Dino Lopez-Pacheco

Robust transport protocol for dynamic high-speed networks: enhancing the XCP approach

Transport protocols have the difficult task of providing reliability and fair sharing of the bandwidth to end-users. In this paper, we focus on very dynamic high-speed networks where the available best-effort bandwidth for regulated traffics can vary over time. Foreseen problems introduced by such highly dynamic environments are inefficiency due to convergence time and high amount of packet losses due to dramatic reductions of the available bandwidth. Therefore end-to-end solutions show their limitations for exploiting the current very high-speed infrastructures. XCP is a promising approach as the evolution of the sender congestion window size is dictated by the routers. However, as the XCP sender relies on the returned ACKs to adapt its congestion window size, XCP performances can be affected if many losses occur on the reverse path. This paper proposes to calculate the congestion window size at the receiver side and to overcome the problem of ACK losses. The result is a more robust XCP transport protocol, capable of achieving a high level of performance in very dynamic high-speed networks, thus able to function in a broader range of network conditions.

NXG03-4: XCP-i : eXplicit Control Protocol for Heterogeneous Inter-Networking of High-Speed Networks

XCP is a transport protocol that uses the assistance of specialized routers to very accurately determine the available bandwidth along the path from the source to the destination. In this way, XCP efficiently controls the senders congestion window size thus avoiding the traditional slow-start and congestion avoidance phase. However, XCP requires the collaboration of all the routers on the data path which is almost impossible to achieve in an incremental deployment scenario of XCP. It has been shown that XCP behaves badly, worse than TCP, in the presence of non-XCP routers thus limiting dramatically the benefit of having XCP running in some parts of the network. In this paper, we address this problem and propose XCP-i which is operable on an internetwork consisting of XCP routers and traditional IP routers without loosing the benefit of the XCP control laws. The simulation results on a number of topologies that reflect the various scenario of incremental deployment on the Internet show that although XCP-i performances depend on available bandwidth estimation accuracy, XCP-i still outperforms TCP on high-speed links.

Too Many SDN Rules? Compress Them with MINNIE

Software Defined Networking (SDN) is gaining momentum with the support of major manufacturers. While it brings flexibility in the management of flows within the data center fabric, this flexibility comes at the cost of smaller routing table capacities. In this paper, we investigate compression techniques to reduce the forwarding information base (FIB) of SDN switches. We validate our algorithm, called MINNIE, on a real testbed able to emulate a 20 switches fat tree architecture. We demonstrate that even with a small number of clients, the limit in terms of number of rules is reached if no compression is performed, increasing the delay of all new incoming flows. MINNIE, on the other hand, reduces drastically the number of rules that need to be stored with a limited impact on the packet loss rate. We also evaluate the actual switching and reconfiguration times and the delay introduced by the communications with the controller.

Understanding the impact of TFRC feedbacks frequency over long delay links

TFRC is a transport protocol specifically designed to carry multimedia streams. TFRC does not enable a reliable and in order data delivery services. However TFRC implements a congestion control algorithm which is friendly with TCP. This congestion control relies in a feedback mechanism allowing receivers to communicate to the senders an experienced drop rate. Although the current TFRC RFC states that there is little gain from sending a large number of feedback messages per RTT, recent studies have shown that in long-delay contexts, such as satellite-based networks, the performance of TFRC can be improved by increasing the feedback frequency. Nevertheless, currently it is not clear how and why this increase may improve the performance of TFRC. Therefore, in this paper, we aim at understanding the impact that multiple feedback per RTT may have (i) on the key parameters of TFRC (RTT and error rate) and (ii) on the network parameters (reactiveness, fairness and link utilization). We also provide a detailed description of the micromechanisms at the origin of the improvements of the TFRC behavior when multiple feedback per RTT are delivered, and determine the context where such feedback frequencies should be applied.

Enabling Large Data Transfers on Dynamic, Very High-Speed Network Infrastructures

High-speed networks are being built using the most upto- date networking technologies such as optical fibers with dense wavelength multiplexing (DWDM) in order to meet the performance constraints of large scientific problems (grid infrastructures) or new multimedia applications. As technologies are moving forwards, many high-speed network infrastructures have amazing capabilities in the order of several gigabits/s. Our work focuses on very dynamic high-speed environments where the available besteffort bandwidth for regulated traffics can vary over time. Foreseen problems introduced by such highly dynamic environments are inefficiency due to convergence time and high amount of packet losses due to dramatic reductions of the available bandwidth. In a previous study, we proposed the XCP-r protocol which is a more robust XCP-based transport protocol. In this paper, we present simulation results comparing XCP-r, XCP, TCP (NewReno) and HSTCP for large data transfers on very dynamic high-speed networks. We show that XCP-r succeeds in providing a high level of performance thus able to function in a broader range of network conditions.

Understanding TCP cubic performance in the cloud: A mean-field approach

Cloud networking typically leads to scenarii where a large number of TCP connections share a common bottleneck link. In this paper, we focus on the case of TCP Cubic, which is the default TCP version in the Linux kernel. TCP Cubic is designed to better utilize high bandwidth-delay product path in an IP network. To do so, Cubic modifies the linear window growth function of legacy TCP standards, e.g., New Reno, to be a cubic function. Our objective in this work is to assess the performance of TCP Cubic in a cloud setting with a large number of long-lived TCP flows.We rely on a mean-field approach leading to a fluid model to analyze the performance of Cubic. After a careful validation of the model through comparisons with ns-2, we evaluate the efficiency and fairness of Cubic as compared to that of New Reno for a set of typical cloud networking scenarii.

Understanding the network level performance of virtualization solutions

Virtualization is the cornerstone of modern private and public cloud solutions. By enabling the consolidation of many virtual machines (VMs) within a physical server, virtualization blurs the frontier between networking and system. The interconnection network now starts within the physical servers, where several (virtual) servers compete to access the network interface cards (NICs). However, our understanding of the networking service offered to VMs by a specific hypervisor is largely unknown. In this work, we adopt an experimental approach to uncover different costs related to accessing a virtual NIC (as exposed by the hypervisor to the VMs), sharing it among VMs and to managing the VMs themselves, which potentially introduce an extra delay at each packet emission. We consider the case of Xen and VMware and propose a technique to identify those different delays through the design of specific experimental scenarios.

Optimal configuration for satellite PEPs using a reliable service on top of a routers-assisted approach

Router-assisted congestion control protocols, also known as Explicit Rate Notification (ERN) protocols, implement complex algorithms inside a router in order to provide both high link utilization and high fairness. Thus, router-assisted approaches overcome most of the end-to-end protocols problems in large bandwidth-delay product networks. Today, router-assisted protocols cannot be deployed in heterogeneous networks (e.g., Internet) due to their non-compliance with current network protocols. Nevertheless, these approaches can be deployed in satellite networks in the context of splitting PEPs. In this work, as router-assisted protocols can use TCP algorithms to enable reliability, we aim at understanding and providing a detailed view of the impact of such algorithms on the performance obtained by router-assisted protocols over satellite links. In particular, we both study XCP and P-XCP proposals over long delay, lossy and asymmetric links and propose an ns-2 implementation of the P-XCP protocol to the satellite community. To the best of our knowledge, this study is the first one which tackles the impact of TCP internal mechanisms on XCP protocol. Our main conclusion is that P-XCP on TCP New Reno Slow But Steady variant is to date, the most optimal configuration for satellite proxies.

Coarse-grained Scheduling with Software-Defined Networking Switches

Software-Defined Networking (SDN) enables consolidation of the control plane of a set of network equipments with a fine-grained control of traffic flows inside the network. In this work, we demonstrate that some coarse-grained scheduling mechanisms can be easily offered by SDN switches without requiring any unsupported operation in OpenFlow. We leverage the feedback loop - flow statistics - exposed by SDN switches to the controller, combined with priority queuing mechanisms, usually available in typical switches on their output ports. We illustrate our approach through experimentations with an OpenvSwitch SDN switch controlled by a Beacon controller.

Understanding synchronization in TCP Cubic

TCP Cubic is designed to better utilize high bandwidth-delay product paths in IP networks. It is currently the default TCP version in the Linux kernel. Our objective in this work is to better understand the performance of TCP Cubic in scenarios with a large number of competing long-lived TCP flows, as can be observed, e.g., in cloud environments. In such situations, Cubic connections tend to synchronize each other and this synchronization is higher than with TCP New Reno. We investigate this phenomenon in detail through experimentations in a controlled testbed, measurements with Amazon EC2s servers, located in the US and simulations. We demonstrate that several factors contribute to the appearance of synchronization in TCP Cubic: (i) the rate of growth of the congestion window when a Cubic source reaches the capacity of the network and its relation to the RTT of the connection, (ii) the way the congestion Cubic tracks the ideal cubic curve in the kernel (as the congestion window grows in a discrete fashion in units of MSS while the cubic curve assumes a fluid window), (iii) the competition among the Cubic sources and the aggressiveness of the sources that did not experience losses during the last loss episode. We also propose and evaluate two propositions to the TCP Cubic algorithm to alleviate the amount of packets lost during the synchronization episodes.