A Fast and Elastic Distributed Matrix Computation Engine using GPUs
- Title
- A Fast and Elastic Distributed Matrix Computation Engine using GPUs
- Alternative Title
- 그래픽 처리 장치를 사용하는 빠르고 탄력적인 분산 행렬 계산 엔진
- Author(s)
- Donghyoung Han
- DGIST Authors
- Donghyoung Han ; Daehoon Kim ; Kyongseok Park
- Advisor
- 김대훈
- Co-Advisor(s)
- Kyongseok Park
- Issued Date
- 2021
- Awarded Date
- 2021/02
- Type
- Thesis
- Table Of Contents
-
Chapter 1. Introduction 1
1.1 Motivation and objectives 1
1.2 Main contributions 8
1.3 Structure of thesis 9
Chapter 2. Background 11
2.1 Matrix Computation Engine 11
2.1.1 Matrix Partitioning Schemes 11
2.1.2 Distributed Matrix Multiplication Methods 12
2.1.3 Operator Fusion 17
2.1.4 Distributed Fused Operator 23
Chapter 3. DistME: Elastic Distributed Matrix Multiplication 28
3.1 Cuboid Matrix Multiplication 28
3.1.1 (P; Q; R)-Cuboid partitioning 28
3.1.2 Optimization of CuboidMM 31
3.2 Acceleration of CuboidMM using GPUs 33
3.2.1 Subcuboid partitioning 33
3.2.2 Optimization of subcuboids for GPU 35
3.2.3 GPU streaming of subcuboids 36
3.2.4 Algorithm 38
3.3 Experimental Evaluation 38
3.3.1 Experimental setup 40
3.3.2 Performance of CuboidMM 41
3.3.3 Performance of DistME 44
3.3.4 Performance of GNMF 47
3.3.5 Comparison with Systems in HPC 48
Chapter 4. FuseME: Operation Fusion Method 51
4.1 Cuboid-based fused Operator 51
4.1.1 Cuboid-based Fusion 51
4.1.2 Cost Optimization 59
4.2 Cuboid-based Fusion Generation 66
4.2.1 Fusion Plan 66
4.2.2 Union and Cut of Fusion plan 71
4.3 Experimental Evaluation 73
4.3.1 Experimental setup 74
4.3.2 Performance of Cuboid-based fusion 75
4.3.3 Optimization of Cuboid-based fusion 78
4.3.4 Performance Depending on Fusion plan . 80
Chapter 5. Related Work 82
5.1 Related Work 82
Chapter 6. Conclusion 84
6.1 Conclusions 84
Bibliography 86
- URI
-
http://dgist.dcollection.net/common/orgView/200000364466
http://hdl.handle.net/20.500.11750/16698
- Degree
- Doctor
- Department
- Information and Communication Engineering
- Publisher
- DGIST
- Related Researcher
-
-
Kim, Daehoon
- Research Interests Computer Architecture and Systems; Virtualization; Cloud Computing
-
- Files in This Item:
-
There are no files associated with this item.
- Appears in Collections:
- Department of Electrical Engineering and Computer Science Theses Ph.D.
Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.