XCo: Explicit Coordination for Preventing Congestion in Data Center Ethernet

Cluster-based storage systems increasingly use commodity communication technologies, such as Fibre Channel over Ethernet (FCoE), for accessing stored data over the network. Data is striped over multiple storage nodes, and storage traffic often shares the network with non-storage traffic. In such conditions, storage clients can experience severely degraded performance, such as TCP throughput collapse and network congestion due to competing network traffic. Furthermore, consolidation of multiple virtual machines (VMs) onto fewer physical nodes can worsen the performance of network storage systems. The root cause of this performance problem is that network traffic from multiple sources can cause transient overloads in the switch buffers. In this paper, we make the case that virtualization opens up a new set of opportunities to alleviate and solve such performance problems experienced by network storage in particular, and data center Ethernet in general. We present an architecture, called XCo, for explicit coordination of network traffic among VMs in a data center Ethernet that is inexpensive, fully transparent, currently feasible, and complementary to any switch-level hardware support. We present experimental evidence via proof-of-concept implementation and evaluation to support this claim and describe the challenges and opportunities in a complete solution.