Varuna: Enabling Failure-Type Aware RDMA Failover

A research team led by Xiaoyang Wang has introduced Varuna, a novel RDMA (Remote Direct Memory Access) failover system that fundamentally changes how data centers handle network link failures. Current production systems use a brute-force approach: when a primary RDMA link fails, they retransmit all in-flight requests over a standby backup path. Varuna's key insight is that this "blanket retransmission" is both inefficient and potentially incorrect. It unnecessarily consumes bandwidth by resending requests the responder has already executed (post-failure requests), and for non-idempotent operations—where repeating an action changes the result—duplicate execution can violate application semantics and break transactional consistency.

Varuna solves this by piggybacking a lightweight completion log on every RDMA operation. When a link failure occurs, this log allows the system to deterministically identify which in-flight requests were truly lost (pre-failure) and which were successfully executed by the responder (post-failure). The system then performs a surgical recovery: it retransmits only the pre-failure subset and fetches the return values for the already-completed post-failure requests. Evaluated using synthetic microbenchmarks and end-to-end RDMA TPC-C transactions—a standard database benchmark—Varuna demonstrated impressive results. It incurs a minimal 0.6-10% steady-state latency overhead in realistic applications, slashes recovery retransmission time by 65%, guarantees transactional consistency, and adds zero connectivity rebuild overhead with negligible extra memory use during failover.