There are some interesting discussions revolving around Robin Harris’s StorageMojo blog—too bad he’s taken a vacation just when the discussion’s getting really good. The topic is SSD architectures and whether disk-form-factor SSD arrays can possibly provide the kind of performance speed-hungry applications require.
At Kaminario we’ve long argued that a disk-form-factor SSD architecture is inherently flawed. While the SSD itself might deliver on performance, inevitably the legacy storage controllers and other legacy components become a performance and reliability bottleneck, increasing latency and reducing throughput as several SSD’s vie for the same limited pipe.
It all starts with Robin’s ’s blog, Are SSD Arrays a Bad Idea? Keep in mind that when Robin uses the term SSD here, he means disk-form-factor arrays, not Flash arrays architected like the K2. In this blog he downgrades disk-form-factor SSD arrays for their inferior performance, reliability, cost, and flexibility. He makes the same argument we do about performance—that SATA/SAS stacks will inevitably become Flash performance bottlenecks. On reliability he demonstrates that disk-form-factor SSD’s retain many of the components that typically fail on disk drives and on price he highlights the extra expense of mounting Flash in a disk form factor compared to on a PC style board. He concludes that such arrays are a somewhat decent temporary architecture solution but will lose out over time to more Flash-centric architectures.
In SSD arrays: limits of architectural critiques, Robin rebuts a Howard Marks’ blog (Are SSD-Based Storage Arrays Doomed?) that tries to shoot down many of his architecture arguments. Howard argues that some vendors have achieved very low latencies with disk form factor arrays. Robin agrees, but argues they are the exception not the rule. Howard seems to agree with Robin’s take on cost, even if he sounds like he doesn’t, and tries to argue that disk failures don’t come from the components Robin points to.
In the end, Robin leaves us in limbo, concluding that architectures are just architectures and that vendors can employ clever engineering to get around an architecture’s drawbacks or screw up a supposedly faster, more reliable architecture to create an inferior product. But if you compare a storage system’s architecture to the foundation of a house, then we all know how that story goes. No amount of home improvements are going to help you if the foundation is cracked.
But, thankfully, Robin comes back to earth in the next blog. In The SSD write cliff in real life, he compares write benchmarks of SSD array products based on the different architectures and proves to his satisfaction that those architected like disks exhibit more latency than those using board mounted Flash. He looks at average response times, 90thpercentile response times, and maximum response times, and shows that non-disk-form factor Flash arrays come out ahead, with one exception.
Conclusion: Architecture matters. Current implementations of disk-form-factor arrays can’t keep up by and large with architectures. I’ve only scratched the surface though. Take a look at these blogs and those that we have written over time here in TheIOStorm. You’ll get an interesting education about SSD array architectures.
Tags: architecture, benchmarks, bottleneck, disk form factor array, Flash, Howard Marks, K2, Kaminario, latency, Robin Harris, SAS, SATA, SSD, SSD array, SSD performance, storage controller, StorageMojo