The top 3 questions to ask when researching SSD performance numbers.

By Andrew Mierau - 2017-03-22

SSD Hero Numbers 

Often when we look at SSD specifications, the first thing that pops out are the ‘hero’ numbers.  These are the numbers that manufacturers and their marketing departments use to promote the superiority of their product. They also typically come with a big, giant asterisk attached to let you know how they got to those numbers.  Every industry has their own hero numbers. For instance, car manufacturers will show you the time it takes their car to get from zero to 60 mph and SSD manufacturers will show you random write performance numbers.

With the release of the 5100 Series SATA SSD, Micron showed the fastest enterprise SATA SSD currently on the market with 74K random write IOPS*.

As a reformed engineer working in marketing, I always check the asterisks on any marketing claims, and you should too.  It helps you understand how those numbers were determined and whether they apply to you.

The simplest way to look at any hero claim is to ask yourself these three questions:

  1. Is it reproducible?
  2. Is it applicable?
  3. Is it the whole story?

In general, SSDs from major manufacturers have no issues hitting their hero numbers, Micron included.  Unlike a car’s hero numbers (where you might need specific tires, on a specific track, at a specific temperature and humidity while be driven by a professional driver) SSD hero numbers are typically reproducible with numerous free benchmarks (IOMeter, FIO, etc.) on any modern hardware.

The bigger question is whether the hero numbers are applicable to your workload and application.  The answer to that is tricky.  In general, very few customers will ever have a consistent workload that matches the hero numbers.  In the same way that most drivers don’t continuously do 0-60 mph runs, most SSD customers do not expose their SSD to 100% random 4KB writes at queue depth 32 for extended periods of time.

You might think this is the end of the story, but like I said, it’s tricky.  In mixed workload environments (those which have a mixture of reads/writes), SSD performance is heavily influenced by write performance.  Historically this is because read performance has been significantly higher than write performance, so increases in write performance will be magnified in mixed workloads when compared to similar increases in read performance.

Just like 0-60 mph is an indicator of a sports car’s performance, random write performance is an indicator as to how an SSD will perform in a given deployment.  They don’t cover every use-case, but contribute in many.

Finally, does random write performance tell the whole story?  Of course it doesn’t.  No single number can ever do that for an SSD or a car for that matter.  But, are there other things that we can look at within that hero number that will give us a more complete picture?  Yes, there are.

For any workload applied to an SSD, you can measure and report a lot of data.  IOPS (input/output operations per second) is a standard metric that the industry uses, but it is only an average and a lot of data can be lost in averages.  You can also measure latency, Quality of Service, consistency and more for a given workload.  This is where Micron has differentiated itself with the 5100 Series.

5100 Write Performance Consistency

Displayed in the image above, the Micron 5100 MAX delivered more IOPS and higher consistency with that performance. When compared to a competitor’s SATA and NVMe SSDs, the 5100 had higher IOPS (hero number!) and it delivered those IOPS more consistently. 

As PC Perspective noted in their review of the 5100, “Performance consistency not only exceeded Micron's specification, it exceeded my own expectations!”  I can’t say I disagree with them.

Whether you launch your car from zero to 60mph in 2.3s, or blast through data at a rate to 74,000 IOPS with your Micron 5100 MAX, keep in mind what those hero numbers mean and enjoy being the fastest (based on whatever information is behind that asterisk).

*Congrats on looking at the fine print! 100% random write performance, 4KB block size, queue depth of 32, measured using FIO at steady state, system variation may affect measured results.

Andrew Mierau

Andrew Mierau