RAID Theory

Need for speed
Because we can’t fill the entire mag with RAID benchmark results (oh I want to, but Dave won’t let me! (there, there, pet, there, there –ed)) we’ll break these down into a number of categories, each focusing on the topics we’ve discussed.

For our tests we used an E8500-based machine on a 780i motherboard that comes with NVIDIA’s MediaShield FakeRAID controller. All of the tests were on the NVIDIA controller, unless otherwise indicated. For storage we used a mixture of 10,000 RPM Western Digital Raptor 150 and Seagate’s latest 7200.11 32M cache 500G drives – all separate from an OS boot drive. All tests were conducted after cold boots and on empty drives and arrays, and so represent the fastest outer tracks.

Although we benchmarked the drives using PCMark05, SiSoft Sandra, ATTO, HDTach and Crysis level loads, the most pertinent benchmarks are PCMark05 and Crysis, since these show off the type of workload you’re likely to see using your PC for working, web surfing, multimedia, and gaming (don’t expect soaring scores here, as real world use looks a lot different to raw throughput benchmarks). The SiSoft Sandra and ATTO results were good for showing read/write ceilings, and they generally reflected the individual results seen in PCMark05. HDTach results were good for showing maximum throughput ceilings.

If you’ve used PCMark05 you’ll know it breaks down the hard drive performance benchmarks into a number of usage categories – perfect for our testing to see how RAID arrays – and especially RAID stripe – affect different workloads.

Three of the tests focus on random access and seek-based workloads (OS Loading, Application Loading, and General Use) while the final two focus on throughput (Virus Scanning and File Write). To keep things simple in the graphs we’ve averaged these results into the categories they represent – seek and throughput. Don’t be put off by the seemingly low transfer speeds for the seek workload based results – PCMark05 is deliberately demanding, and this helps separate the men from the girly-men as it were.

Finally, the Crysis level load time (based on the GPU_Benchmark level) is a good real world test of array performance for a level-loading, gaming workload as distinct from operating system, application and throughput workloads in PCMark05.

First up, let’s establish baseline performance – single drive Raptor 150 and Seagate 7200.11.



As expected the Raptor has a good lead on the access-based workloads, showing off the advantage of its 10,000 rpm spindle speed. At the same time the Seagate 7200.11 has some serious throughput power, pumping over 100MB/s all on its own and outperforming the Raptor. Now let’s answer the question of which is more important to your everyday workload and gaming, seek or throughput? Let’s RAID the Seagate drives and see.



As we can see, even though two of the beefy Seagate 7200.11 drives in RAID provide an even higher throughput score, they still fall behind a single Raptor for everyday use workloads, and they actually load the Crysis level a little slower. Not anything you’d notice, but it’s certainly not as fast as you might expect a RAID 0 array to provide. This is why it’s important to build an array to your workload, or you might not gain much at all.

On that note this graph also shows the importance of choosing a good stripe size – the ‘balanced’ 64k stripe here isn’t helping the Seagates, and as we’ll cover in a moment a smaller stripe would yield better performance for throughput. Still, you can take this away: for general everyday workloads a single 150G 10,000 RPM Raptor is faster than two 500G 7200 RPM Seagates in RAID 0. Even though the Seagate’s density and large cache outperform the Raptor for sheer throughput, the Raptor’s RPM gives it the edge for general PC use.

So what happens if we RAID the Raptor, combining its fast seek capabilities with higher throughput in RAID?



Nice, really nice. We see the RAID Raptors pulling away great scores for the seek-based workload, and the two-drive Seagate RAID still outperforming three Raptors in the throughput-based workload. But now we see something interesting – firstly note how the seek-based strength of the Raptors provides a slightly faster level load over the Seagates, indicating random access is an important factor for the level loads. Also note how the level takes longer to load on a three drive array compared to two drive. Let’s see how different stripes perform for two and three drive Raptor arrays.



Here we see the exact impact stripe size has on seek and throughput-based workloads – as the stripe size increases the seek performance increases while the throughput decreases. In the throughput scores we can see a sweet spot – 16k and 32k stripes provide a much higher throughput compared to 64k and 128k. The drop isn’t linear, and if you look closely you’ll see it’s inversely represented in the seek scores. This gap could be a function of the NVIDIA controller and drivers, but also of the specific files and file size workload PCMark05 is performing here. Which is why it’s handy too look at the Crysis result as well – as in the previous graph we see two drives performing better than three at the large stripe sizes, but the reverse is true for the small stripe sizes. The three drive 16k stripe gets the fastest score of all, revealing that for Crysis, at least, throughput is as important as seek performance.

We’ve got one more combination to try – software RAID.



Well, that’s pretty clear – Windows software RAID owns the onboard RAID for the same stripe size for both the seek and throughput workload scores, a result that’s reflected in a faster Crysis level load score as well.