Cloud computing: How to avoid a network bottleneck

Daniel West, IT director at online fashion retailer Asos, is only too aware of network resilience. The company's global distribution centre is located in the UK and is served by two networks running in opposite directions. He says the site also uses a line of sight wireless link, should the fixed line networks fail. When a business relies 100% on network access, any failure or slowdown will have a material impact.

Cloud computing can put a huge load on a business's network. Organisations need to ensure network access from the user's browser on the local area network (LAN), to a cloud-based datacentre across on the public internet are fit for purpose.

Computer Weekly blogger, Steve Broadhead, founder and director of Broadband Testing, says cloud issues are based around performance and ownership. "Latency is a potential problem here - how many hops is your data travelling across to reach its host/destination? Then there is the actual "ownership" of the cloud element. If it is not a private cloud, who is responsible for that part of the network, who owns it? What happens if there's a problem - who do you point the finger at in the event of a performance issue?"

Broadhead warns another potential problem is at the back-end system. "If the servers are virtualised (which they probably are) there are further potential performance limitations, depending on the capabilities of the servers themselves, the design of the back-end and what bandwidth is available to each virtual server."

Class of service

Richard Thomas is CEO of NetEvidence, a ten year old company which offers a monitoring and service assurance service. NetEvidence sells a tool called Highlight, which tells a business what is happening on their networks and the applications they run and the services they buy. "When preparing a network for the cloud, you have to ensure the network is capable of supporting what you want to do. In cloud computing your area is a lot more sensitive to network performance," he says. "The obvious problem is bandwidth, but different types of services needs different levels of performance. Voice requires a different type of traffic, there is a latency issue and jitter. Consistency of network performance is important."

People put a lot of focus on the LAN and their connections to their other offices. He said: "On the WAN, we find in an alarming number of cases, an organisation can specify classes of services to determine priorities of service, ie how much high priority traffic. These pipes are set up but no one has put traffic in the right class of service. If you have class of service you need to ensure it is actually working."

Cloud performance impacted by WAN

Network performance is affected by bandwidth and the length of the journey between the user's browser and the cloud service. Compuware has set up a community called CloudSleuth, to help organisations get an idea of network latency and performance issues when building cloud applications.

"We have a reference application crafted to look like a static catalogue site - a simple but common-use case," says Doug Willoughby, director of cloud strategy at Compuware. This application is deployed in 17 cloud providers using common virtual machines, based on TomCat application server and Linux.

The Google Application Engine and Microsoft Azure Compuware uses the Google and Microsoft application server infrastructure. CloudSleuth works by running simultaneous transactions from 32 locations around the world and measures the response time of the application hosted at the 17 cloud providers.

Doug Willoughby says: "For our reference application you'd like to see a response time of less than three seconds." Looking at 30 days' worth of measurements, he says the fastest for a user in London were the services that used European datacentres. "CloudSigma is doing a great job and has put a lot of money in its infrastructure. EC2 Europe is based in Ireland. Where there is a large hop between the browser and the application, there will be a material impact on speed of the network."

Cloud service providers Amazon and Rackspace use the Akamai content delivery network to reduce network bottlenecks. The CDN acts as a smart cache, says Bruno Hourdel, product line director at Akamai. The Akamai content delivery network is based on 85,000 servers installed in 950 ISPs. "We have been moving video and distributing applications for 12 years. We developed Sure Route, a protocol optimisation technology and we added HTML file inspection, to prefetch objects in an HTML file such as images."

In the cloud, everything is dynamic. Akamai uses web optimisation to optimise the HTML file such as making Javascript run on the server, to reduce round trips between the browser on a user's PC and the server. A dynamic application comprises HTML, Javascript and images. Akamai can prefetch the objects. He says that Akamai aims to deliver content within two seconds, wherever in the world the user is located.

Tier1Reserach has highlighted another example of where Akamai is being deployed for cloud computing. This time, the service is Force.com, from Salesforce.com. Its Force.com Sites service is enabling customers to rapidly create and launch applications, and Akamai's web application acceleration services are available as an option to customers wanting enhanced performance.

Desktop bottleneck

IT managers looking at desktop services in the cloud will need to consider the virtual bandwidth available to users. A desktop PC may be connected to the LAN via a 1Gbps physical network connection. As Computer Weekly has previously reported, giving every user in a virtual desktop infrastructure deployment 1Gbps will quickly overload servers. The network adapter cards on the servers will quickly be swamped. Paul Robinson, consultant and business development director at 2X Software, warns people often fail to consider how many connections there are between PCs and servers: "Too many connections cause a 'time out' when too many people jump on the network."

A network structure must be designed to accommodate this. If the server has two 1Gbps network interface cards, it can only handle bandwidth of up to 2Gbps, so it only takes a few virtual desktops to saturate the network, especially if each virtual PC expects a 1Gbps connection.

Cloud computing clearly offers businesses a low-cost way to run applications. But as users become dependent on those applications, it is critical the networks that connect the user's browser to the cloud service provider, is resilient. Fashion retailer uses triple redundancy for the network connectivity to its UK-based global distribution centre, which is regarded as business critical. Such resilience does not come cheap and organisations using cloud services will need to assess how much network resilience they need, or how much downtime they are prepared to suffer.

In the past IT managers would have used wide area network optimisation tools and load balancers to ensure internal applications ran optimally across the corporate network. Cloud computing means the CIO has less control over what route network traffic takes into the cloud provider's datacentre. As the reference application from CloudSleuth shows, the distance and the number of network jumps between the a user's web browser and the cloud application, can seriously impact performance. Furthermore, with cloud application platforms like Microsoft Azure and Google Application Engine, it may not be possible to determine where the service is physically located. Longer network hops will lead to latency.

As NetEvidence's Thomas points out, not only should CIOs establish classes of service for their applications - such as giving VoIP a higher priority over email - they should also ensure that these classes of service are maintained, otherwise cloud applications may simply become unusable due to huge network latencies.

Finally, a content delivery network can make a huge difference to the performance of a cloud application, according to data from CloudSleuth. This basically moves the application closer to the end users, which overcomes the latency issues that occur when running cloud services that may span thousands of miles and many network hops.

Case Study Box

John Guest, which manufactures plumbing fittings, brass fittings, valves, manifolds, pipes, hoses and heating systems, is using a network monitoring service from NetEvidence to monitor a new MPLS-based VPN. The MPLS is being used for videoconferencing and up email between the firm's head office in West Drayton, Middlesex and its major sales/distribution offices in the US, France, Germany, Italy and Spain.

Howard Wheatley, IT Manager at John Guest, said: "Highlight enables us to see clearly what is running over the network. It means we are not reliant on our Service Providers to tell us if there is a problem with a link - we have full visibility of how the network and the applications are performing and can resolve minor issues before they become major problems. In fact, within the first few days of using Highlight, it identified an open video link between the UK and US that had been left open for over a week, using around 40% of our bandwidth."

In addition to the new MPLS, John Guest is planning to use Highlight across its whole network. Howard adds: "We are now planning to use it to monitor our leased internet line which is delivered by a second Service Provider and we will also keep a watchful eye over some additional ADSL lines across the UK."