T-RACKs: A Faster Recovery Mechanism for TCP in Data Center Networks

Cloud interactive data-driven applications generate swarms of small TCP flows that compete for the switch buffer space in data-center. Such require a flow completion time (FCT) to be effective. Unfortunately, is myopic with respect composite nature application data. In addition it tends artificially inflate FCT individual by several orders magnitude, because its Internet-centric design, fixes retransmission timeout (RTO) at least hundreds milliseconds. To better understand this problem, paper, we use empirical measurements data center testbed study, microscopic level, effects various types packet losses on TCP's performance. particular, single out impact tail end flows, as well bursty span significant fraction congestion windows, and show non-negligible effect such FCT. Based this, propose so-called, timely-retransmitted ACKs (or T-RACKs), simple loss recovery mechanism conceals drawbacks long RTO even presence heavy losses. Interestingly enough, T-RACKS achieves transparently itself does not any change tenant's virtual machine (VM) or container. T-RACKs can implemented software shim layer hypervisor between VMs server's NIC hardware networking function SmartNIC. Simulation real results remarkable performance improvements.