Figure 2 - uploaded by Wu Feng
Content may be subject to copyright.
Source publication
With the advent of computational grids, networking performance over the wide-area network (WAN) has become a critical component in the grid infras- tructure. Unfortunately, many high-performance grid applications only use a small fraction of their available bandwidth because operating systems and their associated protocol stacks are still tuned for...
Context in source publication
Similar publications
With the advent of computational grids, networking performance over the wide-area network (WAN) has become a critical component in the grid infrastructure. Unfortunately, many high-performance grid applications only use a small fraction of their available bandwidth because operating systems and their associated protocol stacks are still tuned for y...
With the widespread arrival of bandwidth-intensive applications such as bulk-data transfer, multi-media web streaming and computational grids for high-performance computing, networking performance over the wide-area network has become a critical component in the infrastructure. Tragically, operating systems are still tuned for yesterday's WAN speed...
Citations
... Tools for choosing parameter values to maximize file transfer performance have been a frequent area of interest. For TCP, dynamic right-sizing of buffers has been used to improve file transfer throughput [9], [11]. The GridFTP-APT project [14] develops models that identify TCP buffer sizes and number of TCP flows for improved transfer performance [12], and builds tools for dynamically changing the number of connections during a file transfer [13]. ...
File transfers over dedicated connections, supported by large parallel filesystems, have become increasingly important in high-performance computing and big data workflows. It remains a challenge to achieve peak rates for such transfers due to the complexities of file I/O, host, and network transport subsystems, and equally importantly, their interactions. We present extensive measurements of disk-to-disk file transfers using Lustre and XFS filesystems mounted on multi-core servers over a suite of 10 Gbps emulated connections with 0–366 ms round trip times. Our results indicate that large buffer sizes and many parallel flows do not always guarantee high transfer rates. Furthermore, large variations in the measured rates necessitate repeated measurements to ensure confidence in inferences based on them. We propose a new method to efficiently identify the optimal joint file I/O and network transport parameters using a small number of measurements. We show that for XFS and Lustre with direct I/O, this method identifies configurations achieving 97% of the peak transfer rate while probing only 12% of the parameter space.
... The imperfections of the incumbent Internet architecture has fed a growing frustration among the large-scale science community that suffered most acutely from the inability of the Internet to translate tremendous capacities of optical communication media into adequate performance of large-scale scientific applications. Early attempts to remedy the status quo focused on automatic tuning of the Internet protocol suite [18]–[21]. However, accumulated empirical evidence revealed that neither modification of the existing Internet transport protocols nor design of new protocols within the Internet framework is likely to yield satisfactory support for the challenging communication needs of large-scale science. ...
Within the framework of the diversified Internet where architecturally different metanetworks coexist on a shared communication substrate, we present a large-scale scientific metanetwork (LSM) designated for special communication needs of large-scale science. LSM emphasizes performance and security over horizontal scalability and offers two services to users: prompt-delivery service for quick transfers of bulk data and assured-capacity service for applications that require delay or rate guarantees. The prompt-delivery service relies on message-grained scheduling to provide near-minimal average message delay while treating each individual message fairly. In support of the assured-capacity service, LSM maintains differential tree data structures to store and update advance reservations of communication capacities. Explicit accounting for real-world entities involved in communications forms a basis for secure capacity allocation in LSM.
... In a distributed grid computing infrastructure, we have to deal with different computational capacity among the clusters, but also with an heterogeneous network support linking them [Feng et al., 2003, Schulze and Madeira, 2004]. Therefore, a monitoring infrastructure may contribute in supporting a highly dynamic environment where operational conditions are constantly changing [Quéma et al., 2004]. ...
Grid computing and Internet measurements are two areas that have taken off in recent years, both receiving a lot of attention from the research community. In this position paper, we argue that these two promising research areas have a strong synergy that bring mutual benefits. Based on such considerations, we propose a measurement middleware service for grid computing. By defining the architecture and the methods of this service, we show that a promising symbiosis may be envisaged by the use of the proposed measurement middleware service for grid computing.
... According toFigure 4, the time difference between ACKi+1 and the shaded pair is precisely an RTT. The authors of [18] suggested a similar approach to estimate an upper bound of the RTT, the delay between ACKi (instead of ACKi+1) and the packet pair inFigure 4, which can result in higher estimates if the situation explained inFigure 5 occurs. The reliability of our estimates will increase with the number of packet pairs we identify and the large flows in grid will help us to find a large number of packet pairs, resulting in more reliable estimates. ...
For efficient use of geographically distributed resources in a grid, the selection of the optimum site is highly important for an autonomic grid scheduler. We propose a reliable network measurement and prediction architecture that helps its clients with their scheduling decisions by informing them about the minimum time that it will take to transfer certain amount of data. This is achieved by calculating TCP throughput under ideal conditions. Our prediction is based on the "packet pair" measurement method; we introduce a receiver side passive capacity estimation technique which additionally calculates a reliable lower bound of the Round Trip Time. Passive operation is feasible in a grid, where large file transfers between nodes are frequent, and it ensures non-intrusiveness of our architecture; active measurements are only initiated when they are needed by clients.
... @BULLET Simple unreliable datagram delivery (UDP) @BULLET Unreliable congestion controlled datagram delivery (DCCP) · with a choice of congestion control mechanisms · with or without delivery of erroneous data @BULLET Reliable congestion controlled in-order delivery of · a consecutive data stream (TCP) · multiple data streams (SCTP) @BULLET Reliable congestion controlled unordered but potentially faster delivery of logical data chunks (SCTP) This is only a rough overview: each protocol provides a set of features and parameters that can be tuned to suit the environment or application — for example, the size of an SCTP data chunk (or UDP or DCCP datagram) represents a trade-off between end-to-end delay and bandwidth utilization (small packets are delivered faster but have a larger per-packet header overhead than large packets). Work based on TCP parameter tuning is described in [12,13]. Nowadays, the problem of the transport layer is its lack of flexibility. ...
Despite the many research efforts at the transport layer (SCTP, DCCP, etc.), new innovations in that area hardly ever make it into the TCP/IP stacks of standard end systems. We believe that this is due to lack of a flexible interface to the application as well as a lack of transparency – a problem that could be solved by introducing a middleware above TCP/IP. In this paper, we present the architecture of our middleware, and show the importance of congestion awareness by simulations. In addition, we explain how to force congestion control on existing applications using our middleware, show the benefits of doing so with simulations, and finally discuss the impact of our middleware towards a better utilization of the network and a more suitable service for the user software.
... A common optimization for LFNs sets the window size to be at least as large as the bandwidth-delay product. 7,8 However, setting too large a window wastes memory and can severely affect throughput by allowing the sender to overrun the link capacity. ...
Apart form the success in local-area networks (LANs) and system-area networks and anticipated success in metropolitan and wide area networks (MANs and WANs), Ethernet continues to evolve to meet the increasing demands of packet-switched networks. Although the recently ratified 10-Gigabit Ethernet standard differs from earlier Ethernet standards, primarily in that 10GbE operates only over fiber and only in full-duplex mode, the differences are largely superficial. More importantly, l0GbE does not make obsolete current investments in network infrastructure. The 10GbE standard ensures interoperability not only with existing Ethernet but also with other networking technologies such as Sonet, thus paving the way for Ethernets expanded use in MANs and WANs. The world's first host-based 10GbE adapter, officially known as the Intel PRO/10GbE LR server adapter, introduces the benefits of l0GbE connectivity into LAN and system-area network environments, thereby accommodating the growing number of large-scale cluster systems and bandwidth-intensive applications, such as imaging and data mirroring. The 10GbE controller is optimized for servers that use the I/O bus backplanes of the peripheral component interface (PCI) and its higher speed extension, PCI-X.
... To summarize, an Internet application programmer can be expected to face the following choice of transport services (provided by her operating system) within the next couple of years: @BULLET Reliable congestion controlled unordered but potentially faster delivery of logical data chunks (SCTP) This is only a rough overview: each protocol provides a set of features and parameters that can be tuned to suit the environment or application — for example, the size of an SCTP data chunk (or UDP or DCCP datagram) represents a 20/2 trade-off between end-to-end delay and bandwidth utilization (small packets are delivered faster but have a larger per-packet header overhead than large packets). Work based on TCP parameter tuning is described in [12] and [13]. Nowadays, the problem of the transport layer is its lack of flexibility. ...
Despite the many research efforts related to Internet congestion control and Quality of Service, new innovations in these two areas hardly ever make it into the TCP/IP stacks of standard end systems. We be- lieve that this is due to lack of a well-defined interface to the application as well as a lack of transparency - a problem that could be solved by introducing middle- ware above TCP/IP.
... for proper flow-control adaptation. Two approaches to dynamically tuning buffer sizes are auto-tuning [23] and dynamic right-sizing (DRS) [7] [9]. The former is a sender-based approach to flow control, while the latter is a receiver-based approach. ...
Summary form only given. The performance of TCP in wide-area networks (WANs) is becoming increasingly important with the deployment of computational and data grids. In WAN environments, TCP does not provide good performance for data-intensive applications without the tuning of flow-control buffer sizes. Manual adjustment of buffer sizes is tedious even for network experts. For application scientists, tuning is often an impediment to getting work done. Thus, buffer tuning should be automated. Existing techniques for automatic buffer tuning only measure the bandwidth-delay product (BDP) during connection establishment. This ignores the large fluctuation of the BDP over the lifetime of the connection. In contrast, the dynamic right-sizing algorithm dynamically changes buffer sizes in response to changing network conditions. We describe a new user-space implementation of dynamic right-sizing in FTP (drsFTP) that supports third-party data transfers, a mainstay of scientific computing. In addition to comparing the performance of the new implementation with the old in a WAN-emulated environment, we give performance results over a live WAN. In this particular WAN environment, the new implementation produces transfer rates of up to five times higher than untuned FTP.
... One set of solutions calls for enhancing TCP to improve end-to-end throughput, thus limiting upgrades to the end hosts. Such improvements can be made via congestion control [2][3][4] and/or flow control [5][6][7] . A second set of solutions requires upgrades to routers within the Internet. ...
Leveraging the dominance of Ethernet in LANs and SONET/SDH in MANs and WANs, we propose a service called CHEETAH (Circuit-switched High-speed End-to-End Transport ArcHitecture). The service concept is to provide end hosts with high-speed, end-to-end circuit connectivity on a call-by-call shared basis, where a "circuit" consists of Ethernet segments at the ends that are mapped into Ethernet-over-SONET long-distance circuits. This paper focuses on the file- transfer application for such circuits. For this application, the CHEETAH service is proposed as an add-on to the primary Internet access service already in place for enterprise hosts. This allows an end host that is sending a file to first attempt setting up an end-to-end Ethernet/EoS circuit, and if rejected, fall back to the TCP/IP path. If the circuit setup is success- ful, the end host will enjoy a much shorter file-transfer delay than on the TCP/IP path. To determine the conditions under which an end host with access to the CHEETAH service should attempt circuit setup, we analyze mean file-transfer delays as a function of call blocking probability in the circuit-switched network, probability of packet loss in the IP network, round-trip times, link rates, and so on.
As the Internet becomes increasingly heterogeneous, the issue of congestion control becomes ever more important. In order to maintain good network performance, mechanisms must be provided to prevent the network from being congested for any significant period of time. Michael Welzl describes the background and concepts of Internet congestion control, in an accessible and easily comprehensible format. Throughout the book, not just the how, but the why of complex technologies including the Transmission Control Protocol (TCP) and Active Queue Management are explained. The text also gives an overview of the state-of-the-art in congestion control research and an insight into the future. Network Congestion Control: Presents comprehensive, easy-to-read documentation on the advanced topic of congestion control without heavy maths. Aims to give a thorough understanding of the evolution of Internet congestion control: how TCP works, why it works the way it does, and why some congestion control concepts failed for the Internet. Explains the Chiu/Jain vector diagrams and introduces a new method of using these diagrams for analysis, teaching & design. Elaborates on how the theory of congestion control impacts on the practicalities of service delivery. Includes an appendix with examples/problems to assist learning. Provides an accompanying website with Java tools for teaching congestion control, as well as examples, links to code and projects/bibliography. This invaluable text will provide academics and researchers in computer science, electrical engineering and communications networking, as well as students on advanced networking and Internet courses, with a thorough understanding of the current state and future evolution of Internet congestion control. Network administrators and Internet service and applications providers will also find Network Congestion Control a comprehensive, accessible self-teach tool.
