Activeswap: Memory-Rich-Node-Initiated Load Balancing for RDMA-Based Distributed Memory

Remote direct memory access (RDMA)-based distributed memory systems improve cluster-wide memory utilization by allowing for memory-constrained nodes to access unused memory on remote nodes. However, most existing systems adopt a memory-constrained-node-driven architecture, in which resource-constrained nodes must perform cluster-state synchronization, target selection, memory chunk mapping, and metadata management. Such control-path operations further increase CPU pressure on memory-constrained nodes. This paper presents Activeswap, a load-balanced distributed memory system based on memory-rich-node initiative. Memory-rich nodes actively collect cluster-state information through one-sided RDMA operations and initiate memory chunk mappings according to a global-state-aware selection policy, thereby reducing management overhead on memory-constrained nodes and improving the distribution of remote memory chunks. We implement Activeswap in an RDMA cluster and evaluate it using matrix computation, Memcached, and FINCH workloads. Experimental results show that the memory-rich-node-initiated design reduces workload completion time by 6.5% on average and host CPU consumption by up to 59.34% while producing a more balanced distribution of remote memory chunks. These results demonstrate that shifting remote memory control-path operations to memory-rich nodes can improve system efficiency without changing the RDMA data path.