Every once in a while, a major comparative review comes together at exactly the right time, where tests of actual shipping products reveal the impact of just-released technology in no uncertain terms. This review is one of those times. Blade servers from Dell, HP, and IBM -- all three sporting the latest Xeon 5600 (aka Westmere) CPU -- arrived at our test facility at the University of Hawaii before Intel had even officially introduced the chip. As it turns out, our benchmarks reveal that Westmere carries blades to new heights.
Although the start of the InfoWorld blade server shoot-out of 2010 was marked by a natural disaster and hampered by Murphy's Law, we persevered and spent two weeks beating up the best blade systems on the market to find their highs and lows. We built custom power monitoring loops, drank way too many lattes, and worked long into the Hawaiian nights as we ran the Dell, HP, and IBM blade solutions through a gauntlet of tests. They didn't flinch.
We also managed to add a budget-friendly Supermicro blade chassis into the mix. Where a Dell, HP, or IBM chassis with four blades starts upward of $40,000, the Supermicro solution costs a mere fraction of that. It is not in the same class as the big three blade systems, but it's an interesting option for shops that may not need the latest features or maximum performance. Since it was available in the lab, we put it through the same paces.
As with our previous blades shoot-out, all the testing was conducted at the University of Hawaii ANCL (Advanced Network Computing Lab) facilities on Oahu, facilitated by lab director, networking guru, and fellow InfoWorld contributor Brian Chee.
Blade Server Triathlon
The test plan this time around was quite different from the blade server shoot-out we ran in 2007. Gone are the HPC benchmarks, as we opted for a suite of custom VMware tests and an array of real-world performance metrics. The VMware tests were composed of a single large-scale custom LAMP application -- a load-balancer running Nginx, four Apache Web servers, and two MySQL servers -- laid out with sufficient vCPU and RAM resources to oversubscribe the physical and logical CPUs in each blade.
The workload was designed to mimic a real-world Web app usage model, with a weighted mix of static and dynamic content and randomized database updates, inserts, and deletes. This load was generated at specific concurrency levels, starting at 50 concurrent connections and ramping up to 200. The sweet spot on all the blades was the 150 concurrent connections mark, so we used those numbers for the published results. These VMware tests were run first on one, then across two blades. The blades were running VMware ESX 4 and controlled by a dedicated vCenter instance.
The other real-world tests were an array of common single-threaded tasks run simultaneously at levels that met and eclipsed the logical CPU count on each blade, running all the way up to an 8x oversubscription of physical cores. These tests included LAME MP3 conversions of 155MB WAV files, MP4-to-FLV video conversions of 155MB video files, gzip and bzip2 compression tests, and MD5 sum tests -- the same benchmarks we used last week to compare the new Intel Westmere-EP and prior generation Nehalem-EP CPUs. They represent a good cross-section of actual workloads and proved to be consistent markers of overall blade performance.
In addition to those tests, we rounded out the benchmarks by using Ixia's IxChariot tools to gauge network throughput between blades in the same chassis. Those tests were based on a few common loads, including iSCSI 32K and 256K read tests, a packet blaster load, and a BitTorrent P2P load that gauged connection rates. We also used Ixia's IxLoad testing tool to generate load on the LAMP application for preliminary testing, prior to using the benchmark's built-in client.