• 
• 

 1 of 2

Gallery

Cool servers are coming

Posted on 18 Aug 2009 at 10:33

Jon Honeyball ponders the next generation of Intel servers and Hummers with supercomputers in the back

We're finally starting to see some real innovation in the server hardware space. It's been obvious for some time that we need a new generation of low power consumption, high storage density servers, aimed at handling normal but not outrageous workloads, and capable of doing so while putting out the minimum amount of waste heat.

The Core 2 processing platform used for laptops is a good starting point, but it could be argued that these chips aren't feature-rich enough for the server environment that will be needed in 2010. The Intel Atom CPU would be even more ideal, except that it doesn't support virtualisation, and VMs are an absolute must in the 2010 timeframe, even if there's a 1:1 relationship between processor cores and VM.

A few people have been trying the approach of stacking devices such as Apple Mac minis into rack space, which can provide a solution even if it's really the wrong tool for the job. So I was most interested to see Dell's announcement regarding fitting multiple servers into a small rack space. Called the DCS (Data Center Solutions), it's built around a high-density processing module that's a complete computer on a slide-in board, based around the VIA chipset, which both supports 64-bit processing and virtualisation too. The rack case is just 2U in height (3.5in) and can take 12 such units, but its killer feature is its power consumption of below 30W per unit.

Next-generation Intel servers

At the recent Mix09 conference in Las Vegas, I saw some very interesting technology coming down the line at the opposite end of the power scale - the next generation of Intel servers. At the moment, there are architectural limits on the power supply rails that can be routed to each CPU chip in a server, but the future direction is to enable you not only to drop the voltage to each individual chip but also to each individual core within each chip, and to adjust each individual core's clock frequency.

A few people have been trying the approach of stacking devices such as Apple Mac minis into rack space, which can provide a solution even if it's really the wrong tool for the job

This would allow you to dynamically scale your power consumption and headroom so that you always have just enough processing power to fulfil the current tasks, but can immediately scale it up within the server to handle new workloads as they arrive. When coupled with the sort of intelligent operating system facilities that we're seeing from software such as System Center, it's clear that you ought to be able to do some very interesting and aggressive power-management in the future.

For example, if your London office is unmanned at night, why not move all its VMs onto a low-power server and shut down all the unused servers completely? Start them up again in the morning and move the VMs off the slow "overnight" server back into the live workload environment. It's clear that there will be big benefits from scaling the amount of power consumption to the necessary load, and ensuring that nothing is wasted.

Another scenario is the one that covers power failure and system shutdown. If your mains power has failed and you've moved over to a battery-powered UPS, it would make sense to shift all your server loads onto slower chipsets to allow a more orderly shutdown. And to be able to start to squeeze cycles down on all the cores to use the least power possible. And to recognise when you've moved on again from the battery backup onto the emergency generator.