FORGE: Mitigating Synchronization Amplification for Memory-Disaggregated Caching Systems

Disaggregated Memory (DM) architectures offer caching systems the potential for elastic scaling and improved resource utilization by decoupling compute and memory. However, this advantage is undermined by costly cross-node synchronization, which exacerbates the overheads of critical cache operations, including hotness tracking, eviction coordination, and memory defragmentation. To address this challenge, we present FORGE, a caching system tailored for DM that prioritizes synchronization efficiency. FORGE groups cached objects based on similarity and performs group-level synchronizations to amortize overheads. It evicts cold groups via a contention-free and hotness-aware FIFO queue, efficiently sustaining high hit ratios while mitigating memory fragmentation. Leveraging the predictability of FIFO evictions, FORGE adopts a lazy synchronization strategy that updates hotness metrics just-in-time for eviction and offloads this process to on-chip memory in RDMA NICs for acceleration. Extensive evaluations on YCSB and real-world workloads demonstrate that FORGE achieves up to 4.5× higher throughput, 4.0×/7.5× lower P50/P99 latency, and an average of 1.14× higher cache hit ratio compared with state-of-the-art systems.