As the price of oil looks set to push through the $150 per barrel threshold, firms running large datacentres are facing increasingly onerous energy costs. As companies try to scale their datacentres out to increase processing capacity and network connectivity, the power consumption problem becomes more acute, while even tougher problems lay in wait for firms planning new datacentre infrastructure.

One firm that believes it has an answer to these woes is Blade Network Technologies (BNT). The vendor has developed switches that sit inside datacentre racks alongside servers and storage, thus obviating the need to run network connections through power-hungry, chassis-based core switches.

BNT’s chief executive Vikram Mehta has 20 years experience in the network switch sector, and has led the company since it was spun out from Nortel in 2005.

Mehta’s view on the traditional model for building out datacentres is unequivocal: “It’s broken, and if the same old strategy is continued, firms will have massive problems,” he said. “You can’t have an exponential increase in your carbon footprint, personnel and datacentre floor space.”

The reason for the big increase in demands being placed on datacentres is a combination of new multicore processors, a move to server and storage virtualisation, and the new breed of Web 2.0 applications that have a high input/output (I/O) requirement. This in turn has led to a big increase in demand for network bandwidth. “When I take CIOs and IT managers through a server rollout in their datacentre, the first thing the IT manager says is, ‘I’m out of power, and I’m being expected to do a lot more with a lot less’,” said Mehta.

A typical server rollout can involve initial network connectivity using anything between four and six Ethernet network interface cards (NICs), according to Mehta. After this there could also be host bus adaptors (HBAs), Fibre Chan nel (FC) fabric switching, and if the datacentre manager is using InfiniBand host channel adapters (HCAs) for server interconnects, servers might end up with 10 network interfaces. “If we’re looking at a 10,000-server environment, then you could have around 100,000 cables to deal with,” he said.

Reducing network costs
Mehta estimated that datacentre network infrastructure amounts to around 20 per cent of the total costs, and that firms should be able to reduce this significantly through virtualising network connections. The problem with virtualisation, however, is that it is not network-aware. “So switches need to talk the language of virtual machines (VMs),” said Mehta.

Because switches are not “VM aware”, network policies have to be applied to the port, and BNT’s recently launched system, SmartConnect, is designed to address just this problem. “SmartConnect says you can put network switches anywhere and create logical groups that can operate over different physical ports, allowing you to move VMs anywhere ­ and not just from one side of the datacentre to another,” said Mehta.

BNT’s SmartConnect enables firms to thread multiple hypervisors together to create what Mehta called a “megavisor”. Mehta also said that while BNT’s 10 Gigabit Ethernet (GbE) switch hardware has a slight performance disadvantage against InfiniBand, it uses 60 per cent less power.

A system like this simplifies not just deployment, but also control and policy administration, which in turn can lead to a significant reduction in the total cost of ownership. The answer here is network virtualisation. “Think about it: if you have Ethernet, FC and InfiniBand connectivity, you have three types of network administration to deal with, so the obvious plan would be to simplify your I/O infrastructure,” Mehta said.

BNT recently launched hardware and software designed to do this in an effort to help firms in their quest to tame datacentre costs. The RackSwitch GS8000 is a 44-port gigabit copper aggregation switch with four GbE uplinks, while the RackSwitch GS8100 is a 24-port 10Gigabit Ethernet (10GbE) switch specifically designed to reside in the rack alongside the blade servers it is providing connections for.

Current datacentre configurations have server and storage network connections going through power-hungry core switches, which need additional cooling, said Mehta, but the RackSwitch kit fits into the same rack and has similar cooling characteristics. “Air flows in the same direction as that required to cool server and storage hardware ­ from back to front. Traditional switches with rearward facing ports pull in air from the front and pump it out at the back. This creates ‘hot loops’ and leads to inefficient cooling,” he said.

This strategic solution is something Mehta refers to as Rackonomics, whereby the 42U rack becomes the basic unit of provisioning, so that datacentres add extra capacity by putting in another rack full of servers, storage and switches. “First, look to converge your network technologies. Stop thinking about deploying compute capacity one server at a time, think of deploying capacity a rack at a time. Then, address network virtualisation, because you can’t address server and storage virtualisation otherwise ­ that’s Rackonomics,” he said.