Exploring remote memory for buffer cache extension: a preliminary investigation

Bygari, Ramya

Permalink

https://hdl.handle.net/2142/124717 Copy

Description

Title

Exploring remote memory for buffer cache extension: a preliminary investigation

Author(s)

Bygari, Ramya

Issue Date

2024-04-30

Director of Research (if dissertation) or Advisor (if thesis)

• Ganesan, Aishwarya
• Alagappan, Ramnatthan

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Distributed Systems
• Storage Systems
• buffer cache
• remote memory
• disaggregated memory
• RDMA

Abstract

Data-intensive and memory-intensive applications demand high performance from buffer caches to efficiently manage large datasets and minimize access latencies. Traditional buffer caches, confined within a single machine’s expensive DRAM, encounter limitations in capacity and scalability, leading to performance bottlenecks as a result of costly disk fetches. Conversely, datacenters often face underutilization challenges. By circumventing traditional network protocol overheads, RDMA offers low-latency, high-throughput access to remote memory pools, thereby minimizing performance penalties. This study investigates the integration of RDMA (Remote Direct Memory Access) with buffer caches to overcome these challenges. RDMA enables memory disaggregation, allowing buffer cache capacities to extend beyond individual machine constraints by leveraging memory resources across multiple servers. This study assesses the efficacy of RDMA-enabled disaggregated buffer caches in enhancing application operations by expanding distributed capacity, leading to improved performance characterized by reduced latency for file accesses and enhanced throughput. This thesis serves as a preliminary study and demonstrates seamless integration with existing applications, eliminating the need for modifications to the application stack. Performance evaluation using various data-intensive POSIX application-level databases, such as LevelDB, SQLite, RocksDB, and WiredTiger, illustrates significant reductions in access latencies, enhancements in throughput, and improvements in cache hit rates. These findings affirm the viability of RDMA-based buffer caching for read-intensive operations in data center environments. Furthermore, this research paves the way for exploring support for complex read-write operations and expanding the scope of memory disaggregation applications, promising widespread adoption and adaptation in various computing environments requiring optimal performance from buffer caches.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Ramya Bygari

Owning Collections