In a network it is often a good idea to know where server and storage are located. This way, the network path can be optimised to minimise the number of jumps (ie the number of network switches and routers that IP traffic must pass through) between the various server and storage components. A more direct path between server and storage components in the datacentre leads to better performance.
But to get the most from server and storage virtualisation, the physical location of a virtual machine should not be tied down, as this would affect the flexibility of the virtual infrastructure. As a result, a traditional Ethernet network cannot easily be optimised for virtualisation and cloud computing.
A poll conducted at the 2010 Gartner Data Center Conference (held in December 2010) found that 83% of respondents were using mobility to reassign new locations or shift workloads, or were using policy-based software rules for optimisation.
Computing fabric for virtual machines
Gartner believes an approach called "computing fabric" will be required to support the dynamic allocation of virtual machines, where the network, server and storage act as a single unit connected using a switch. "Cloud computing and virtualisation make networking difficult. Modern blade platforms, such as HP Virtual Connect and Cisco UCS [unified computing system], are integrated with switches [to simplify networking]," said Andy Butler, a distinguished analyst at Gartner. "Fabric computing relies on the network switch being integrated with the server, the network and the storage."
According to Brocade, Ethernet networks are not designed for cloud computing. While network managers have previously been able to optimise networks by managing performance at the network's core, Marcus Jewell, regional sales director at Brocade, says virtualisation means network traffic becomes unpredictable.
Duncan Hughes, systems engineer at Brocade, said: "For the past 20 years we have been using a three-layer hierarchical layer comprising the access layer, aggregation layer and the core layer. Routing would only be performed at the core layer, so if you needed to communicate with a server on a different part of the network [subnet] you would need to travel the network up through the three layers [three network boxes] and back down again." This is inefficient and does not copy well when used in a virtualised environment.
Hughes says that if this approach to networking is used in a virtualised environment, network traffic between virtual machines will bounce up and down the network.
Brocade and other network equipment makers are now selling the idea of a network fabric, which overcomes the network problems caused by virtualisation. The storage networking company has developed what it calls a flat-layer, two-network, self-healing fabric, which it says overcomes the limitation of using a three-layer network topology.
Case study: Ethernet fabric boosts performance for De Persgroep
The Brocade network infrastructure is helping Belgian media company De Persgroep manage its datacentre networking. The company, which has grown rapidly through acquisition, has seen the number of servers in its datacentre grow from 300 in 2007 to around 1,000 virtual servers today. Along with office automation, the main applications are editorial and advertising systems, plus its websites, some of which support two million users at peak times, with bandwidth of 2Gbps.
De Persgroep's previous network was experiencing capacity issues due to the growth in virtual servers, and limitations on the number of physical network ports was affecting network performance. Throughput bottlenecks meant that back-up schedules would fail, and overall switch performance and stability was no longer acceptable. Following a competitive tender, the company selected the Brocade VDX 6720 as the foundation to build an Ethernet fabric for its datacentre.
A total of 72 Brocade VDX 6720 switches have been deployed across the fabric-based architecture, creating a single logical chassis with a single distributed control plane across multiple racks of servers. This design has provided compelling reductions in capital and operating costs, while simplifying virtual machine (VM) migration. The deployment is already delivering the desired performance and resilience for De Persgroep.
The Ethernet fabric uses dual 24-port network switches fitted to the top of its server racks to enable each rack to operate its own dual redundant network. Servers in the racks are equipped with two network cards each to support the dual redundancy. The racks are connected to aggregator switches which bring the networks from the servers together and connect them to another datacentre, which is located a few hundred metres away. This provides a so-called active-active system, where both datacentres are operational.
"On our old system you could lose several seconds if a switch failed," said Wim Vanhoof, ICT infrastructure manager at De Persgroep. Such a delay would be enough to cause an application to crash. "On the new fabric, which uses a virtual switch configuration, fail-over is instant," he added.
Another benefit of the approach De Persgroep has taken is that its datacentres no longer require complex network wiring of a massive switch to handle all the connections.