.png){kind=icon}
.png){kind=icon}
Return
Improving SQL Join Algorithms for Distributed Systems: A Case Study of Compute Express Link-Based Multihost Shared Memory
.png){kind=icon}
PRE
AI
DOI：10.1109/MM.2025.3574357.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The advent of compute express link (CXL) has introduced the possibility of multihost shared memory architectures. Despite this advancement, there has been limited exploration of shared memory at the application layer. Traditional distributed systems typically partition data across multiple servers, enabling independent processing. However, cross-partition operations, such as joins, require data repartitioning, leading to significant communication overhead. To address this challenge, we propose merge hash join (MHJ), a novel structured query language join algorithm that leverages shared memory to eliminate the need for repartitioning. By storing the joining table in shared memory and making it directly accessible to all servers, MHJ significantly reduces communication overhead. To validate our approach, we implemented MHJ and the necessary shared memory functionalities on a CXL-based shared memory prototype. Extensive evaluations using the industry-standard TPC-DS benchmark demonstrate that MHJ achieves up to 1.5× performance improvement compared to conventional join algorithms.
Keywords:
Sparks
Partitioning algorithms
Distributed databases
Servers
Hash functions
Clustering algorithms
Buildings
Training
Structured Query Language
Random access memory