CIDER: Boosting Memory-Disaggregated Key-Value Stores with Pessimistic Synchronization

A research team led by Yuxuan Du, Xuchuan Luo, Xin Wang, Yangfan Zhou, and Jiacheng Shen has developed CIDER, a novel framework that dramatically improves performance in memory-disaggregated key-value (KV) stores. These systems, which separate compute and memory resources across networks, have traditionally suffered from severe I/O bottlenecks due to redundant data synchronization operations. The researchers identified that existing optimistic synchronization approaches create massive network traffic when handling concurrent workloads, making the limited network bandwidth between compute and memory pools a critical performance constraint.

CIDER addresses this fundamental mismatch by introducing pessimistic synchronization combined with innovative optimization techniques. The framework employs a global write-combining method that significantly reduces cross-node redundant I/O operations by intelligently merging writes. Additionally, it features a contention-aware synchronization scheme that maintains performance even in low-contention scenarios where pessimistic approaches typically struggle. Experimental results demonstrate remarkable improvements, with CIDER boosting throughput of state-of-the-art memory-disaggregated KV stores by up to 6.6× under the YCSB benchmark.

The paper, accepted for publication at VLDB'26 (the premier database research conference), represents a significant advancement in distributed systems architecture. By fundamentally rethinking how synchronization occurs in disaggregated memory environments, CIDER enables more efficient utilization of network resources while maintaining data consistency. This breakthrough could have substantial implications for cloud database providers and large-scale web services that rely on high-performance key-value stores for real-time applications.