Architecture Support for Fine-grained Parallelism
Event details
Date | 26.06.2024 |
Hour | 13:00 › 15:00 |
Speaker | Jiawei Lin |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Paolo Ienne
Thesis advisor: Prof. Thomas Bourgeat
Co-examiner: Prof. Caglar Gulcehre
Abstract
The rapid evolution of computing demands has driven the need for more efficient and versatile processing frameworks. Modern applications such as machine learning, gaming, and real-time data processing can be accelerated by GPUs and domain-specific accelerators. However, irregular applications, such as Graph Neural Networks (GNNs), remain challenging to accelerate due to workload imbalances and the need for fine-grained parallelism.
This paper explores enhancements to the Simultaneous and Heterogeneous Multi-threading (SHMT) framework to address these challenges. We propose integrating dynamic parallelism capabilities and an efficient on-chip synchronization mechanism. Our enhanced SHMT framework aims to improve resource utilization and overall system efficiency, offering better support for fine-grained and dynamic parallel computations.
Background papers
Exam president: Prof. Paolo Ienne
Thesis advisor: Prof. Thomas Bourgeat
Co-examiner: Prof. Caglar Gulcehre
Abstract
The rapid evolution of computing demands has driven the need for more efficient and versatile processing frameworks. Modern applications such as machine learning, gaming, and real-time data processing can be accelerated by GPUs and domain-specific accelerators. However, irregular applications, such as Graph Neural Networks (GNNs), remain challenging to accelerate due to workload imbalances and the need for fine-grained parallelism.
This paper explores enhancements to the Simultaneous and Heterogeneous Multi-threading (SHMT) framework to address these challenges. We propose integrating dynamic parallelism capabilities and an efficient on-chip synchronization mechanism. Our enhanced SHMT framework aims to improve resource utilization and overall system efficiency, offering better support for fine-grained and dynamic parallel computations.
Background papers
- Can Dataflow Subsume Von Neumann Computing? https://ieeexplore.ieee.org/document/714561
- Simultaneous and Heterogenous Multithreading https://dl.acm.org/doi/10.1145/3613424.3614285
- MaxK-GNN: Extremely Fast GPU Kernel Design for Accelerating Graph Neural Networks Training https://dl.acm.org/doi/10.1145/3620665.3640426
Practical information
- General public
- Free