For proper performance, data centers have to match the rack cooling system's abilities to server density, or vice versa.
If you graph data center energy use over time, you would see an upward-trending sawtooth curve. Higher computing demands mean more servers, pushing energy demand up until a new generation of more energy-efficient servers is introduced, which drops energy use down until more demand starts pushing the energy demands up again.
We're refreshing some servers and wondering about the current or optimal density of server racks. Has rack density changed?
Today, the power density for fully loaded server racks is 8 kW to 10 kW, with blade systems in the 12 kW to 16 kW range. Aging data centers and server rooms may have difficulty cooling these rack densities, in which case you should run some racks less than full and plan to incorporate new cooling infrastructures in the future.
Server rack energy density has increased from the early days of roughly 1 kW or 2 kW of power per server rack. These servers were often single-processor systems in relatively large (2U or 4U) rack enclosures. Servers evolved quickly, and the combination of multi-processor and multi-core systems, higher levels of integration with system components and the shift of storage from local disks to centralized storage-area networks or network-attached storage allowed for smaller, 1U enclosures. This pushed average rack densities into the 5 kW to 7 kW range, with some as high as 10 kW.
Although experts and industry pundits predicted rack densities of 20 kW to 25 kW with the introduction of high-density blade systems, these deployments never reached mainstream adoption. Such rack densities produce hot spots that are extremely difficult to cool with computer room air conditioners and other standard data center cooling technologies. At higher rack densities, options include direct cooling, as with liquid cooling options and specialized racks.
Today's servers do a lot more work with lower energy draw than previous generations. As processor vendors invest in features -- smaller chip fabrication sizes, clock throttling -- that lower the servers' power profiles and thermal envelopes, server vendors are building highly integrated system-on-a-chip platforms and other energy-efficiency technologies into new server platforms. As compute capacity grows, cooling can stay about the same.
Virtualization has had almost no effect on rack power densities. A rack full of virtualized servers can run many more workloads than the same rack of non-virtualized servers, but both racks use the same amount of energy to operate the servers. Virtualized data centers will need fewer racks for the same workloads, lowering overall energy demand.
When a new data center build is simply not in the budget, rely on judicious data center renovations that support your technology upgrades, preserving the operational integrity of current systems and workloads.
Dig Deeper on Data center design and facilities
Related Q&A from Stephen J. Bigelow
Regression tests and UAT ensure software quality and both require a sizeable investment. Learn when and how to perform each one, and some tips to get... Continue Reading
Learn the meaning of functional vs. nonfunctional requirements in software engineering, with helpful examples. Then, see how to write both and build ... Continue Reading
Just because software passes functional tests doesn't mean it works. Dig into stress, load, endurance and other performance tests, and their ... Continue Reading