Building a new data center? Think WAN
- 17 July, 2008 09:50
Practically every company I talk with is consolidating data centers, constructing new ones, or both. These aren't the old "glass house" models of the 1980s and 1990s: They're next-generation designs with racks of blade servers, virtualized clusters and storage-area networks.
What's the hardest part of engineering such a new data center? Often, it's managing power and HVAC requirements. Blade servers can suck up power at a density that's beyond the capacity of many facilities -- as much as 30Kwatt/rack, once you factor in cooling.
Another challenge is deciding where to put the data center. Many IT organizations are considering hosting or outsourcing, which limits the range of physical locations (you can only put the data center where the providers have facilities) but raises the question of which hosting providers to consider.
Whether hosted or not, though, it's important to consider the data center's WAN architecture. Here are a few thoughts to consider:
Bandwidth. With server consolidation, an increasing percentage of the company's traffic is traveling to and from the data center. Companies are reporting a significant uptick in bandwidth requirements across the board -- the typical branch office has a WAN link between T1 (1.544 Mbps) and fractional T3 (up to 45 Mbps). Translating branch office bandwidth to data center WAN requirements can be tricky, but it's safe to say that most companies are looking at T3 to OC-3 speeds, at minimum. Moreover, companies say their bandwidth requirements are growing between 50% and 100% year over year (the median growth reported in a recent Nemertes Research study was 99%) -- so don't assume that today's bandwidths will suffice for tomorrow.
Latency. Consolidated data centers means the company is taking servers that used to be down the hall from users and putting them hundreds (or thousands) of miles away. At those distances, latency can really hurt application performance: many applications assume that network latency will be on the order of a couple of milliseconds, rather than the 150 ms that's standard for a WAN link. This is particularly important for services such as VoIP, which is highly latency-sensitive, so if your data centers contain VoIP servers, you must consider latency. Ask for -- and confirm -- service-level agreeements containing latency from your provider. And where necessary, build a network of data centers to ensure that data centers are close enough to users (a strategy called "regionalization". )
Service type. Although MPLS is the standard WAN service (73% of companies I work with say they've deployed it), organizations are looking at carrier Ethernet (VPLS) or straight-up optical when it comes to data center WANs. Why? See above -- it's easier (and less expensive) to get the ultra-high bandwidths and low latencies from these services.
Redundancy. With consolidated data centers, your eggs truly are in one basket. So as Mark Twain once said, make sure you watch that basket: ensure that you've got redundant power as well as carrier connectivity. Losing either can take down your data center, even if the equipment's fine.
Bottom line: When architecting data centers, consider the WAN.