Many applications in public clouds require a high fan-in, many-to-one type of data communication (known as TCP incast) in modern Data Center Networks (DCNs). Such communication could cause severe incast congestion in switches and result in TCP throughput collapse, substantially degrading the application performance. The root cause of throughput collapse is the Retransmission Timeouts (RTO) due to packet losses in congested switches. Tenants in public clouds can opt to use a variety of TCP versions. However, the existing solutions rely on modifications of TCP protocols and specific techniques from switches, and thus these existing solutions are not always feasible for public clouds. In this paper, we are inspired by the emerging virtualization and network softwarization technologies to develop a novel scheme called Retransmission timeout Avoidance by Packet Intelligent Discarding (RAPID) using software switches. RAPID considers the number of packets of each incast flow, buffered in the switch to selectively discard some packets, and ensures that the Fast Retransmission/Fast Recovery rather than RTO is invoked at the sender(s) in response to packet loss. Thus, the long idle period of a timeout and the throughput drop are avoided. We prove that, given a predetermined minimum switch buffer space, dedicated to the incast application, RAPID can prevent RTO in all the incast senders. We also present a low-complexity heuristic version of RAPID named RAPID-ED, which combines the principles of RAPID and early detection and is extremely easy to implement on today's software switches. We evaluate the two proposed schemes in a data center network testbed built on NS-3 simulator. The simulation results confirm the theoretical expectation, and show that the RAPID and RAPID-ED perform very well to prevent RTO of TCP incast flows and hence the throughput collapse. Compared with other incast solutions, RAPID and RAPID-ED do not modify TCP protocols and therefore are more suitable in public clouds.

Many applications (e.g., cluster based storage and MapReduce) in modern data centers require a high fan-in, many-to-one type of data communication (known as TCP incast), which could cause severe incast congestion in switches and result in TCP goodput collapse, substantially degrading the application performance. The root cause of such a collapse is the long idle period of the Retransmission Timeout (RTO) that is triggered at one or more senders by packet losses in congested switches. In this paper we develop a packet labelling scheme PLATO, which improves the loss detection capabilities of NewReno using an innovative packet labelling system. Packets carrying this special label are preferentially enqueued, at the switch. This allows TCP to detect packet loss using three duplicate acknowledgements, instead of the time expensive RTO; thus avoiding the goodput collapse. PLATO makes minor modifications to NewReno and does not alter its congestion control mechanism. The implementation and simulations have been done in Network Simulator 3 (NS3). PLATO's performance is significantly better than NewReno as well as state-of-art incast solutions Incast Control TCP (ICTCP) and Data Center TCP (DCTCP). We also show that TCP PLATO can be implemented using commodity switches with Weighted Random Early Detection (WRED) function.

Incast applications have grown in popularity with the advancement of data center technology. It is found that the TCP incast may suffer from the throughput collapse problem, as a consequence of TCP retransmission timeouts when the bottleneck buffer is overwhelmed and causes the packet losses. This is critical to the Quality of Service of cloud computing applications. While some previous literature has proposed solutions, we still see the problem not completely solved. In this paper, we investigate the three root causes for the poor performance of TCP incast flows and propose three solutions, one for each at the beginning, the middle and the end of a TCP connection. The three solutions are: admission control to TCP flows so that the flow population would not exceed the network's capacity; retransmission based on timestamp to detect loss of retransmitted packets; and reiterated FIN packets to keep the TCP connection active until the the termination of a session is acknowledged. The orchestration of these solutions prevents the throughput collapse. The main idea of these solutions is to ensure all the on-going TCP incast flows can maintain the self-clocking, thus eliminates the need to resort to retransmission timeout for recovery. We evaluate these solutions and find them work well in preventing the retransmission timeout of TCP incast flows, hence also preventing the throughput collapse.

The recent literature on data center networks often propose the use of centralized control server to manage resources or for coordination. These controllers are always a single omniscient device. While we see no problems in practice yet as it works for small scale networks, there is a scalability concern. This paper proposes the idea of devolved controllers, namely, a number of controllers that each with only limited information about the network, but they together can replace a omniscient controller. Thus it permits the controllers' workload to scale up. We use dynamic flow reroute as an example, to study how we can build a network with devolved controllers, how to configure them, how they operate, and especially how they provide resilience. We show that these devolved controllers can remove the scalability concern and provide redundancy to each other; externally, they are as easy to use as a single omniscient controller. We describe how these controllers are prepared and used to prove that devolved controllers is feasible idea, and provide a precursor to a new direction on the control aspect of networking.

We propose a fast reroute scheme for the single link failure protection of IP multicast network. A multicast network connects nodes as a spanning tree, which is vulnerable to link failures. Our scheme uses ring topologies to ensure a backup path always exists to any single link failure and only a few nodes in the neighborhood of the failed link have to react to the failure. Compare to other solutions, we have four advantages: (1) The multicast tree can be arbitrary. (2) There is only a minimal disruption and there would be no packet loss. (3) Failure protection is available to a wider set of network topologies. (4) Only a small number of multicast nodes are involved and the number of spare links used are also small.

Admission control has been shown to be a preferred alternative to TCP-friendly congestion control for inelastic flows in heterogeneous networks shared by elastic and inelastic traffic [1]. However, it is possible for an inelastic flow to adopt different level of aggressiveness in implementing the admission control. How these different levels of aggressiveness affect the system performance remains an open issue. In this paper, we evaluate a full spectrum of (abstract) admission control algorithms in terms of their aggressiveness towards elastic flows. A totally aggressive version would admit an inelastic flow even if this means elastic flows' fair bandwidth share is reduced to close to zero. In the other extreme, a TCP-friendly version would only admit an inelastic flow if its desired rate is no higher than what the elastic flows will receive after its arrival. We show that the performance of inelastic flows is asymptotically insensitive to their aggressiveness without strong assumptions about flow file size or holding time distributions. This makes a strong case for adopting a less aggressive, yet TCP-friendly admission control in a heterogeneous network. Extensive simulations are carried out to validate the performance, stability and asymptotic behavior the the proposed TCP-friendly admission control policy.