ROLEX - A Scalable RDMA-oriented Learned Key-Value Store for Disaggregated Memory Systems

Explore a conference talk on ROLEX, a scalable RDMA-oriented learned key-value store designed for disaggregated memory systems. Delve into the innovative approach that separates monolithic servers into compute and memory nodes, leveraging RDMA for direct remote memory access. Learn how ROLEX addresses challenges in existing ordered key-value stores by implementing a retraining-decoupled learned index scheme, enabling efficient data modifications and asynchronous model retraining. Discover the performance improvements achieved by ROLEX, particularly in dynamic workloads, and understand its potential impact on high-performance computing and data management in disaggregated environments.

Intro
Disaggregated Memory Systems (DMS)
Computing Bottleneck
Overloaded Bandwidth
Inconsistency Issues
Memory Pool Stores Data
Learned Model - Piecewise Linear Regression Models
Learned Model - Leaf Table
Asynchronous Retraining - Consistency Guarantee
Experimental Setup
Performance on YCSB
Scale with CPU cores on compute nodes
Training Latency
Memory Overhead