Bifrost: Alibaba's Next-Generation VPC Network with High-Performance Multipath Reliable Transport

Virtual Private Cloud (VPC) has become an increasingly important infrastructure-level service that provides isolated virtual networking environments for cloud tenants. In Alibaba Cloud, over 80% of tenant applications require reliable data transport, which is currently implemented in VPC networks through guest-side TCP. However, TCP often suffers significant performance degradation when facing network instabilities in production clouds, especially for tail-latency sensitive applications such as Redis and Nginx. In this paper, we present BIFROST, the next-generation VPC network in Alibaba Cloud that provides high-performance multipath reliable transport. BIFROST employs RTT-aware multipath packet spraying to bypass failures and mitigate elephant flows, ensures in-order delivery via in-place guest reordering, achieves end-to-end reliability with delayed bitmap ACKs, and performs efficient reliability state management to support large-scale deployment. Extensive evaluation shows that BIFROST reduces tail latency by up to 307× for Redis and 66× for Nginx, achieves millisecond-level failure recovery, and supports O(100k) concurrent connections per SmartNIC.