Distributed Systems on Modern Data-Center Architectures
Event details
| Date | 23.05.2023 |
| Hour | 09:00 › 11:00 |
| Speaker | Clément Burgelin |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Paolo Ienne
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Edouard Bugnion
Abstract
Modern data centers offer fast network fabric and specialized hardware to make efficient use of this network. Such technologies enable the emergence of new data center architectures. Disaggregated Memory (DM) is one such architecture that allows for better resource utilization and scalability of cloud services such as Key-Value Stores (KVSes). Yet, designing services for DM is not a trivial task, and existing DM-based in-memory KVSes suffer from high latency overhead and low throughput compared to traditional ones. We study recent contributions made in the field of DM-based services and in-memory KVSes, and how new designs could further improve upon existing approaches.
Background papers
1- Design Guidelines for High Performance RDMA Systems (USENIX ATC’16): https://www.usenix.org/conference/atc16/technical-sessions/presentation/kalia
2- RACE: One-sided RDMA-conscious Extendible Hashing (ACM Transactions on Storage '22): https://dl.acm.org/doi/full/10.1145/3511895
3- FUSEE: A Fully Memory-Disaggregated Key-Value Store (FAST '23) : https://www.usenix.org/conference/fast23/presentation/shen
Exam president: Prof. Paolo Ienne
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Edouard Bugnion
Abstract
Modern data centers offer fast network fabric and specialized hardware to make efficient use of this network. Such technologies enable the emergence of new data center architectures. Disaggregated Memory (DM) is one such architecture that allows for better resource utilization and scalability of cloud services such as Key-Value Stores (KVSes). Yet, designing services for DM is not a trivial task, and existing DM-based in-memory KVSes suffer from high latency overhead and low throughput compared to traditional ones. We study recent contributions made in the field of DM-based services and in-memory KVSes, and how new designs could further improve upon existing approaches.
Background papers
1- Design Guidelines for High Performance RDMA Systems (USENIX ATC’16): https://www.usenix.org/conference/atc16/technical-sessions/presentation/kalia
2- RACE: One-sided RDMA-conscious Extendible Hashing (ACM Transactions on Storage '22): https://dl.acm.org/doi/full/10.1145/3511895
3- FUSEE: A Fully Memory-Disaggregated Key-Value Store (FAST '23) : https://www.usenix.org/conference/fast23/presentation/shen
Practical information
- General public
- Free