Despite a growing world-wide thirst for bandwidth, supply has outpaced demand by a wide margin. During the rapid expansion of the Internet in the 1990's, the data communications industry created an infrastructure that was capable of delivering cheap bandwidth in high volumes. In fact, bandwidth has become so plentiful that even the affects of Metcalfe's Law are insufficient to consume available capacity for many years to come. The result of this imbalance has been the commoditization of bandwidth, rapidly declining bandwidth prices, and a vendor environment that has actively promoted the myth that high bandwidth can address almost any performance problem.

But as enterprise application deployments have expanded to the wide area, an environment where bandwidth is sometimes as plentiful as on the LAN, IT managers have witnessed a dramatic decrease in application performance. They scratch their heads and wonder, "Why would two networks, the LAN and the WAN, with identical bandwidth capacities, deliver such different performance results?"

The answer is that application performance is affected by many factors, associated with both network and application logic, that must be addressed in order to achieve satisfactory application performance results. At the network level, application performance is limited by high latency (the effect of physical distance), jitter, packet loss, and congestion. At the application level, performance is further limited by natural behavior of application protocols (especially when faced with latency, jitter, packet loss, and congestion at the network level), application protocols that engage in excessive handshaking across the network links, and the serialization of the applications themselves.

The goal of this article is to hopefully shed light on the issues affecting application performance in the wide area, and to give IT managers the knowledge required to design strategic enterprise application acceleration and deployment solutions.

Common Application Performance Myths

Myth #1: Application Performance Depends Only On Bandwidth

Application performance and throughput are influenced by many factors. Latency and packet loss have a profound effect on application performance. Little's Law, the seminal description of queuing theory and an equation that models the effects of physical distance (latency) and packet loss, illustrates the impacts of these two factors on application performance. As the round trip time (RTT) of each request increases, the congestion window must increase or TCP throughput will decrease. Unfortunately, TCP does not effectively manage large windows. As a result, even small amounts of latency and packet loss can quickly drop network performance for a given application to less than 1 megabit per second. Even if bandwidth capacity were to be increased to 100Mbps, the application would never consume more than 1 percent of the total capacity. Under these conditions, managers who add network capacity waste money on a resource that cannot be consumed.

In wide area networks, sources of high round trip times (i.e. latency) include physical distance, inefficient network routing patterns, and network congestion--elements that are all present in abundance on the WAN. Today, many TCP protocol stacks are highly inefficient when it comes to managing retransmissions. In fact, some stacks may have to retransmit the whole congestion window if a single packet is lost. They also tend to back off exponentially (i.e. reduce congestion windows and increase retransmission timers) in the face of network congestion, a behavior that is detected by TCP as packet loss. And while loss is often insignificant in frame relay networks (less than .01% on average), it is very significant in IP VPN networks that go into and out of certain markets like China, where loss rates commonly exceed 5%. Under this latter scenario, high loss rates can have a catastrophic effect on performance. When packet loss and latency effects are combined, the performance drop-off is even more severe.

Myth #2: TCP Requires Aggressive Back-Off To Ensure Fairness

Many network engineers believe that aggressive back-off in the face of congestion is necessary to keep network access fair. While in some cases that statement is true, in others it is not. Where congestion control is the responsibility of each host on a network, an environment where each host has no knowledge of the other host's bandwidth needs, aggressive back-off is necessary to ensure fairness. However, if congestion is managed within the fabric of the network, by a system that sees all traffic on a given WAN connection, then much greater and more efficient throughput is possible--and aggressive back off is NOT required. Standard protocol behavior specifies that when hosts consume bandwidth, they must do so independent of:

• The requirements of the application
• The amount of bandwidth is available
• The amount of competition that exists for that bandwidth.
  
  

The result is a situation where applications are often starved for bandwidth resources at the same time that the network is largely unused. This situation is obviously highly inefficient. A much better solution to the TCP fairness problem is allow individual hosts to consume as much bandwidth as they need, so long as all other hosts receive adequate service when they need it. This can be accomplished by implementing a single congestion window, shared by all hosts, that is managed within the network itself. The result is a system where hosts get the bandwidth they need in periods of light competition, and all hosts get sufficient bandwidth when competition is more intense. This single window method delivers consistently higher utilization and greater overall throughput. Hosts each see a clean, fast network that never loses packets (and therefore doesn't diminish TCP performance--see myth #1), and cumulative traffic demands are matched to the overall buffering capability of the network.

As a result, IT managers experience optimally utilized networks, under the broadest range of network latency and loss conditions. Single window solutions can be constructed that are completely transparent to client systems. Components of such solutions may include TCP technologies such as selective acknowledgement, local congestion window management, improved retransmission algorithms, and packet dispersion. These capabilities are then combined with other technologies that match the throughput requirements of applications to the availability of network resources, and that track the bandwidth requirements of all hosts utilizing the network. By aggregating the throughput of multiple, parallel WAN links, this technology can achieve even greater throughput and reliability.

Myth #3: Packet Compression Improves Application Performance

While common packet compression techniques can reduce the amount of traffic on the WAN, they often impede application performance since they tend to add latency to application transactions. These techniques requires that packets be queued up, compressed, transmitted, decompressed on the receiver, and then retransmitted--all of which can take substantial resources and add substantial latency, actually slowing down the very applications that need acceleration. Next-generation application performance solutions combine protocol streamlining with transparent data reduction techniques. Compared to packet based solutions, next-generation solutions dramatically reduce the amount of data that needs to be transmitted, eliminate latency that is introduced by protocol behavior in the face of physical distance, and can drive wide area network performance at gigabit speeds.

Myth #4: Quality of Service Technology Accelerates Applications

Quality of Service (QoS), if used properly, is a highly beneficial technology that can be helpful for managing application performance. However, the only thing that QoS can do is divide existing bandwidth into multiple virtual channels. QoS does nothing to move more data or streamline protocol behavior. QoS simply decides, in an intelligent way, which packets to drop. And while it is better to drop packets in a controlled way than to leave it to chance, dropping packets does not accelerate applications.

Many QoS implementations rely on port numbers to track applications. Since applications often negotiate port assignments dynamically, these mechanisms have to be configured to reserve a large port ranges to ensure coverage of the ports actually used by the application. For QoS to be most effective, it should be dynamic. Dynamic QoS solutions ensure that bandwidth is reserved only when applications can use it.

The Solution

So how can a business overcome these challenges? One such way is to install a WAN application acceleration product. These products can deliver dramatic application performance and greatly reduced WAN costs by monitoring the limiting effects of network conditions, adjusting protocol behavior, and by managing all levels of the protocol stack, from the network layer through to the application layer. It's important that these products are built around an architecture that can recognize the critical interdependence between application-level and transport-level behavior. It should integrate advanced transport acceleration technologies such as adaptive TCP acceleration and session-aware quality of service, with application acceleration technologies that include dynamic (XML) object caching, application proxies, and application encryption. The system should also be supported by a statistics generation and monitoring engine that enables real-time management of application network behavior.

These types of products can essentially deliver LAN-like application performance over the WAN, accelerating applications such as ERP, CRM, email, file transfer, data replication, and other applications, resulting in predictable, fast performance for all WAN users.

Summary

Application performance on the WAN is affected by a large number of factors in addition to bandwidth. The notion that bandwidth solves all, or even most application performance problems is a complete myth.

At the network level, application performance is limited by high latency (the effect of physical distance), jitter, packet loss, and congestion. At the application level, performance is likewise limited by factors such as: the natural behavior of application protocols that were not designed for WAN conditions; application protocols that engage in excessive handshaking; and the serialization of the applications themselves.

A solid application acceleration solution will be able to recognize the critical interdependence between application-level and transport-level behavior. The best of these solutions can deliver predictable application performance, increase throughput many times over, and improve application performance on networks as diverse as premium-quality, class-of-service managed networks to commodity, best efforts-based Internet IP VPN's. The result is increased application performance, massive scalability, and a return on investment that is measured in months instead of years. ENS