Technology, G. Johnson Research Note 15 September 2003 Optimize Your Broadband WAN When migrating from narrowband to broadband WANs, be aware that TCP's limitations can severely affect network performance, despite investment in more capacity. Multiple network performance management tools are needed. Core Topic Enterprise Networking: Network Equipment Vendors and Technologies Key Issues During the five-year planning period, which networking equipment vendors and technologies will enterprises use to successfully build their networks? Which WAN technologies will have the greatest impact on enterprises? Strategic Planning Assumptions Through 2008, more than 80 percent of enterprises that migrate from narrowband WANs to broadband WANs will face poor network performance because of latency (0.7 probability). By 2006, less than 50 percent of enterprises will understand TCP's limitations in long-distance WANs, and less than 20 percent will have implemented network traffic management solutions (0.8 probability). After investing in broadband bandwidth to create high-speed WANs, enterprises are discovering that their network application performance has only marginally improved, and real-time, consistent performance is not assured. The limitations of TCP must be better understood now that the use of TCP/IP to reach over global distances is pervasive in internetworking and broadband WANs. TCP "governs" traffic with send, acknowledge, resend and congestion controls. HTTP and FTP are carried over TCP. In the LAN, bandwidth is inexpensive and traffic does not suffer from latency because of the short distances; overprovisioned bandwidth creates negligible delay. However, in the WAN, network latency increases with distance. Life can be terrible on global WANs, especially for large file transfers over different hemispheres. Network Performance Optimization Options, Techniques and Actions There are many ways to optimize the performance of networks. All are piecemeal solutions that solve particular problems for specific applications and the unique types of network traffic that they generate. Logical or physical separation of network infrastructure for disparate low-latency and high-latency applications, such as voice and video vs. data communications, can minimize the effects of competing and interacting traffic over a high-latency data communications infrastructure. This is essential where applications that demand low latency (such as voice and video) are carried on a data communications network that has high latency, and expects to buffer traffic and retransmit to handle delay or lost packets. A common example is the carriage of voice over frame relay (not often used), where voice is carried over its Gartner Reproduction of this publication in any form without prior written permission is forbidden. The information contained herein has been obtained from sources believed to be reliable. Gartner disclaims all warranties as to the accuracy, completeness or adequacy of such information. Gartner shall have no liability for errors, omissions or inadequacies in the information contained herein or for interpretations thereof. The reader assumes sole responsibility for the selection of these materials to achieve its intended results. The opinions expressed herein are subject to change without notice.
own permanent virtual circuit (PVC), and pure data traffic remains on a separate PVC to limit interaction effects. The design issue is to mask "a little" voice inside "a lot" of data. This approach to modifying WAN performance relies on the "assurance" that each type of application will be carried on its own dedicated link, which is sufficiently optimal for the type of traffic offered. Action Item: Review the scope of logically or physically separate networks, and compare their relative total costs of ownership. Merely adding bandwidth uses brute force on the operational symptoms and ignores the network performance problem. Web caching and local replication of databases, such as Lotus Notes, reduce WAN loading by decreasing frequent "chatty" traffic over long-haul networks. Pipe size is fundamental, but it must be understood, rather than abused by overcapacity. The complexity of multiple contending applications on IP networks requires informed iterative design, not "guesstimates." Adding bandwidth does not address other issues, such as congestion control and traffic management. Because there is no bandwidth reservation in TCP/IP, large file transfers will always swamp time-sensitive usage, such as voice and video, on the same path. Mission-critical networks can't survive without some congestion control and application priority management for critical traffic. Increasing bandwidth capacity also does not offer insight into the demand and usage of bandwidth by various applications, some of which may be bandwidth "hogs." Action Item: Identify network segments where adding bandwidth is a legitimate approach (for example, LANs and campus LANs). Examine alternatives for other network segments. Web caching and replication reduce response times by locating data repositories as close to the local user as possible to reduce latency and multiple messages over long-haul networks. Caching is particularly suitable for Web applications, but it does not guarantee network availability, or reserve or restrict network resources to suit critical applications. Major Web caching vendors include 3Com, Cache Flow (now Blue Coat Systems), Cisco Systems, Dell, Equiinet, F5 Networks, Hewlett-Packard, InfoLibria, Network Appliance and Sun Microsystems. Action Item: Design caching and replication server locations that are properly suited to the traffic, applications and capacities of your network. 15 September 2003 2
Bandwidth management solutions can provide congestion control by guaranteeing network resources to applications or users, and restricting the impact of other (foreign) users by limiting use of available bandwidth. Bandwidth management can add significant network overheads and relies on proprietary hardware to manage networks. Bandwidth management vendors include Allot Communications (NetEnforcer), Orchestream, NetScreen Technologies, Packeteer, Sun (SunSoft bandwidth allocator) and Visual Networks. Action Item: Ensure that bandwidth managers are used to create an understanding of an application's peak and intermittent bandwidth and latency needs, as well as its interactivity and impact on other traffic. Network management software alerts operators to heavy traffic usage on links or aggregate routes. These alarms do not improve network performance, but they offer strategic and tactical advice that can be used as a warning to redesign the network, especially its use of various protocols. Tools from BMC Software, Computer Associates, HP OpenView, IBM Tivoli and other vendors increase functionality, but they do not necessarily provide analysis or insight into the end-to-end performance of an application, such as the tools provided by Compuware and others do. Action Item: Leverage outputs from alerts, event alarms and traffic performance monitoring into tactical fixes and strategic dimensioning of network capacity and management of application deployments. Router-based queuing relies on router hardware to apply traffic queuing algorithms that may suit a combination of application traffic types. Queuing solutions require a deep technical understanding of the protocols at work and their impact on the application mix on the network. Each queuing variant solves a problem, and it has its own implementations and parameter settings to tune network performance. Variants include: Traffic prioritization (weighted random early detection) Rate-limiting (random early detection) Weighted fair access sharing Hierarchical class-based queuing, where packets are broken down in a hierarchical tree that allows data packets to borrow bandwidth from sister branches when capacity is used The router queuing approach adds a heavy overlay of technical complexity in router configuration and management. It also requires expert knowledge of the design and its maintenance, as 15 September 2003 3
well as adding additional cost because new software and hardware must be applied to deliver this solution. Cisco dominates the enterprise router market and first-generation queuing solutions, although there are alternatives to its IOS and proprietary hardware. Enterasys Networks, Nortel Networks and other router vendors have router queuing solutions. Action Item: Investigate router queuing optimization, but consider the complexity that is created in configuration management as a trade-off. Intelligent routing can dynamically direct traffic across a number of internetworking links to arbitrate between network performance and available link prices. Intelligent routing assumes that an enterprise will have multiple links to the Internet, possibly through meshed networks, where intelligent routing can dynamically allocate IP traffic to the best available links. This approach does not guarantee network availability, or provide congestion control or limit network resources to particular applications. However, it will make the best of the available paths. A popular rule of thumb is, "Switch where you can; route where you must." Vendors include F5, netvmg, Proficient Networks, Radware, RouteScience Technologies and Sockeye Networks. Action Item: Midsize and large enterprises will have sufficient meshing in their networks to justify the expense and overhead of intelligent routing. Bandwidth compression is a favorite technique for voice and data communications. Bandwidth compression reduces the total number of bits offered. It is a well-established method of making the best of thin routes or restricted capacity in enterprise networks. These solutions reduce the total amount of traffic to be carried on a WAN link and effectively increase the capacity of otherwise limited pipes. Savings can occur based on the lower volume of traffic transmitted across the network. Although compression does not offer sophisticated availability guarantees, congestion control or undesirable usage control mechanisms, it is an important element in other hybrid network performance solutions. Vendors such as Expand Networks and Peribit Networks offer hybrid compression and latency optimization tools (see "Time to (Re)evaluate Network Compression"). Action Item: Evaluate compression and related hybrid technologies for their substantial improvement potential in some 15 September 2003 4
network segments, but understand these technologies' limitations and the costs of proprietary equipment and software. Augmentation of the standard TCP protocol stack with trafficshaping algorithms can dramatically improve throughput in saturated high-latency networks. Augmenting the standard TCP protocol stack considers the limitations of TCP. The use of proprietary mathematical algorithms can reproduce traffic streams identically without much of the original protocol overhead. This requires a proprietary client and server, with both using the optimized protocol stack. Augmentation mitigates the impact of transporting TCP-based applications (especially large FTP files), particularly over highlatency, global WANs. Its network performance gains are most spectacular when proprietary algorithms relieve the saturation effects of high-latency, round-trip times and packet loss in large file transfers, such as multimedia or database replication (see Figure 1). Digital Fountain and Foursticks are examples of protocol augmentation algorithm vendors. Maximum TCP Throughput (Mbps) 45 Figure 1 Maximum Throughput for TCP-Based Applications TCP Augmented Protocol 40 35 30 25 Traffic saturation in TCP with high latency. T3 barely delivers T1. 20 15 10 5 0 1.5 5 10 15 20 25 30 35 40 45 (T1 speed) WAN Bandwidth (Mbps) (T3 capacity) Source: Gartner Research (August 2003) Action Item: Enterprises that shift large files or data transfers over continental, international or global networks likely will discover significant bandwidth cost and throughput 15 September 2003 5
improvements using suitable TCP-augmented protocols that are deployed over proprietary servers and clients. Other techniques such as forward error correction help networks with very-high-performance requirements, such as air traffic control and military applications. Bottom Line: As WAN bandwidth becomes less expensive and broadband capacity routinely is used, your largest operating exposure may be the limitations of TCP/IP in WANs. Simply adding bandwidth to achieve high performance will work in LANs, metropolitan-area networks and short-haul WANs, but not in long-haul networks, where the latency caused by distance severely affects network throughput. Audit your network for performance and evaluate WAN traffic management options based on the design problem or limitation to ensure that your investment in fast network capacity is not wasted, and that IT applications do not run poorly. 15 September 2003 6