Online Vector Scheduling Debmalya Panigrahi Duke University slides - PowerPoint PPT Presentation

Online Vector Scheduling Debmalya Panigrahi Duke University slides es Work done with: Maryam Shadloo Nat Kell Sungjin Im Janardhan Kulkarni (MSR  UMN) (UC Merced) (Duke) (UC Merced)

Online Load Balancing [Graham ’66] Job: 8 1 2 3 4 5 6 7 Load: 1.1 1.3 1.2 0.8 2 1.2 0.9 1.2 (processing time) Machine 1 Machine 2 Online Vector Scheduling 2 Machine m

Online Load Balancing [Graham ’66] Job: 8 1 2 3 4 5 6 7 Load: 1.3 1.2 0.8 2 1.2 0.9 1.2 1.1 (processing time) Machine 1 Machine 2 Online Vector Scheduling 3 Machine m

Online Load Balancing [Graham ’66] Job: 8 1 2 3 4 5 6 7 Load: 0.8 2 1.2 0.9 1.2 1.1 + 1.2 = 2.3 (processing time) Machine 1 (load of a machine is the sum of its job loads) 1.3 Machine 2 Online Vector Scheduling 4 Machine m

Online Load Balancing [Graham ’66] Job: 8 1 2 3 4 5 6 7 Load: 0.8 2 1.2 0.9 1.2 Online problem: cannot see future jobs. 1.1 + 1.2 = 2.3 (processing time) Machine 1 (load of a machine is the sum of its job loads) 1.3 Machine 2 Online Vector Scheduling 5 Machine m

Online Load Balancing [Graham ’66] Job: 8 1 2 3 4 5 6 7 Load: Machine 1 Objective: minimize the makespan of the schedule (maximum load) Machine 2 Algorithm performance benchmark: Competitive ratio Online Vector Scheduling 6 Machine m

Online Load Balancing [Variants] Job: 8 1 2 3 4 5 6 7 Load: Machine 1 Objectives: minimize the p-norm of the machine loads (makespan is the ∞ -norm) Machine 2 [CW ’75, CC ’76, AAGKKV ’95, AAS ’01, C ’08, CFKKM ’11] Machine models: - Identical machines (load = p j ) [G ’66, FKT ’89, BKR ’94, BFKV ’95, KPT ’96, A ’99, FW ’00, GRTW ’00, R ’01, AAS ’01] - Related machines (load = p j / s i ) [AAFPW ’97, BCK ’00] - Unrelated machines (load = p ij ) [CW ’75, CC ’76, AAGKKV ’95, AAFPW ’97, C ’08, ANR ’95, CFKKM ’11] Online Vector Scheduling 7 Machine m

How do we load balance simultaneously on multiple resources (e.g., in data centers)? Online Vector Scheduling 8

Online Vector Scheduling 2 3 4 1 Jobs: (2, 1.5, 1) (1, 1.5, 1.3) (1, .8, .9) (2, 2.8, 1.3) Machine 1 Machine 2 Online Vector Scheduling 9 Machine m (processor) Dimension 2 (storage) Dimension 3 (network) Dimension 1

Online Vector Scheduling 2 3 4 1 Jobs: (2, 1.5, 1) (1, 1.5, 1.3) (1, .8, .9) 2 2.8 1.3 Machine 1 Machine 2 Online Vector Scheduling 10 Machine m (processor) Dimension 2 (storage) Dimension 3 (network) Dimension 1

Online Vector Scheduling 2 3 4 Jobs: (1, .8, .9) 2+1 = 3 1.3 + .9 = 2.2 2.8 + 1.5 = 4.3 (loads accumulate Machine 1 in each dimension) 1 2 1.5 Machine 2 Online Vector Scheduling 11 Machine m (processor) Dimension 2 (storage) Dimension 3 (network) Dimension 1

Online Vector Scheduling makespan: maximum over makespan in individual dimensions Jobs: Machine 1 Machine 2 Online Vector Scheduling 12 Machine m Dimension 2 Dimension 3 Dimension 1

Online Vector Scheduling p-norms: maximum over p-norms in individual dimensions Jobs: Machine 1 Machine 2 Online Vector Scheduling 13 Machine m Dimension 2 Dimension 3 Dimension 1

Summary of Results Makespan p-norm minimization minimization Identical machines O(log d) [Azar et al ’13, Our result: Meyerson et al ’14] Θ( (log d/log log d) 1-1/p ) Our result: Θ (log d/log log d) (Im-Kulkarni-Kell-P. Unrelated machines O(log d + log m) [Meyerson et al ’14] FOCS ’15) (machine dependent Our result: Our result: loads) Θ (log d + p) Θ (log d + log m) Related machines Later… Later… (Im-Kell-P.-Shadloo (non-uniform ’17) machine speeds) Online Vector Scheduling 14

Summary of Results Makespan p-norm minimization minimization Identical machines O(log d) [Azar et al ’13, Our result: Meyerson et al ’14] Θ( (log d/log log d) 1-1/p ) Our result: Θ (log d/log log d) (Im-Kulkarni-Kell-P. Unrelated machines O(log d + log m) [Meyerson et al ’14] FOCS ’15) (machine dependent Our result: Our result: loads) Θ (log d + p) Θ (log d + log m) Related machines Later… Later… (Im-Kell-P.-Shadloo (non-uniform ’17) machine speeds) Online Vector Scheduling 15

Identical machines algorithm: First attempt Greedy assignment (minimize maximum load across all machines and dimensions) unbalanced loads on dimensions …can be as bad as poly(d) -competitive Online Vector Scheduling 16

Identical machines algorithm: First attempt Greedy assignment Random Assignment (minimize maximum load (assignment uniformly at random) across all machines and dimensions) Chernoff bounds: unbalanced loads on dimensions O(log(dm))-competitive …can be as bad as poly(d) -competitive (optimal for unrelated machines) Online Vector Scheduling 17

Algorithm: Random and Greedy Assign uniformly at random Online Vector Scheduling 18

Algorithm: Random and Greedy Assign uniformly at random log d/ log log d log d/ log log d log d/ log log d Online Vector Scheduling 19

Algorithm: Random and Greedy Assign uniformly at random log d/ log log d log d/ log log d log d/ log log d Online Vector Scheduling 20

Algorithm: Random and Greedy Exceeds threshold Assign uniformly at random log d/ log log d log d/ log log d log d/ log log d Online Vector Scheduling 21

Algorithm: Random and Greedy Assign uniformly at random log d/ log log d log d/ log log d log d/ log log d Greedy schedule (minimize max over all machines and dimensions) Online Vector Scheduling 22

Algorithm: Random and Greedy Assign uniformly at random log d/ log log d log d/ log log d log d/ log log d E[greedy volume] < volume/poly(d) (Chernoff bounds) Greedy schedule E[greedy makespan] = O(1) (minimize max over all machines and dimensions) Online Vector Scheduling 23

Algorithm: Random and Greedy Competitive ratio: O(log d/log log d) Best we can do? Turns out yes: implies Ω (log d/log log d) lower bound Coloring lower bound For vector scheduling Online Vector Scheduling 24

Online Monochromatic Clique Clique: Given fixed of t colors: red, blue, and green. (here t = 3) Objective: minimize the largest monochromatic clique. 2 5 1 4 6 7 3 i th vertex arrives: online algorithm gets adjacencies with vertices 1, …, i -1 Online Vector Scheduling 25

Online Monochromatic Clique Clique: Given fixed of t colors: red, blue, and green. (here t = 3) Objective: minimize the largest monochromatic clique. Good Bad i th vertex arrives: online algorithm gets adjacencies with vertices 1, …, i -1 Online Vector Scheduling 32

The Game : Bins versus Colors (…or robots versus blue devils) Number of colors: t = 4 Online Algorithm Adversary (us) 3 My turn! Bin 1 Bin 2 1. Adversary defines adjacencies with prior vertices. 1 2. Algorithm places vertex in a bin (ALGO’s color). 3. Adversary colors the vertex (OPT’s decision) Bin 3 Bin 4 2 Bins = algorithm’s coloring Online Vector Scheduling 33

The Game : Bins versus Colors (…or robots versus blue devils) Number of colors: t = 4 Online Algorithm Adversary (us) 3 My turn! Bin 1 Bin 2 1. Adversary defines adjacencies with prior vertices. 1 2. Algorithm places vertex in a bin (ALGO’s color). 3. Adversary colors the vertex (OPT’s decision) Bin 3 Bin 4 2 Bins = algorithm’s coloring Online Vector Scheduling 34

Online Vector Scheduling Debmalya Panigrahi Duke University slides - PowerPoint PPT Presentation

Online Vector Scheduling Debmalya Panigrahi Duke University slides es Work done with: Maryam Shadloo Nat Kell Sungjin Im Janardhan Kulkarni (MSR UMN) (UC Merced) (Duke) (UC Merced) Online Load Balancing [Graham 66] Job: 8 1

Ponchatoula High School Scheduling for your Junior Year 2015-2016 Scheduling Procedures Online

Challenges of mixed-width vector code generation and static scheduling in LLVM (for VLIW

New Online Algorithms for Story Scheduling in Web Advertising Susanne Albers Achim Paen TU

CPU Scheduling CPU Scheduling CPU Scheduling 101 CPU Scheduling 101 The CPU scheduler makes a

CPU Scheduling CPU Scheduling CPU Scheduling 101 CPU Scheduling 101 The CPU scheduler makes a

The Sparse Vector Technique and online query answering

Online Scheduling of Parallel Programs on Hybrid Architectures

Module 6: CPU Scheduling Basic Concepts Scheduling Criteria Scheduling Algorithms

7 th Grade Counselor LYNN TOSI Online Scheduling at New Albany Middle School 7 th Grade

Scheduling Parallel DAG Jobs Online Ben Moseley (CMU) Joint work with: Kunal Agrawal (WahsU)

CJHS Scheduling incoming 9th graders TIMELINE January 3, 2019- Scheduling Presentation to

Chapter 6: CPU Scheduling Basic Concepts Scheduling Criteria Scheduling Algorithms

Streaming Algorithms for Bin Packing and Vector Scheduling Graham Cormode and Pavel Vesel y

CPU Scheduling Schedulers in the OS Structure of a CPU Scheduler Scheduling =

Module 5: CPU Scheduling Basic Concepts Scheduling Criteria Scheduling Algorithms

CPU Scheduling - I Different Scheduling Algorithms FCFS SJF Priority RR

Scheduling OS Lecture 11 UdS/TUKL WS 2015 MPI-SWS 1 Scheduling What is scheduling and

NOW Handout Page 1 1 Styles of Vector Architectures Components of Vector Processor Vector

CPU Scheduling Questions Why is scheduling needed? CSCI [4|6] 730 What is

CPU Scheduling Schedulers in the OS Structure of a CPU Scheduler Scheduling =

Module 5: CPU Scheduling Basic Concepts Scheduling Criteria Scheduling Algorithms

ScheduleSource ISU Dining Scheduling Program What is ScheduleSource? ScheduleSource is an

Vector addition: The zero vector The D -vector whose entries are all zero is the zero vector ,

1 CPU Scheduler Preemptive Scheduling Also referred to as Short Term Preemptive