Do IOPS matter?

I’ve been designing VMware vSphere clusters for the last 10 years know. And with every design, the storage part is one of the most challenging. A improper storage design, results in an poor virtual machine performance.

Over the years, storage vendors added all kind of optimizations to their solution, in the form of cache. Almost every vendor added flash as an caching tear. Some only do read, but most of them do read and write.

With this cache, most vendors claim that there solution can handle 100.000 IOPS or more. But we all know, adding an flash drive who can handle 100.000 IOPS, won’t give you 100.000 IOPS in your vSphere environment.

We also have to deal with different block sizes, read/write ratio and write penalties.

What I see, is that most of the time, the storage processor, or the storage area network is the bottleneck.

In the near future, NVMe will be common in our datacenter, and the successor 3DXpoint will follow shortly. While NVMe can deliver 1,800 MB/sec read/write sequential speeds and 150K/80K random IOPS, 3DXpoint is a 1000x faster. Both solutions have an average latency of less than 1ms.

This is going to change the way we design our storage solution!

@vcdxnz001 wrote a great article about the storage area network and de speeds that are involved with the upcoming flash devices.

So, to get back to the question: Do IOPS matter?

If we have a full flash array, consisting of SSD, NVMe or 3DXpoint, IOPS are no longer the problem. All types of flash can deliver plenty of IOPS.

What matters is, latency. Ok, all flash devices are low latency (<1ms). But all IOPS have to be processed by the storage processor (in your array, or if you’re running a hyper-converged solution by your CPU) and by the storage area network. So these 2 components will determine your latency, thus the performance of your virtual environment.

Therefore IOPS are no longer an concern when you’re designing your storage.

When you design your storage solution, determine the highest latency you want to encounter and monitor this carefully. Design you whole stack, from HBA to the disk for low latency. If your latency is low, IOPS won’t be a problem.