combinatorial algorithms for minimizing the weighted sum ......singleiteration(jobsetj) increase j...

22
Combinatorial Algorithms for Minimizing the Weighted Sum of Completion Times on a Single Machine James Davis 1 Rajiv Gandhi 2 Vijay Khothari 1 Department of Operations Research Cornell University 2 Department of Computer Science Rutgers University - Camden November 8, 2010 James Davis (Cornell University) Machine Scheduling 1 / 22

Upload: others

Post on 31-Jan-2021

0 views

Category:

Documents


0 download

TRANSCRIPT

  • Combinatorial Algorithms for Minimizing the WeightedSum of Completion Times on a Single Machine

    James Davis 1 Rajiv Gandhi2 Vijay Khothari

    1Department of Operations ResearchCornell University

    [email protected]

    2Department of Computer ScienceRutgers University - Camden

    [email protected]

    November 8, 2010

    James Davis (Cornell University) Machine Scheduling 1 / 22

  • Single Machine Scheduling (Off-line Problem)

    Input:J, set of jobspj , processing timesrj ≥ 0, release dateswj , weights

    Goal:Schedule jobs on a single machineNon-preemptive scheduleNo job can be scheduled before it’s release dateMinimize

    ∑j wjCj , weighted sum of completion times

    James Davis (Cornell University) Machine Scheduling 2 / 22

  • Single Machine Scheduling (Off-line Problem)

    James Davis (Cornell University) Machine Scheduling 3 / 22

  • Single Machine Scheduling (Off-line Problem)

    Cost = 1 · 1+ 6 · 8+ 9 · 4+ 13 · 3

    James Davis (Cornell University) Machine Scheduling 4 / 22

  • Single Machine Scheduling (Online Problem)

    Input:Jobs, Jpj , rj , wj ∀j ∈ J

    Goal:Schedule jobs on a single machineNon-preemptive scheduleMinimize

    ∑j wjCj , weighted sum of completion times

    Aware of job j at time rjSchedule fixed at time t without knowledge of jobs s.t. rj > t

    James Davis (Cornell University) Machine Scheduling 5 / 22

  • LP Formulation

    Notation (S ⊆ J):Sum of processing times

    p(S) =∑j∈S

    pj

    Sum of squared processing times,

    p2(S) =∑j∈S

    p2j

    James Davis (Cornell University) Machine Scheduling 6 / 22

  • Simple LP Formulation

    Cj denotes the completion time of job j

    min∑j∈J

    wjCj

    subject to Cj ≥ rj + pj, ∀j ∈ J∑j∈S

    pjCj ≥p(S)2 + p2(S)

    2, ∀S ⊆ J

    Cj ≥ 0, ∀j ∈ J

    James Davis (Cornell University) Machine Scheduling 7 / 22

  • Previous Work

    Single Machine Scheduling:NP-Hard (Lenstra et al.)Off-line PTAS known (Afrati et al.)Off-line 1.6853-Appx. Alg. via LP rounding (Goemans et al.)Online 1.6853-Appx. Alg. via LP rounding (Goemans et al.)

    Using Simple LP (provide upper bound on int. gap):Off-line 3-Appx. Alg. via LP rounding (Hall et al.)

    James Davis (Cornell University) Machine Scheduling 8 / 22

  • Our Contribution

    Use Simple LP (provide upper bound on int. gap):Off-line (1+

    √2)(≈ 2.42)-Approximation Algorithm

    Online 3-Approximation Algorithm

    James Davis (Cornell University) Machine Scheduling 9 / 22

  • Main Criteria

    Recall Notation:p(S) =

    ∑j∈S pj

    New Notation:j∗ ∈ J has highest release date (rj value)j′ ∈ J has lowest wjpj value

    Main Criteriarj∗ > p(J)

    James Davis (Cornell University) Machine Scheduling 10 / 22

  • List Algorithm

    Off-line List AlgorithmJ ′ ← Jwhile J ′ 6= ∅ do

    j∗ ← job with largest rj valueif rj∗ > p(J ′) thenRemove j∗ from J ′

    else if rj∗ ≤ p(J ′) thenj′ ← j ∈ J ′ with lowest wjpj valueRemove j′ from J ′

    end ifend whileSchedule jobs in the reverse order that they were removed from J ′

    James Davis (Cornell University) Machine Scheduling 11 / 22

  • Intuitive Analysis

    If rj∗ > p(J ′) then Cj∗ is approximately dominated by rj∗

    rj∗ = 11 > 10 = p(J ′)

    James Davis (Cornell University) Machine Scheduling 12 / 22

  • Intuitive Analysis

    If rj∗ > p(J ′) then Cj∗ is approximately dominated by rj∗

    rj∗ = 11 > 10 = p(J ′)

    James Davis (Cornell University) Machine Scheduling 13 / 22

  • Intuitive Analysis

    If rj∗ ≤ p(J ′) then Cj∗ is approximately dominated by p(J ′)

    rj∗ = 2 ≤ 8 = p(J ′)

    James Davis (Cornell University) Machine Scheduling 14 / 22

  • Intuitive Analysis

    Smith’s RuleSchedule the most useless job last (lowest wjpj )

    j′ is job 1 (w1p1 = 1)

    James Davis (Cornell University) Machine Scheduling 15 / 22

  • Intuitive Analysis

    Final Schedule

    James Davis (Cornell University) Machine Scheduling 16 / 22

  • Simple LP Formulation

    Cj denotes the completion time of job j

    min∑j∈J

    wjCj

    subject to Cj ≥ rj + pj, ∀j ∈ J∑j∈S

    pjCj ≥p(S)2 + p2(S)

    2, ∀S ⊆ J

    Cj ≥ 0, ∀j ∈ J

    James Davis (Cornell University) Machine Scheduling 17 / 22

  • Simple Dual

    Introduce βS ∀S ⊆ J and αj ∀j ∈ J

    max∑j∈J

    αj(rj + pj)+∑S⊆J

    βS

    (p(S)2 + p2(S)2

    )subject to αj + pj

    ∑S:j∈S

    βS ≤ wj, ∀j ∈ J

    αj ≥ 0, ∀j ∈ JβS ≥ 0, ∀S ⊆ J

    James Davis (Cornell University) Machine Scheduling 18 / 22

  • Primal-Dual AlgorithmSimplified Setting:

    Single Iteration (job set J)Increase αj∗ or βJ

    rj∗ > p(J ′)Increase αj∗αj∗ = wj∗

    rj∗ ≤ p(J ′)Increase βJβJ =

    wj′pj′

    Dual Constraintαj + pj

    ∑S:j∈S βS ≤ wj

    James Davis (Cornell University) Machine Scheduling 19 / 22

  • Primal-Dual Algorithm

    Off-line Primal-Dual AlgorithmJ ′ ← Jwhile J ′ 6= ∅ do

    j∗ ← job with largest rj valueif rj∗ > p(J ′) thenαj∗ ← wj∗ − pj∗

    ∑S:j∗∈S

    βS

    Remove j∗ from J ′else if rj∗ ≤ p(J ′) then

    j′ ← j ∈ J ′ with lowest wjpj valueβJ′ ←

    wj′pj′−∑

    S:j∈S βS

    Remove j′ from J ′end if

    end whileSchedule jobs in the reverse order that they were removed from J ′

    James Davis (Cornell University) Machine Scheduling 20 / 22

  • Primal-Dual Algorithm

    TheoremUsing the off-line primal dual algorithm:

    Cost of Schedule ≤ (1+√2) · Dual Feasible Solution ≤ (1+

    √2) · OPT

    James Davis (Cornell University) Machine Scheduling 21 / 22

  • Thank You!

    James Davis (Cornell University) Machine Scheduling 22 / 22