Are SSDs worth the extra cost?

Whether solid-state drives are worth considering over disks will come down to cost calculations involving performance and the nature of your datasets.

Can SSDs be as cost-effective as rotating media? That's the question posed by a thought-provoking analysis from George Crump, senior analyst at Storage Switzerland.

Your first reaction may be, "Of course not." But many find it's justified to pay for the extra performance, typically just for the data that's worth speeding up. For example, in a datacentre, you might speed up just a database's working set or, on a personal machine, just the OS and applications. Having the OS or a complex application such as Photoshop pop up in seconds is worth the price of entry.

As Crump points out, most cost-justification in the enterprise for SSDs is done on a cost per IOPS, not cost per GB. That way, you compare it with the cost of short-stroked disks, of whose capacity you might only be using 50 percent. The rest of the disk space goes to waste, along with the cost of spinning and cooling the things. SSDs score here.

According to Crump, they also score when you front-end a spinning disk system with an SSD cache but the performance calculation is more complex. Flash-storage vendors like to talk about how many extra IOPS you get and the time you save when using SSDs as a cache but never, in my experience, point out that this technique only works when the proportion of cache hits reach a certain level.

If there's a cache miss, you're into negotiations between disk systems:

Have you got this bit of data?
OK. Do you know where I can find it?
Try over there.

Not pretty or desirable. But throw in some deduping and compression and, with luck and a following wind, you might claw back some half-decent capacity savings, though nowhere near the volumes you save when backing up multiple VMs or client PCs. That might bring the justification of SSDs closer into line.

In the longer term, and as flash memory prices fall, I wonder if we might not see more systems using some form of execute-in-place technology, where the distinction between memory and storage becomes more blurred, in much the same way as it is in mobile phones. I appreciate that's more in the way that it's described and used, rather than in reality.

In the meantime, flash-memory vendors mostly sell systems that use a disk interface between the computer — that is, the CPU and memory subsystem — and storage. Not only does this practice add upfront cost, as Crump points out, but it also adds to your opex, as you have to house and power those disk controllers, as well as keep their drivers up to date.

Alternatives are emerging, such as that from the likes of Fusion-io, whose PCI-E-connected flash memory storage goes some way towards mitigating this problem.

But for most organisations, this approach is both way too expensive and potentially limiting — just a single vendor sells the product so you'll need a strong justification.

Instead, when considering SSDs to boost performance, ask if it really is cost-justified, and whether your datasets are suited to the kinds of usage that SSD vendors envisage when showering you with shiny PowerPoint presentations.