Cornell University
Scaling Distributed Systems with RDMA in Post-Moore's Law Datacenters
Abstract
dc:description.abstractIn recent years, datacenter architectures have been significantly reshaped by the convergence of high-speed networks and the limitations of Moore's Law. This evolution necessitates new strategies to enhance the performance and scalability of distributed systems. Traditional CPU-centric optimizations are increasingly inadequate, leading to the exploration of hardware offload solutions such as RDMA (Remote Direct Memory Access) and SmartNICs. This thesis investigates the integration of RDMA-based network offloads and addresses the challenges of adapting distributed protocols to leverage these technologies effectively. We introduce PRISM, a novel RDMA interface designed for efficient offloading of distributed systems, and propose SLIM (Shared Log Interface to Memory), a simplified log API that bridges the abstraction gap between distributed programming models and specialized hardware accelerators. SLIM offers a higher-level instantiation of the traditional shared memory model, aiming to enhance programming outcomes by providing a more intuitive and effective approach to reasoning about concurrency and consistency in complex RDMA-based offloads.
Degree
thesis:*- Name thesis:degree_name
- Ph. D., Computer Science
- Level thesis:degree_level
- Doctor of Philosophy
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Cornell University
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Dharanipragada, Sowmya
- Chair dc:contributor.chair
-
- Alvisi, Lorenzo
- Committee members dc:contributor.committeemember
-
- McCrea, Lawrence
- Ristenpart, Thomas
Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Bibid
- 16612008
- Dc Identifier Other
-
ProQuest Submission ID: 14599
ProQuest Publication ID: 31489085