ฃ๐๐ ๐๐ก ๐ฃ๐ฆ, ๐๐กโ๐๐๐ค๐๐ ๐
Generalization:๐๐ฃ ๐ฅ, ๐ฆ โ Service cost of ๐ฃ if ๐ฅ servers and ๐ฆ requests at ๐ฃ
The function has to satisfy some natural properties: Monotonicity (in ๐ฅ and ๐ฆ)
Effect of adding additional servers to a node ๐ฃ
โข should become smaller (convexity in ๐ฅ)
โข should not decrease if #requests gets larger
The upper-bound result holds for this generalization
Both lower-bound results even hold for the previous service cost
4. Future Work
With respect to minimizing the #movements:
โข Study randomized online algorithms
With respect to minimizing the combined cost:
โข Study the online variant of mobile facility location problem (OMFLP) in general metrics
OMFLP definition
our lower-bound already holds for any det. online algorithm that solves OMFLP
Thanks for your attention