As the United States looks to approve the START treaty, I thought it was time to propose that SSD manufacturers enter their own strategic arms reduction treaty to control the rampant and destructive proliferation of million IOPS marketing.
The road to an IOPS arms race began innocently enough, SSD manufacturers had a novel story to tell. The IOPS from hard disk drives have been atrocious since the dawn of computer time. Even today, the lowly hard drive can only squeak out 300 random IOPS. From the earliest days of the SSD, IOPS marketing was a big part of the story. The beauty of a solid state storage device was that it could move more data (IOPS) to a processor faster (less latency) than traditional disk. This simple story has been at the core of the SSD value proposition for 30 years.
Admittedly, I fired the first shots (and probably the second, third, fourth….) in the escalating IOPS arms race in 2001 when Texas Memory Systems announced the RamSan-520 a 5U monster of a system with all of 128GB of RAM capacity. This system with fifteen 1Gbit Fibre Channel ports was said to deliver 750,000 random IOPS. Do you have any idea what kind of reaction that generated at storage conferences in 2001. Wow! Impossible, most would say. This process led to TMS proudly declaring itself the “World’s Fastest Storage®”. After firing this weapon for nearly ten years, I have to admit the time has come to stop the IOPS marketing arms race. The challenge in 2001, as it is today, is finding the customer that can drive 1,000,000 IOPS. Actually, in 2001 finding a server to drive that many IOPS was impossible. The processors, operating systems, host bus adapters, etc. were all too slow. Fortunately, the imposition of Moore’s law on electronics led to breakthrough after breakthrough enabling SSD manufacturers to demonstrate high IOPS with single server configurations.
I like to think, but hesitate to admit, that in addition to pioneering million IOPS marketing, TMS also drove the widespread use of IOMeter (a tool used to test storage devices by generating IO) at trade shows. As a storage marketer my grandest dream was to find a customer that needed to run IOMeter as their business application. What a perfect customer this would be. I searched the world over. Strangely, the financial exchanges didn’t need IOMeter to complete trades. Telecom companies didn’t need it to bill cellular customers. Who could possibly need IOMeter for their business? Imagine my glee when host bus adapter manufacturers and switch manufacturers started caring about IOPS as a marketing tool. Finally, my dream customer had arrived. My apologies to storage industry exhibit hall wanderers; I too have tired of seeing IOMeter.
This brings us back, admittedly after a brief tangent, to million IOPS claims. I hesitate to examine the number of vendors that are persistently and proudly proclaiming profound performance. One million IOPS. Yawn! Is that all you’ve got! In fact, I would argue if the extent of your marketing message is your IOPS you don’t have enough… marketing talent.
Using a solid state storage device is about a lot of things: application acceleration, lower power consumption, enabling business growth and solving mission critical problems. Tell us the customer stories. How many 1,000,000 IOPS customer stories have you read? Hmmm. Fair enough, 1 million IOPS sounds like so many that people will stop worrying about whether the storage device can meet their production performance requirements. But can we stop at 1 million? Perhaps we should shoot for 2 million. No. It is time for the SSD IOPS marketing proliferation to be stopped while the customers still care. Start designing systems that satisfy the range of customer buying requirements: low latency (the number one reason most customers benefit from SSD), good-enough IOPS, bandwidth suitable to the application’s goal, five 9’s reliability, low mean time to repair, low power consumption, interoperability and low total cost of ownership.
Couldn’t agree more.
application performance trumps everything. Most of the 1m iops benchmarks are for read only small blocks and are using so much CPU that there is none left to do any meaningful processing with the data read:
Thanks for the article, I am an orthodontist with a Dell 2900 III running my office that I want to speed up. Kind of your lower lvel but maybe more numerous type customer in the future. I just want my 300 gig database to go faster. Would be nice for folks like me if there was something easy to figure out about how much speed help we might see if we bought something like the Fusion IO 320 gig PCIe card for our server. To sell a lot of these little business like mine will need to start wanting to buy them and need some way to understand them.
Thank you for submitting your comment. You are right that there are a whole bunch of people on the sideline deciding when to adopt SSD for their environment. The answer that you want to hear is that a PCI SSD will make your application faster. Technically, by storing your database on SSD, reads and writes will be quicker. Let’s say that today you store your 300GB database on two mirrored SATA hard drives (your environment could be better or worse…). In a mirrored configuration, those two hard disks cannot provide more than around 150 random reads per second. A good PCI SSD can provided tens of thousands of reads/writes per second. Hard drives also have very high latency compared to SSD. For example, reading or writing from a PCI SSD can happen 20 to 100 times faster than from a hard disk drive. So, if you were to monitor your application from say a Windows performance monitor you would notice that the storage was not a bottleneck for your application. The big question, and one that is hard to address without more data, is whether the people using this database would notice the performance increase. If you have a lot of users doing things simultaneously or you have some complicated reports you have a better chance of noticing the improvement from SSD.
The other things to keep in mind as you consider SSD is that whatever means you use to protect your data today, such as mirroring of disk drives, backups, etc. you would want to continue using.
If you would like, email me and I can send you a document that explains how to monitor your existing solution to determine if you have a performance problem that SSD can solve. SSD is not a silver bullet solution for every application performance problem, but for applications with alot of disk IO, it can remove one of the most common bottlenecks in database environments.