Wednesday, November 3, 2010

Latency and Data Center Siting

Today, there is much emphasis on the ability of cloud services to be located in mammoth data centers ideally suited to maximize low cost power and climatic or environmentally favorable cooling methodologies.

So, why don't we move everything to such facilities? There are a number of reasons:
  • Companies whose system require care and feeding by a staff in a specific geographic location.
  • Privacy laws that require the data be kept within certain boundaries.
  • Disaster recovery and accessibility concerns.
  • Physical latency which is the time that it takes to communicate with the facilities.
Today, I am going to address the issue of latency, which at its core is limited by physical laws of nature.

Latency is the amount of time, generally measured in milliseconds (ms) which represent 1/1000 of a second. It is generally expressed as "round trip latency", or the time it takes the information to make it from point a to point b and back to point a. This is in part because of the traditional tools (i.e., "ping") used to measure latency and also to simulate the effective time to make and confirm a transaction.

Round trip latency between New York and London is about 100 ms, and New York to Los Angeles is about 70 ms.

This delay is important in that it impacts the performance and business aspects of many transactions. Program trading operations, for example, want to minimize latency as much as possible and, as such, want to locate within 20 miles of the "trading" center and have latency that is below 2 ms. Many other applications perform best with latencies that are less than 20 ms, implying a distance that is less than 700 miles.

The tolerance of applications is increasing and many have become virtually latency insensitive. It is this class of applications that can easily be moved to a data center optimally located for power and cooling.

The more sophisticated IT departments now segregate their applications by latency, as well as other issues. This allows them to group what must be done at local vs remote data centers, and may include several different gradations along the way - such as one data center within 20 miles, one within 250 miles, one within 500 miles, etc.

1 comment: