Workloads Matter

By Andrew Mierau - 2016-06-07

Specifications are, in a word, specific. They are the act of describing something precisely. While we like to boil down specifications to a single number, like IOPS or MB/s, there are many qualifying details that give those numbers meaning. For SSD performance, the workload being applied is the single biggest factor in determining a product specification. A workload is typically made up of the following information:

  • Read/Write Mix 
  • Random/Sequential Mix
  • Queue Depth, or number of outstanding input/output operations
  • Block or Transfer Size
  • Entropy, or the randomness of the data

When looking at SSD specifications, only very specific workloads are used to determine performance, otherwise the datasheet would be thousands of pages long. Unlike many protocols and standards, SSD performance is not governed by any standards body. So what does that mean for you? It means that in order to compare different SSDs, you have to look at all of the workload details in order to get an "apples to apples" comparison.

Latency is one area where there isn’t a great deal of alignment within the industry on what workload should be used in the product specification. Here are 3 NVMe products and their read/write latency specification:

   Product 1 Product 2  Product 3 
Read Latency  20 µs  20 µs  120 µs
Write Latency  20 µs  20 µs  30 µs

 Obviously, Products 1 and 2 are superior because their latency is so much lower than Product 3, right? Not so fast, my friend. Let’s breakdown the workloads:

  Product 1  Product 2  Product 3 
Read/Write Mix   100% Read/100% Write  100% Read/100% Write  100% Read/100% Write
Random/Sequential Mix  100% Random  100% Sequential  100% Random
Queue Depth  1 1 1
Block Size  512 Bytes  4KB  4KB
Entropy  100% Random  100% Random  100% Random

Products 1 and 2, by varying the workload used to determine latency, are able to show a lower (and better) specification than Product 3. This isn’t to say that either product is right or wrong, they are just different. To put this in perspective, Product 2 also specifies latency at 100% Random. In that case, the workload is identical to Product 3 and its read/write latency is 120/30 µs, exactly the same as Product 3.

If you look at the product specifications for all 3 products, the workloads details are there, but to many customers, they are obfuscated. The key, when you see any SSD performance specification, is to ask “what is the workload”, because workloads matter.

I want to hear from you. Leave a comment below and connect with us on Twitter @MicronStorage and on LinkedIn

Andrew Mierau