KDB: A Scalable Persistent Key-Value Store with Atomic Batches and Snapshots

SOS: A High-Performance Distributed Key-Value Store for Large-Scale Online Services

Cutting to the Chase: NVMe-oF Fast-track for Replicated Persistent Key-value Stores

Adopting replicated persistent key-value stores (RPKVSs) as metadata backends is becoming increasingly popular in modern large-scale storage systems. Conventional RPC-based KV replication is infamously subject to its replication overheads due to the inherent architectural mismatch between the application-level network stack and storage-prioritized operations. In this paper, we shift toward a new paradigm of storage-native block-based replication to fundamentally address the limitations of conventional replication. This shift is made possible by using widely deployed NVMe-oF technology to unify replication and storage efficiency under integrated block semantics. We present NoFDB, an NVMe-oF-enabled RPKVS design, that eliminates redundant index management and reduces replication latency, and introduces a set of system-level optimizations to address the challenges introduced by NVMe-oF and unlock its performance potential. Our evaluation shows that NoFDB improves random write throughput by up to 2.85 × and macrobenchmark throughput by up to 3.10 × over the best-performing baseline in each setting.

Poster: AutoThermKV: An Efficient User-Transparent In-Memory Management of Hot Data for Key-Value Stores

Log-structured merge-tree (LSM-tree) key-value stores rely on caching to mitigate multi-component lookups and long tails, yet block and KP caches are prone to compactioninduced expiry, and all three cache types suffer from scan pollution under LRU eviction. We present AutoThermKV, a two-level in-memory architecture that decouples fast admission from hot-item protection. The front tier (HotFilter) uses a 64 Baligned bitmap/key/value page layout with Quick Space Reservation for near-constant admission cost; the back tier (HotCommitted) retains proven hot items via a lightweight LRU and a secondtouch promotion rule. Implemented atop RocksDB and evaluated on YCSB-like workloads with controllable temporal locality, AutoThermKV achieves up to 2.58 × throughput speedup, 2.60 × average-latency reduction, and tail-latency improvements of up to 20 × (p99) and 36 (p999) under modest memory budgets.

MiTDM: Eliminating False Conflicts in Scalable MVCC in Disaggregated Memory

Disaggregated memory architecture has gained wide adoption in cloud and high-performance systems [18, 31, 43] due to its decoupled resource model, elasticity, and low-latency access. In such architectures, transaction mechanisms must ensure atomic and consistent access to remote memory. Prior designs use array-based version structures to improve locality, but their fixed capacity introduces false conflicts, especially for long transactions that require accessing historical versions, limiting concurrency and throughput. We present MiTDM, a transaction system tailored for disaggregated memory environments. MiTDM introduces a hierarchical block version chain that combines the benefits of array and chained structures, supporting dynamic version expansion while maintaining low-latency access. A lightweight Block Index Table accelerates remote reads, and an asynchronous index update mechanism synchronizes indexes across nodes with minimal overhead. To manage version growth, MiTDM employs a hybrid garbage collection strategy aware of transaction states, enabling safe reclamation and reuse of version slots, particularly in hot spots. A one-sided optimistic transaction protocol ensures consistency and atomicity under high concurrency. Experimental results on a disaggregated memory cluster show that MiTDM outperforms state-of-the-art systems. Under the TPC-C workload with 128 clients, it achieves up to 80.1% higher throughput compared to FaRMV2 and 42.7% higher than Motor, while significantly reducing P50 and P99 latencies.

S !"#$ : A Scalable and Resize-optimized Hash Index on Disaggregated Memory

MEMONDEMAND: A Memory Management System for Large-Scale Enterprise Data

On EnterpriseRAG-Bench, MEMONDEMAND outperforms the strongest published LB#1 result at every evaluated scale from 10M tokens through the complete 618M- token collection, and results on FinanceBench, HotpotQA, and FRAMES further show strong performance across financial, multi-hop, and fact-retrieval settings.