Overcoming the root causes of slow and unpredictable web-enabled application response times is a business-critical challenge for today's IT manager. Poor performance leads to low adoption rates, ineffective business processes, high abandonment rates and poor user satisfaction. Malcolm Rowe examines why application performance solutions address the root causes of 'middle mile' bottlenecks to transform the internet into a high-performance and reliable application delivery platform.

The performance and availability of web-based enterprise applications frequently falls far short of end-user expectations. Users remotely located from an application server can experience application response times 10 times greater than the sub-second response times experienced by close proximity users. The result is a poor application experience that's potentially unusable for a globally distributed user base.

The performance degradation caused by the internet's routing protocols - BGP and TCP - and the way these interact with the web's HTTP communication protocols are just part of the conundrum. With web applications becoming increasingly dynamic and interactive, IT organisations face additional implementation challenges when attempting to improve the performance of dynamic applications and web services. Firstly, dynamic content does not benefit from caching; a technique typically employed by traditional content delivery networks (CDNs) to improve response times associated with static or frequently used content. Secondly, unpredictable usage can result in an overwhelmed or excessively built-out application infrastructure.

Exploring the cause of 'middle mile' bottlenecks

The inherent design of the internet, which is in essence 'a network of networks', is the primary cause of application delivery problems to a global user base. Information sent across a network incurs some degree of latency as it travels from source to destination - measured as internet protocol (IP) round-trip time (RTT). Network bottlenecks associated with service provider peering points and congestion blackouts and brownouts all add additional latency and packet loss to each RTT. Across a single geographic region - such as the US or Europe - latency and packet loss can exceed 100msec with a 10 per cent packet loss, while between two geographies - for example, the US and Asia Pacific regions - RTT rises to over 500msec, with a 20 per cent packet loss.

The middle-mile bottleneck is further compounded by the 'RTT multiplier' effect resulting from the multiple back-and-forth round trips between client and server that take place during the web page download; as many as 30-50 round trips are standard for many web-enabled applications. This results in a significantly increased RTT multiplier for users residing further away from the origin server. A typical web page - containing an HTML body and numerous embedded images and objects - uses several separate internet transport protocol (TCP) connections; even with no network disruptions, this results in a large RTT multiplier, the scale of which is dependent on page size, number of objects and their size, as well as browser and server technologies.

RTT and RTT multiplier are the key culprits responsible for the 'middle-mile' performance bottlenecks that limit acceptable application response times to a distributed user base. Two other causes include the extreme sensitivity of application response times to internet congestion and distribution - which manifests as packet loss - and poor internet route selection. Lost or out-of-order packets result in retransmissions and TCP timeouts, both of which further increase the RTT multiplier.

'First-mile' bottlenecks, that is the ability of the origin server infrastructure to scale to meet end-user demand, such as TCP/IP connection terminations, server load balancing, content compression and SSL encryption/decryption, can be addressed by origin server offload techniques.

Tackling the challenge - the expensive way

The 'bricks and mortar' approach to solving poor performing web applications is to build out bigger and more data centres, adding servers and bandwidth to handle demand while moving applications and content closer to geographically dispersed users. Building an infrastructure capable of handling peak demand is costly, and often results in periods of under-utilised capital assets. Furthermore, this approach introduces the need for data replication and synchronisation - all of which adds further cost, and introduces both additional complexity and compliance risk.

While acceptable for intranets with a small number of locations, a bi-nodal WOC solution is incapable of supporting extranet applications or web service queries on the internet, where the end-user community is outside the corporate WAN environment.

For enterprises deploying application delivery appliances, upfront capital costs are just the beginning of the total cost of ownership (TCO) calculation. Other costs include hardware maintenance, updates, replacement costs, the need for additional IT staff, extended deployment time, increased time to value and technical obsolescence.

A better way

Application delivery network (ADN) services have emerged as a comprehensive internet-based platform for improving web application performance. By implementing an overlay network to the internet, ADN service providers use internet intelligence combined with techniques employed by traditional CDNs and application delivery appliances, to transform the internet into a high-performance delivery platform.

The outcome is a local response time for global users, high availability, on-demand scalability and enterprise security - with no changes to applications or data centre infrastructure. Alongside the guaranteed application performance and availability that's enshrined in service level agreements, the TCO of an ADN service is lower than appliance-based approaches, providing a predictable monthly service cost.

ADN delivers dramatic performance gains

Addressing the shortcomings of the core internet protocols, ADN accelerates delivery by tackling the causes of first and middle-mile internet bottlenecks through a purpose built bi-nodal overlay network of highly distributed specialised servers. Located in close proximity to end-users, as well as the origin infrastructure, this creates a direct mapped network between centralised applications and users across the edge of the internet.

Locating servers as close as possible to end-users effectively minimises latency and packet loss by optimising the RTT and RTT multiplier that typically generates bottlenecks. End-user access requests to a web application can be processed dynamically and directed to the nearest server; the ADN network's user and origin servers interject themselves into the path between the client and application server, with no change in the transfer process or application infrastructure.

Optimised routing is enabled by algorithms that provide a real-time 'weather map' of the internet to support performance-based routing decisions that overcome the inefficiencies inherent in BGP and ensuring an optimal RTT for every round-trip; end-user requests reach the application server, regardless of internet bottlenecks caused by service provider blackouts, de-peering, network outages or even natural disasters such as earthquakes. The RTT multiplier effect is reduced further through the substitution of specialist communication protocols between the origin server and the overlay network's end-user servers; traffic is only converted back to standard HTTP/TCP at the end destination server.

TCP inefficiencies are managed through guaranteed pipelining between the network's servers and the origin server's data centre, intelligent retransmission to overcome the TCP timeout parameter issue, and addressing HTTP's requirement for multiple TCP connections for page delivery through a combination of intelligent pre-fetching - to ensure all page elements are delivered as if the origin server is a few milliseconds away - data compression to reduce bandwidth usage, and the storage of any cacheable content close to end-users.

To ensure application performance and scale, first mile performance techniques are used to 'offload' process intensive application server tasks; TCP/IP connections and SLL processing are offloaded to the ADN platform, enabling the origin server's processing resources to be devoted to serving content. Scalability and performance is further reinforced through data compression to reduce the payload to be transferred across the internet, and the caching of static content in close proximity to end-users, by ADN servers.

A complete solution

By addressing the internet's root protocol problems, ADN services improve both the performance and scalability of dynamic applications and web services. By intelligently placing itself between the end-user and origin data centre, ADN delivers the benefits of application delivery controllers, WAN optimisation controllers and content delivery networks rolled into a single platform - with the added benefit of enterprise security.

ADN eradicates the cost and management complexity that's associated with data centre deployments, offering predictable capacity and LAN-like response times for global users. With no requirement for application or infrastructure modifications, ADN offers a low-risk, low-cost solution that enables enterprises to centralise their infrastructure and fully realise global delivery goals for its web and IP-enabled applications.

Malcolm Rowe is regional manager Northern Europe at Akamai Technologies.