CloudSleuth, a resource from application performance player, Compuware, has released the results of a year-long study of the fastest response times among a number of major cloud service providers.
The independent tests, which were carried out using 30 testing nodes around the world to monitor performance once every fifteen minutes. To save you the calculator-pull, that is 515,000 tests overall for a year from August 2010 to July 2010. Each of these tests involved loading a simulated retail shopping site of two pages, one page loading 40 item descriptions and small JPEG images, the second pulling open a sample image of 1.75 MB.
The winner, in a rather hands-down sort of way, was Microsoft Azure, which beat out competitors Amazon EC2, Google’s App Engine, Rackspace and a handful of others.
As ArsTechnica noted in a recent analysis of the test:
“The Windows Azure data center in Chicago completed the test in an average time of 6,072 milliseconds (a little over six seconds), compared to 6.45 seconds for second-place Google App Engine. Both improved steadily throughout the year, with Azure dipping to 5.52 seconds in July and Google to 5.97 seconds. Also scoring below 7 seconds for the whole year were the Virginia locations of OpSource and GoGrid along with BlueLock in Indiana. Rackspace in Texas posted an average time of 7.19 seconds, while Amazon EC2 in Virginia posted a nearly identical 7.20. Amazon’s California location scored 8.11 seconds on average.”
As Jon Brodkin stated, however, the tests have a number of potential weak points. For instance, as he wrote today, “Although Compuware tries to make the tests expansive by spreading nodes throughout the world, the results are still highly affected by location. For example, both Azure and Amazon posted poor scores in their Singapore data centers (16.10 seconds for Azure and 20.96 milliseconds for Amazon, the worst time in the survey) but the discrepancies between North America and Asia are due in large part to limitations in the Compuware testing network.”
Brodin went on to note that “Within Asia, the performance is generally abysmal by North American standards,” says CloudSleuth product manager Lloyd Bloom. But the measurements are skewed because “most of our measurement points are not in Asia.”
While the cloud needs for a retail shopping site versus an HPC application are about as closely tied as night and day in many cases—especially given the relatively small compute and storage demands involved here—it nonetheless points to general speed for projects like testing and development, an area that many HPC users attempt to outsource to free up their clusters for bigger crunching and much higher dependence on high performance, low-latency networks.