Treating the Storage Stack Like a Network

摘要: In a data center, an IO from application to distributed storage traverses not only the network but also several software stages with diverse functionality. This set of ordered is known as or stack. Stages include caches, hypervisors, schedulers, file systems, and device drivers. Indeed, in typical number these often larger than hops destination. Yet, while packet routing fundamental networks, no notion exists on The path endpoint predetermined hard coded. forces different needs (e.g., requiring caching replica selection) flow through one-size-fits-all stack structure, resulting ossified article proposes sRoute, architecture that provides abstraction for sRoute comprises centralized control plane “sSwitches” plane. sets forwarding rules each sSwitch route requests at runtime based application-specific policies. A key strength our it works unmodified applications Virtual Machines (VMs). shows significant benefits customized center tenants: example, factor 10 tail latency, more 60% better throughput replication protocol, 2 caching, enabling live performance debugging running system.