SF-STACK: Streamlining RDMA for Heterogeneous Telecom Storage

Remote Direct Memory Access (RDMA) significantly enhances performance in distributed storage systems. However, deploying RDMA in heterogeneous environments remains a major challenge. Existing solutions from hyperscalers rely on full control over network configurations to achieve uniform deployments—conditions rarely met in telecom settings. Without infrastructure modifications, real-world deployments suffer from excessive connection overhead on resource-constrained servers, inconsistent path selection due to varying Linux bonding implementations, and degraded performance caused by congestion control sensitivity to diverse switch configurations. We present SF-STACK, a portable RDMA framework designed for fragmented storage clusters characterized by het-erogeneity and limited administrative control. SF-STACK introduces: (1) a thread-shared eXtended Reliable Connection (XRC) Queue Pair (QP) mechanism to reduce memory consumption; (2) a dual-IP routing strategy to ensure path consistency; and (3) a lightweight RTT-based congestion control algorithm that operates independently at the end hosts. Deployed across over 20 production clusters with diverse network architectures at a major telecom provider, SF-STACK achieves 70% memory consumption reduction and 50% bandwidth utilization improvement, requiring no hardware or network modifications.