Leveraging Remote Direct Memory Access (RDMA) for Designing Efficient and Fast Distributed Systems
Event details
| Date | 17.06.2024 |
| Hour | 09:00 › 11:00 |
| Speaker | Beatrice Shokry |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Edouard Bugnion
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Thomas Bourgeat
Abstract
Building efficient and fast distributed systems is inherently related to overcoming communication overhead. Recently, Remote Direct Memory Access (RDMA) technology has presented an opportunity for distributed systems to accomplish this goal. In particular, RDMA allows remote memory access operations between physically separated nodes without involving the Central Processing Units (CPUs) on both sides, thereby minimizing latency and enhancing the overall network performance. By enabling efficient data transfers between remote nodes, RDMA plays a pivotal role in high-performance computing, cloud computing, and data center environments. Nevertheless, there are several challenges associated with RDMA adoption, including the impact of RDMA's complex hardware architecture on the design choices and accordingly the overall system's performance. Such design choices have not been thoroughly investigated in literature. In this proposal, we present three works from literature that have addressed the issue of building RDMA-aware systems aiming to maximize the usage of existing RDMA capabilities for achieving higher performance. Finally, our proposed research is discussed.
Background papers
1- Design Guidelines for High Performance RDMA Systems
2- Using One-Sided RDMA Reads to Build a Fast, CPU-Efficient Key-Value Store
3- RDMA is Turing complete, we just did not know it yet!
Exam president: Prof. Edouard Bugnion
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Thomas Bourgeat
Abstract
Building efficient and fast distributed systems is inherently related to overcoming communication overhead. Recently, Remote Direct Memory Access (RDMA) technology has presented an opportunity for distributed systems to accomplish this goal. In particular, RDMA allows remote memory access operations between physically separated nodes without involving the Central Processing Units (CPUs) on both sides, thereby minimizing latency and enhancing the overall network performance. By enabling efficient data transfers between remote nodes, RDMA plays a pivotal role in high-performance computing, cloud computing, and data center environments. Nevertheless, there are several challenges associated with RDMA adoption, including the impact of RDMA's complex hardware architecture on the design choices and accordingly the overall system's performance. Such design choices have not been thoroughly investigated in literature. In this proposal, we present three works from literature that have addressed the issue of building RDMA-aware systems aiming to maximize the usage of existing RDMA capabilities for achieving higher performance. Finally, our proposed research is discussed.
Background papers
1- Design Guidelines for High Performance RDMA Systems
2- Using One-Sided RDMA Reads to Build a Fast, CPU-Efficient Key-Value Store
3- RDMA is Turing complete, we just did not know it yet!
Practical information
- General public
- Free