RLSM: A Disaggregated LSM-Based Key-Value Store with Resource-Aware Compaction

Abstract

LSM-based storage engines are increasingly deployed in cloud storage systems and are moving toward disaggregated architectures. When compute and storage resources are separated, LSM compaction becomes a cross-node resource-control problem. Compute-side compaction can generate heavy RDMA network traffic when fetching remote SSTables, while near-data compaction may exhaust limited storage-side service capacity. Existing compaction schedulers primarily rely on CPU signals, but CPU availability is no longer the dominant bottleneck in disaggregated settings. We propose RLSM, a disaggregated LSM-based key-value store with critical-resource-aware compaction placement. RLSM jointly considers storage-side service conditions for near-data compaction admission and available RDMA bandwidth for CN-side compaction scheduling. Our evaluation shows that RLSM improves write throughput by 31%–53% over the strongest baseline while maintaining competitive throughput under read-heavy workloads.