Solid-state disks: moving from luxury to necessity.

Even if you've managed to avoid all of the storage trade publications for the past decade, you've probably heard of solid-state disks. Chances are, though, you haven't given them much more than a passing thought. Sure, they have a high "cool" factor, being that they're made of chips instead of spinning platters. They've always been so expensive though--everyone's got a budget, and these days it's probably a lot less than in previous years. Spending a big part of that budget on whiz-bang disks just wouldn't make sense--or would it? Naturally, to justify the purchase, a legitimate need for solid-state disks must be shown. To begin, let's first recap the changes in computing and storage over the years.

The Past

Long ago, CPUs were slow, and disks were even slower. To get around this performance disparity, striping (later to be called RAID, which is an acronym for Redundant Array of Inexpensive Disks) was introduced. This allowed us to add spindles to a volume to increase the number of simultaneous I/O operations the storage system could perform. As CPU speeds got faster, more storage tricks were used--bigger stripes, different types of RAID, big front-end caches on the arrays and even on disks themselves.

All the while, solid-state disks were available. Sure, they were only used for very specific purposes. Their capacity was low, and their cost was astronomical. Because of this, they've always been considered the "last resort" for solving performance problems. I remember selling a 4MB solid-state disk in 1988 that cost $125,000! Unless you were running something like a nuclear simulation there was little chance you could swing that kind of purchase price.

However, times have changed due to shifts in both memory and CPU technology. First, the price of memory chips has plummeted beyond anyone's expectations. Additionally, the capacity has gone way up while form factor has gotten smaller. This makes solid-state disks much more affordable than they used to be. Second, technology advances have caused an explosion in CPU performance. While it's true that CPUs have always been faster than disks, this performance gap has dramatically widened in the past couple of years. CPUs have moved from 15 to 100 to 500MHz, and now past 3GHz! No matter how you look at it, this is a phenomenal performance progression.

Progress

Although disk technology has progressed, it's been mostly in terms of capacity and form factor instead of performance. The 5MB 5-inch disk grew to 512MB, then to gigabyte capacities in a smaller package, and now we see disks over 150 and 200GB in a 3-inch form factor. It's true that rotational speed has increased, and so has seek performance. However, the performance improvements haven't been dramatic. We've seen disks with an average access time of 5-10 milliseconds (ms) for many years now. True, some disks are now advertising a 3.2ms access time, but that assumes optimal conditions are maintained. More important than a disk's access time is the service time: A disk's service time is the time required for the CPU to get the data to or from the disk. It's a combination of the time the I/O request has spent in the disk queue, access time of the disk, and the data transfer time of the disk and bus. On a heavily utilized disk, it's common to see service times of 100-200ms and up! A solid-state disk is so fast that the I/O requests are handled before a long disk queue can build, thus keeping the service times tremendously small. Even if a disk could sustain its advertised access time, it's hardly the "multiple order of magnitude" jump in performance we've seen with CPUs.

Present Day

Today we can buy a server with one or more 3GHz CPUs, and a number of disk drives that still get about a 5ms access time. Do you know what the result of this is? A CPU that idles along at under 20% utilization while I/O beats the disks to death. Yes, the disk drives are effectively holding back that blazing fast CPU from achieving anywhere near its potential. Imagine dictating a novel to someone "writing" on stone with a hammer and chisel. Get the picture?

What's the Solution?

It could very well be solid-state disks. Not for all applications of course, but there are quite a few that really need the performance advantages--and these applications are much more common than one might expect. Database logs and indexes are huge I/O bottlenecks, as are mail server queues. And there are many other everyday applications that are impacted by I/O performance to rotating disks.

What about the cost? Sure, memory prices have fallen tremendously, but is it enough to make the solid-state disks affordable? In a word, yes. Prices have come down so far that solid-state disks can often be the fastest and cheapest way of solving an I/O problem. Is the cost per megabyte as low as a traditional disk? No. That's because of economics of quantity. There are thousands of rotating disks shipped for each dozen solid-state disks. That won't always be the case, but it is for the time being. While the cost of a solid-state disk is roughly 500 times that of a rotating disk, its performance is around 1000 times better. A disk's access time of even 5ms is an eternity compared to the access time of a solid-state disk, which is measured in microseconds (1 ms is 1000[micro]s). That performance can unlock the horsepower of those fast CPUs you've just purchased.

We haven't just experienced changes in technology--we're also seeing paradigm shifts in computing in general. Traditionally, the server has been the most expensive device in the machine room. That's not necessarily true anymore, the cost structures have changed. It may be unfamiliar to think of data storage costing more than the server, but now it's a fact of life. Don't think of a solid-state disk as being an unreasonable purchase because it's "more expensive than rotating disks." Think of it as a specialized acceleration device that's needed to fully realize the computer's performance. With new servers and storage (and their pricing), you're still getting much more power for the dollar than you used to.

Is it feasible to replace all of the rotating disks in an IT environment with solid-state disks? Certainly not, at least not today. Fortunately, there currently isn't a need to do so. Typically, only about 5% of the data in a given application is a "bottleneck." It's the data in those small key files that should be moved off of rotating storage. Adding solid-state disks to an IT environment has benefits beyond just speeding up accesses to those key files previously mentioned. When the "hot" files receiving most of the I/O are moved off of the traditional storage arrays, the cache in those arrays becomes much more efficient, since the "hot" data doesn't keep pushing other useful data out of it.

If that's not enough, think of the alternative ways of achieving more performance:

Adding systems: The Gartner Group has shown that only 17% (and now that's probably less due to the past two years' price drops) of an IT budget is hardware costs. What adds up faster is the cost of additional software licenses, hardware maintenance, system administration, consumables (backup tapes, etc.), facilities (power, cooling, floor space) and so on.

Adding more disk spindles: While this has been a tried-and-true method for enhancing performance, it has limiting returns. The more disks that are added to a stripe, the lower the MTBF goes. Unlike a solid-state disk, a rotating disk has moving parts and thus a much higher failure rate. The greater the number of drives in a stripe, the lower the stripe's MTBF. Write performance can suffer if using a RAID level that has redundancy built in. Calculating data parity can add a tremendous amount of overhead. This is particularly so when writing less than an entire stripe as data must be read, parity recalculated, then re-written. In a random I/O environment, no number of added disks will help--assuming the data's not in cache, accessing the data means that a disk still must spin, and a head must still seek.

Adding memory to a server: This can help in some cases, but it also involves certain risks. Should the server go down for any reason, the data held in system memory is lost. Depending on the business, the effect can range from "inconvenience" all the way to "corporate catastrophe."

Modifying the application: It may be possible to enhance the application to run more efficiently. However, these changes can take a long time and potentially introduce bugs into an otherwise working program. Even then, the performance gains are often incremental at best. Besides, good software developers and tools can often cost far more than a solid-state disk.

www.bitmicro.com

Kelly Cash is the technical evangelist for BiTMICRO Network (Fremont, CA)