Server testing means digging into benchmark suites. We'll expand this test suite as time goes on, but for this roundup of workhorse machines, we decided to stick with basic CPU and memory stress testing, coupled with a networked file system test to see how the server fared under varying degrees of disk loading over the network.

After examining several options, we settled on two benchmark test suites. The first is Geekbench 2.0.10 from Primate Labs, a Canadian software development firm. I like Geekbench because it does a thorough job of stressing both the disk and the memory subsystems. Also, it's available across Windows, Linux, OS X, and even Solaris, and it can run in either 32-bit or 64-bit mode, with specific support for both multiprocesser and multithreaded operation. Geekbench returns an overall Geekbench score compiled from all its tests. With only a few machines tested, this number is of limited importance, but it will grow in comparative value as I review more servers. Meanwhile, you can check out existing Geekbench scores on Primate Labs' results browser (a searchable record of submitted scoring results that's already several hundred machines thick).

In addition to the overall Geekbench score, I consider specific scores for CPU integer and floating-point performance, as well as for memory speed. Geekbench tests each of these subsystems with a variety of single- and (where applicable) multithreaded tests and returns an overall score in each category.

For disk tests, we decided to stick with the same benchmark suite we've been using for network attached storage (NAS) testing: the IOzone network file system benchmark (iozone.org ). Nineteen years in the making, IOzone is a sophisticated I/O benchmark test that evaluates an overall server solution from the application or file level rather than just from the raw hardware level. That means we configure our IOzone tests to run an I/O stream in a variety of file and block sizes. In this way we can see results in which the client and server caches play a role, as well as instances when files are so large that caches are overloaded, and we can get a sense of raw hardware performance.

I set IOzone to run files ranging in size from 32MB to as big as each server's RAM complement. Block sizes ranged from 4K to 128K. The traffic this produced included a range of small and medium-size files, forcing the server and its operating system to show their caching smarts. It also let me see what would happen when whopping data files stressed the servers at the bedrock hardware level.

We'll be tweaking our performance testing over time, especially the IOzone test, which we may run in a multinodal fashion in later reviews to simulate much larger user loads in stepped iterations. If you've got any requests or suggestions along these lines, feel free to drop me an e-mail, at oliver_rist@ziffdavis.com.

Benchmark Test Results
Check out the SMB Sever test scores.