Rethinking Congestion Management in Intra-Host Networks

Conventional wisdom believes that the intra-host network provides ample and stable band-width. However, to meet the growing demand of datacenter applications, modern commodity servers are equipped with numerous heterogeneous hardware accelerators. These devices consume substantial intra-host network resources (e.g., memory bus bandwidth, cache sizes, etc.), while technology trends in these areas have stagnated. As a result, congestion emerges in the intra-host network and leads to performance jitters. In this article, we revisit existing works in intra-host congestion management and find that these works typically focus on only partial root causes of the congestion and lack a holistic view of the intra-host network. Then, we propose host congestion management (HCM), a holistic congestion management architecture for intra-host networks. HCM consists of three components: a holistic intra-host status collector, a congestion risk level monitor, and a resource allocation manager. HCM directly utilizes hardware/software counters or implements state-of-the-art architectural innovations to realize a fine-grained system while ensuring high performance. We implement HCM on commodity servers and conduct multiple testbed experiments. The results show that HCM can consistently maintain high performance even with a high degree of intra-host congestion. Finally, we analyze and outline the future of congestion management in intra-host networks.