NetXen claims to be able to virtualise a network interface within the network card itself, rather than relying on the hypervisor to do it.

NetSlice gives each virtual machine (VM) a virtual network card of its own with the result that it saves on processing power and allows more data-intensive applications to run on virtual systems, the company said.

Companies typically use VMs today for relatively lightweight apps, such as webservers, but the next phase of virtualisation will be used to run applications such as data warehousing, business intelligence or maybe Microsoft Exchange, said Vik Karvat, NetXen's senior marketing director.

"Looking forward, we will soon have quad-CPU servers with quad-core processors, that's 16 cores so we will see heavier apps and many more VMs per machine - and those are going to start pushing a lot of I/O," he said. "Even if each is only a few hundred megabytes, it soon adds up to gigabytes, and there's a finite limit to how much you can get through the hypervisor."

At the moment, all I/O is routed from the VM's network driver to a virtual network within the hypervisor, which controls the physical network interface. "That needs a high number of cycles and a lot of software touch, and because all the packets are routed through the hypervisor, it adds a limit on the bandwidth available to the VM," Karvat said.

The current planned fix for this is to build network adapters that are virtualisation-aware, and the PCI-special interest group (SIG) is developing standards for this. "We have been working with PCI-SIG on NIC
(network interface card) virtualisation, but standards take time," said Karvat - so the company has decided to introduce its own version now. "Everything we're doing ties in with the SIG work though," he added.

NetSlice isn't aimed at supporting multiple physical servers, like the Neterion approach which allows blade servers to virtualise and share a NIC. Instead it allows virtual servers on the same physical system to have one or more virtual instances of a NIC each.

NetXen gives each VM its own physical NIC driver, allowing the guest operating system to bypass the hypervisor and drop packets directly to the hardware. This requires an intelligent NIC capable of memory-mapping in hardware. NetXen will provide a firmware upgrade to add NetSlice support, Karvat said. "We support 1024 virtual NICs. Our view is yes, you do need that many, because the VMs could have multiple traffic types - such as data, management, iSCSI - and each traffic type will have its own priorities and so on, so it will want its own NIC.

"Now you can for example load Microsoft's TCP Chimney [TCP offload] stack at the guest O/S level and do a complete protocol stack bypass too," he said.

Karvat noted that NetSlice still requires some intervention from the hypervisor to control DMA remapping, so the hypervisor needs software to support the NetXen card. "We have delivered our first firmware to the virtualisation vendors and they are tinkering with it," he added. "It's up to them to deliver support for it - but pretty much everything we're doing, they'll have to support for the SIG anyhow."