Underutilized servers continue to plague the datacenter, taking up valuable space, cranking out heat, and devouring kilowatts while doing little to no work in return. (Replace "kilowatts" with "kibble," and the previous description can be applied to my parents' Garfield-esque cat Marmalade.)
Fortunately, advances in server and server-management technology are making it easier for datacenter operators to not only determine precisely how much and what type of work (if any) a server is doing but also to reduce the amount of power a server consumes when it's working or standing by for a new workload. Moreover, thanks to innovations in virtualization and cloud computing, the day may soon come when the prospect of powering down servers between uses won't send chills of terror through the spine of the IT admin.
First, let's recap the significance and costliness of underutilized servers. A study by McKinsey & Company titled "Revolutionizing Data Center Efficiency" found that among a total of 458 servers at four production datacenters, 32 percent (146 in all) were running at or below 3 percent utilization. Meanwhile, analysts estimate that more than 15 percent of the servers at an average company run 24/7 without being used actively on a daily basis. On top of that, research has found that servers burn more than half of their rated power while idling -- which in turn amps up the need for additional cooling, burning further power. Depending on the size of your datacenter, underutilized servers represent potentially thousands, tens of thousands, or even hundreds of thousands of dollars wasted each month on power and cooling.
Servers: Prove your worth
One of the more interesting solutions to curbing server waste comes from 1E. The company recently released a Server Edition of NightWatchman, 1E's popular PC-power management product. Using a complex algorithm, NightWatchman SE employs agents to probe and determine whether a server is drawing energy to perform useful work -- for example, crunching data and fulfilling end-user requests -- or wasting energy on non-useful work, such as self-maintenance tasks that can be put on hold.
Using that data, the software can generate detailed analysis to show datacenter operators how much power a given server is consuming to perform what tasks. Additionally, admins can see what level of work a server is doing at any given time of day or night. The offering can report not only on energy consumption, but also cost, efficiency and CO2 emissions for all servers or group by location, department, application, and more. In short, you have a detailed view of precisely how productive (or unproductive) the servers are on a 24/7 basis.
According to 1E, the solution is available for all types of server platforms. Interestingly, it can track the performance of not only individual machines but also of virtual machines. This will gain increasing importance as virtual server sprawl becomes as problematic as physical server sprawl.
Armed with that information, an admin can make informed decisions about what to do with a given server. Machines achieving zero useful work can be decommissioned. The workloads of underutilized machines can be combined intelligently; for example, tasks that run during off-peak hours could be moved to a server doing productive work during peak hours only.
You are getting drowsy
Finally -- and this is especially interesting to me -- an IT admin can use NightWatchman SE to automatically put a machine into Drowsy Server mode (a form of dynamic power management) when it's not performing useful work.
Note that 1E makes an effort here to emphasize that the server isn't being powered down completely. 1E CEO Sumir Karayi stressed the point as well during my briefing about the product. The reason is understandable: The typical IT admin is very wary of powering down a server completely, for fear that it won't come back up quickly enough, if at all. "Powering down is not a method that goes down well with datacenter managers," he said, "They are really obsessed about service levels."
Notably, admins can set NightWatchman Server Edition to power machines off completely when they aren't doing anything useful, but Karayi said this application is for very special niche cases, such as test and development machines.
What's unfortunate here is the fact that Drowsy Server yields just a 12 percent reduction in power consumption. Perhaps that's nothing to sneeze at; every watt saved translates to money saved and carbon emissions reduced. But that 12 percent figure pales in comparison to the potential savings from putting a machine into a deeper sleep mode. As a point of comparison, a PC that's up and running might use 70 to 100 watts but only 5 watts in standby. That's a power savings of more than 90 percent.
This all leads to a bigger question: Will system administrators ever be prepared to allow their servers to take an occasional deep snooze when they're not doing anything useful? A company called Cassatt attempted to entice datacenter operators to power machines off and on as needed using its Active Response package -- but it didn't gain traction.
Meanwhile, other organizations are dabbling in powering down servers. Cisco, for example, reports that it has developed tools in-house for powering non-mission-critical servers off and on, such as those for testing and development, as well as applications such as payroll, which are run on a regular schedule but not needed daily.
But the day may come when powering on and off production servers is no longer taboo, thanks to cloud computing and virtualization, according to Douglas Alger, IT architect for physical infrastructure at Cisco and author of "Grow a Greener Data Center: A guide to building and operating energy-efficient, ecologically sensitive IT and Facilities infrastructure."
Alger says, "We'll reach the point where the datacenter consists of large virtualized resource pools, and any single box isn't particularly important because it's just one of 100, or 200, or 500 [running a given service]. It seems we'll be able to power machines down when they're idle without fear that they won't come back up. Even if most come back up, you'll still be able to work with that pool."
A dynamic virtualized datacenter? Sounds promising.