on some stochastic load balancing problems
play

On Some Stochastic Load Balancing Problems Anupam Gupta Carnegie - PowerPoint PPT Presentation

On Some Stochastic Load Balancing Problems Anupam Gupta Carnegie Mellon University Joint work with Amit Kumar, IIT Delhi Viswanath Nagarajan, Michigan Xiangkun Shen, Michigan (appeared at SODA 2018) 1 / 26 Optimization under Uncertainty


  1. On Some Stochastic Load Balancing Problems Anupam Gupta Carnegie Mellon University Joint work with Amit Kumar, IIT Delhi Viswanath Nagarajan, Michigan Xiangkun Shen, Michigan (appeared at SODA 2018) 1 / 26

  2. Optimization under Uncertainty Question: How to model/solve problems with uncertainty in input/actions? ◮ data not yet available, or obtaining exact data difficult/expensive ◮ actions have uncertainty in outcomes ◮ (my talk) we only have stochastic predictions 2 / 26

  3. Optimization under Uncertainty Question: How to model/solve problems with uncertainty in input/actions? ◮ data not yet available, or obtaining exact data difficult/expensive ◮ actions have uncertainty in outcomes ◮ (my talk) we only have stochastic predictions Goal: get algos making (near)-optimal decisions given predictions. 2 / 26

  4. Optimization under Uncertainty Question: How to model/solve problems with uncertainty in input/actions? ◮ data not yet available, or obtaining exact data difficult/expensive ◮ actions have uncertainty in outcomes ◮ (my talk) we only have stochastic predictions Goal: get algos making (near)-optimal decisions given predictions. worst-case analysis (vs. queueing perspective, cf. Mor’s talk) ◮ Relate to performance of best strategy on worst-case instance 2 / 26

  5. Optimization under Uncertainty Question: How to model/solve problems with uncertainty in input/actions? ◮ data not yet available, or obtaining exact data difficult/expensive ◮ actions have uncertainty in outcomes ◮ (my talk) we only have stochastic predictions Goal: get algos making (near)-optimal decisions given predictions. worst-case analysis (vs. queueing perspective, cf. Mor’s talk) ◮ Relate to performance of best strategy on worst-case instance given predictions (vs. all-adversarial model as in competitive analysis) 2 / 26

  6. Approximation Algorithms for Stochastic Optimization High Level Model: ◮ Inputs random variables X 1 , X 2 , . . . with known distributions ◮ (for today) assume discrete distributions, explicit access to them 3 / 26

  7. Approximation Algorithms for Stochastic Optimization High Level Model: ◮ Inputs random variables X 1 , X 2 , . . . with known distributions ◮ (for today) assume discrete distributions, explicit access to them ◮ outcomes not known a priori, revealed over time ◮ want to optimize, say, E [ objective ]. 3 / 26

  8. Approximation Algorithms for Stochastic Optimization High Level Model: ◮ Inputs random variables X 1 , X 2 , . . . with known distributions ◮ (for today) assume discrete distributions, explicit access to them ◮ outcomes not known a priori, revealed over time ◮ want to optimize, say, E [ objective ]. ◮ many different models ⋆ e.g., adaptive vs. non-adaptive ⋆ e.g., single-stage vs. multi-stage 3 / 26

  9. Approximation Algorithms for Stochastic Optimization High Level Model: ◮ Inputs random variables X 1 , X 2 , . . . with known distributions ◮ (for today) assume discrete distributions, explicit access to them ◮ outcomes not known a priori, revealed over time ◮ want to optimize, say, E [ objective ]. ◮ many different models ⋆ e.g., adaptive vs. non-adaptive ⋆ e.g., single-stage vs. multi-stage Today: centralized load-balancing problem, minimizing E [ makespan ]. 3 / 26

  10. today’s problem 4 / 26

  11. (Classical) Load Balancing Problem Schedule n jobs on m machines to minimize makespan . Graham list-scheduling from 1966. A B C D 5 / 26

  12. (Classical) Load Balancing Problem Schedule n jobs on m machines to minimize makespan . Graham list-scheduling from 1966. A B C D 5 / 26

  13. (Classical) Load Balancing Problem Schedule n jobs on m machines to minimize makespan . Graham list-scheduling from 1966. A B C D 5 / 26

  14. (Classical) Load Balancing Problem Schedule n jobs on m machines to minimize makespan . Graham list-scheduling from 1966. A makespan B C D 5 / 26

  15. much work on the deterministic problem Simplest Model: Identical machines: ◮ List Scheduling: 2-approx [Graham 66] , PTAS [Hochbaum, Shmoys 87] . 6 / 26

  16. much work on the deterministic problem Simplest Model: Identical machines: ◮ List Scheduling: 2-approx [Graham 66] , PTAS [Hochbaum, Shmoys 87] . Most General: Unrelated machines: ◮ Jobs have different sizes on different machines. 6 / 26

  17. much work on the deterministic problem Simplest Model: Identical machines: ◮ List Scheduling: 2-approx [Graham 66] , PTAS [Hochbaum, Shmoys 87] . Most General: Unrelated machines: ◮ Jobs have different sizes on different machines. ◮ 2-approx [Lenstra, Shmoys, Tardos 90] , better for special cases. 6 / 26

  18. Stochastic Load Balancing Job j on machine i takes on size X ij (r.v. with known distribution) Today: these r.v.s are independent Find an assignment to minimize expected makespan:   m �  . E max X ij  i =1 j ∈ J i ? ? ? ? ? 7 / 26

  19. Stochastic Load Balancing Job j on machine i takes on size X ij (r.v. with known distribution) Today: these r.v.s are independent Find an assignment to minimize expected makespan:   m �  . E max X ij  i =1 j ∈ J i ? ? ? ? ? 7 / 26

  20. Stochastic Load Balancing Job j on machine i takes on size X ij (r.v. with known distribution) Today: these r.v.s are independent Find an assignment to minimize expected makespan:   m �  . E max X ij  i =1 j ∈ J i ? ? ? ? ? ? ? ? ? ? 7 / 26

  21. Stochastic Load Balancing Job j on machine i takes on size X ij (r.v. with known distribution) Today: these r.v.s are independent Find an assignment to minimize expected makespan:   m �  . E max X ij  i =1 j ∈ J i ? ? ? ? ? 7 / 26

  22. Stochastic Load Balancing Job j on machine i takes on size X ij (r.v. with known distribution) Today: these r.v.s are independent Find an assignment to minimize expected makespan:   m �  . E max X ij  i =1 j ∈ J i ? ? ? ? ? minimize expected makespan 7 / 26

  23. Related Work: Stochastic Job Sizes #P-hard to evaluate objective exactly O (1)-approximation for identical machines [Kleinberg, Rabani, Tardos 00] . Better results for special classes of job size distributions [Goel, Indyk 99] . ◮ Poisson distributed job sizes: 2-approximation. ◮ Exponential distributed job sizes: PTAS. 8 / 26

  24. Related Work: Stochastic Job Sizes #P-hard to evaluate objective exactly O (1)-approximation for identical machines [Kleinberg, Rabani, Tardos 00] . Better results for special classes of job size distributions [Goel, Indyk 99] . ◮ Poisson distributed job sizes: 2-approximation. ◮ Exponential distributed job sizes: PTAS. What about general unrelated case? 8 / 26

  25. Main Result Theorem An O (1) -approx algo for minimizing E [ makespan ] on unrelated machines. 9 / 26

  26. Roadmap and Assumptions Identical machines case Ideas needed for unrelated machines (and sketch of proof) 10 / 26

  27. Roadmap and Assumptions Identical machines case Ideas needed for unrelated machines (and sketch of proof) By scaling, assume E [ OPTmakespan ] = 1. 10 / 26

  28. Roadmap and Assumptions Identical machines case Ideas needed for unrelated machines (and sketch of proof) By scaling, assume E [ OPTmakespan ] = 1. Assume each job is “small”: Pr[ size > E [ OPTmakespan ]] = 0. ◮ Easy extension to general sizes 10 / 26

  29. Deterministic Surrogate Find deterministic quantity as a surrogate for each r.v. Do optimization over these deterministic quantities Surrogate = expected size? 11 / 26

  30. Deterministic Surrogate Find deterministic quantity as a surrogate for each r.v. Do optimization over these deterministic quantities Surrogate = expected size? (No!) 11 / 26

  31. Deterministic Surrogate Find deterministic quantity as a surrogate for each r.v. Do optimization over these deterministic quantities Surrogate = expected size? (No!) Bad Example Type 1: size 1 (deterministic). 1 Type 2: size Bernoulli(0 , 1) r.v. with p = √ m . log m E [ mkspan ] = log log m m machines m − √ m jobs: expectation 1 1 m jobs: expectation √ m 11 / 26

  32. Deterministic Surrogate Find deterministic quantity as a surrogate for each r.v. Do optimization over these deterministic quantities Surrogate = expected size? (No!) Bad Example Type 1: size 1 (deterministic). 1 Type 2: size Bernoulli(0 , 1) r.v. with p = √ m . m − √ m machines m − √ m machines log m E [ mkspan ] = m − √ m jobs: expectation 1 √ m log log m √ m machines √ m 1 m jobs: expectation √ m 11 / 26

  33. Deterministic Surrogate Find deterministic quantity as a surrogate for each r.v. Do optimization over these deterministic quantities Surrogate = expected size? (No!) Bad Example Type 1: size 1 (deterministic). 1 Type 2: size Bernoulli(0 , 1) r.v. with p = √ m . m − √ m machines m − √ m machines log m E [ mkspan ] = m − √ m jobs: expectation 1 √ m log log m √ m machines √ m 1 m jobs: expectation √ m 11 / 26

  34. Deterministic Surrogate Find deterministic quantity as a surrogate for each r.v. Do optimization over these deterministic quantities Surrogate = expected size? (No!) Bad Example Type 1: size 1 (deterministic). 1 Type 2: size Bernoulli(0 , 1) r.v. with p = √ m . m − √ m machines m − √ m machines m − √ m machines m − √ m machines log m E [ mkspan ] ≤ 2 E [ mkspan ] = √ m log log m √ m machines √ m machines √ m 1 m jobs: expectation √ m 11 / 26

Recommend


More recommend