FineMem - Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained Disaggregated Memory Management

Learn about FineMem, an innovative RDMA-connected remote memory management system presented at OSDI '25 that solves the critical trade-off between allocation overhead and memory waste in disaggregated memory architectures. Discover how researchers from the University of Science and Technology of China and Google developed a solution to enable high-performance, fine-grained memory allocation in RDMA-enabled memory disaggregation systems. Explore the technical challenges of remote memory allocation and deallocation in modern data centers, where existing systems rely on coarse-grained allocations measured in gigabytes, resulting in significant memory waste. Understand FineMem's innovative approach to removing RDMA memory region registration costs from allocation paths through per-compute node MR pre-registration, while maintaining remote memory isolation using RDMA memory windows and trusted allocation services. Examine the lock-free, one-sided RDMA-based protocol that enables allocation of memory chunks as small as 4KB and 2MB without involving the memory node's CPU, and learn how the system maintains metadata consistency during compute node failures through logging mechanisms. Analyze the impressive performance improvements, including up to 95% reduction in remote memory allocation latency compared to state-of-the-art systems, and see how FineMem enables memory malloc systems, key-value stores, and swap systems to achieve low memory waste with minimal overhead in disaggregated memory environments.