Cloud data centers host diverse applications, mixing workloads that require small predictable latency with others requiring large sustained throughput. In this environment, today's state-of-the-art TCP protocol falls short. We present measurements of a 6000 server production cluster and reveal impairments that lead to high application latencies, rooted in TCP's demands on the limited buffer space available in data center switches. For example, bandwidth hungry "background" flows build up queues at the switches, and thus impact the performance of latency sensitive "foreground" traffic.

Abstract—Many applications require fast data transfer over high speed and long distance networks. However, standard TCP fails to fully utilize the network capacity due to the limitation in its conservative congestion control (CC) algorithm. Some works have been proposed to improve the connection’s throughput by adopting more aggressive loss-based CC algorithms. These algorithms, although can effectively improve the link utilization, have the weakness of poor RTT fairness. Further, they may severely decrease the performance of regular TCP flows that traverse the same network path. On the other hand, pure delay-based approaches that improve the throughput in high-speed networks may not work well when the traffic is mixed with both delaybased and greedy loss-based flows. In this paper, we propose a novel Compound TCP (CTCP) approach, which is a synergy of delay-based and loss-based approach. Specifically, we add a scalable delay-based component into the standard TCP Reno congestion avoidance algorithm (a.k.a., the loss-based component). The sending rate of CTCP is controlled by both components. This new delay-based component can rapidly increase sending rate when network path is under utilized, but gracefully retreat in a busy network when bottleneck queue is built. Augmented with this delay-based component, CTCP provides very good bandwidth scalability with improved RTT fairness, and at the same time achieves good TCP-fairness, irrelevant to the windows size. We developed an analytical model of CTCP and implemented it on the Windows operating system. Our analysis and experiment results verify the properties of CTCP. Index Terms—TCP performance, delay-based congestion control, high speed network I.

Several methods have been developed for joint working and spare capacity planning in survivable wavelength-division-multiplexing (WDM) networks. These methods have considered a static traffic demand and optimized the network cost assuming various cost models and survivability paradigms. Our interest primarily lies in network operation under dynamic traffic. We formulate various operational phases in survivable WDM networks as a single integer linear programming (ILP) optimization problem. This common framework avoids service disruption to the existing connections. However, the complexity of the optimization problem makes the formulation applicable only for network provisioning and offline reconfiguration. The direct use of this method for online reconfiguration remains limited to small networks with few tens of wavelengths. Our goal in this paper is to develop an algorithm for fast online reconfiguration. We propose a heuristic algorithm based on LP relaxation technique to solve this problem. Since the ILP variables are relaxed, we provide a way to derive a feasible solution from the relaxed problem. The algorithm consists of two steps. In the first step, the network topology is processed based on the demand set to be provisioned. This preprocessing step is done to ensure that the LP yields a feasible solution. The preprocessing step in our algorithm is based on: (a) the assumption that in a network, two routes between any given node pair are sufficient to provide effective fault tolerance and (b) an observation on the working of the ILP for such networks. In the second step, using the processed topology as input, we formulate and solve the LP problem. Interestingly, the LP relaxation heuristic yielded a feasible solution to the ILP in all our experiments. We provide insights into why the LP formulation yields a feasible solution to the ILP We demonstrate the use of our algorithm on practical size backbone networks with hundreds of wavelengths per link. The results indicate that the run time of our heuristic algorithm is fast enough (in order of seconds) to be used for online reconfiguration.