RSS
Email
Print
ShareThe Truth About Shelf Based SSDs
Posted by George Crump on September 15, 2009
Shelf based SSDs are SSDs designed to be installed into storage system manufacturers' current drive expansion shelves that they currently use for mechanical hard disk drives. It seems like a good strategy and a quick path to an SSD offering for storage manufacturers. The truth is, however, that either manufacturers or the storage manager need to be aware of the ramifications of loading up a shelf of SSD technology.
Companies like STEC and Pliant Technology offer drive slot compatible SSD drives. In fact, Pliant just announced a high performance, dual ported drive that is at the top end of performance for the category. There is nothing wrong with the technology itself, instead the challenge is going to be with the storage OEMs and how they implement the technology.
These are more than just faster hard drives, they are 15X to 30X faster, depending on the application and your read-to-write ratio. As we outline in our Visual SSD Readiness Guide, determining what applications can take advantage of SSD takes some investigation on your part, but it is not difficult. Knowing the read-to-write ratio is important in determining what SSD technology to use and what kind of performance you can expect. For example, some will provide better performance on mixed read/write workloads than others.
Storage system manufacturers have to break out of the old way of thinking; filling up a shelf with many SSD drives is not like filling it up with mechanical drives. In the mechanical world, the I/O capabilities of the shelf were never an issue. In the SSD world, a half dozen drives may max out the I/O capabilities of the shelf, add more drives, and you won't see any more performance. A full shelf SSD will still outperform a shelf of mechanical drives, but it will be nowhere near what you are paying 15X more for. I think customers are going to want to get all the performance coming to them.
Storage systems manufacturers are going to have to give customers detailed best practices guides. Do they implement half full shelves and build the array vertically up the rack? Can you then mix slower drives in the same shelf for capacity centric applications? Will the dramatic difference in performance cause issues? This is more complicated than when we were mixing SATA and fibre mechanical drives for the first time. Could it be a viable alternative to come out with lower drive slot count but high performance shelves that are only for SSD?
Comment by Captain Obvious on September 15, 2009 9:47 PM
What makes SSD any different? Over the years we've seen mechanical drive updates that have often rendered prior generation storage systems a bottleneck. The cycle has been slower but still relevant.
We've just seen a bit of a jump in the cycle due to SSD's coming to market but it's the same problem we've had for a long time.
Reply to this comment
Comment by George Crump on September 15, 2009 9:58 PM
Captain, I must have missed your point. We have seen no significant drive performance increase in almost 10 years since the 15K RPM drive came out. Yes we have seen other work arounds, like fine grain virtualization and wide array stripes, but they can only carry 10 year old technology so far. While SSD's are not new, the price point and capacities of the systems are.
SSD are a massive upgrade in performance, a delta that we have not seen in years. And now the reliability significant as well:
See "Flash SSD is Reliable Now" http://bit.ly/BaO7r
Reply to this comment
Comment by Mike Young on September 16, 2009 1:02 PM
George,
I agree that drives haven't gained in performance like they've gained in density. But there are many who feel SSDs are just a magic pill. Now, I'm not suggesting you're in that camp. I know better than that.
I went ahead and took a server and simply swapped the SAS drives with SSDs to illustrate the point. The application is Samba's smbtorture, which tends to stress much more than the typical Iometer style of testing.
Anyway, you can see the results here: http://cachengo.com/blog/2009/09/ssd-usage-shouldnt-we-consider-these-significantly-faster-drives/.
The reason I find this relevant is there are SMB customers asking about this stuff. But the bang-for-the-buck value proposition is quite low for such customers.
Mike Young
CEO, Cachengo
http://cachengo.com/blog
Reply to this comment
Comment by Merrill on September 17, 2009 12:22 PM
Another consideration is software readiness for improved storage I/O performance. Applications are written in very abstract but inefficient languages; they run on multilevel middleware stacks; and these run in VMs on guest OSs on host OSs. This has all been encouraged by having a lot of processor cycles available per I/O operation.
As processor cycles per I/O drop, there will have to be a renewed emphasis on software design for efficient execution.
Reply to this comment
Comment by Faster RAID is needed on September 17, 2009 2:28 PM
Basically, if you RAID controller is already running at its fstest speed with HD technology, SSDS may not buy you much in read/write speeds.
But, you can max out your RAID speeds with a lot fewer drives, if you don't need the additional storage of the HDs.
And now, with Adaptec's new 800 MB/sec RAID that embeds an intel SSD (32 or 63 GB, I don't remember which) you may be able to get the performance of a high cost SSD system for chump change.
For example, suppose you are running a web server. The Adaptec RAID server will gradulaly load most of that web site into its 32 GB 'cache', and now you will be serving the site for the SSD with much higher IOPS.
This won't work for every application, but there are a lot of servers where retrofitting this raid controller is going to make a big, big difference in performance and capacity. I think the list price is about $1299.
Reply to this comment