Jeff Baxter

And that's a little more complicated than that because depending upon your workload mixes and block sizes and all sorts of other things.

But basically, if we put an SPC1 benchmark out there, which we have in our system, you can see that's the end capacity and performance capacity that we end up advertising.

So whether that's 200,000 IOPS or 250,000 IOPS or a cluster doing a million IOPS, that's the point at which we say this is how much we're basically saying the system can do.

Now, storage systems can always do more than what you say they can, but we call that the knee of the curve.

And by that, I mean you start to see the latency, which up to that point may have been, you know, on our AF system might be 0.5 milliseconds latency, 0.6, 0.7, and it's roughly going up towards that one millisecond line.

I think with our SPC1, we basically capped it out at 0.8 milliseconds.

And then when you go past where the vendor says that the system is going to be able to perform,

Basically, what's happening is for every extra IOP you're getting, the amount of latency you're adding is increasing, and the line is sharply turning vertical, if you can imagine in your head, right?

So you may get another 10 IOPs out, but you get another 0.01 milliseconds of latency, and then you go farther along, and then for every 10 IOPs, you're getting 0.1 milliseconds of latency, and then for every 10 IOPs, and so on and so forth.

So that curve starts to go vertical.

So the optimal point is the point at which the system is still basically scaling linearly up to that point.

It's predictable performance up to that point.

And that's what we call 100% of the performance capacity of the system.

So then we go ahead and we take the workloads that are already running on top of it, so the performance capacity that's already being used on the system.

We subtract that from the optimal point, and that gives us a metric called headroom.

And Headroom is built directly into ONTAP 9.

It's exposed through different tools like the On-Command Performance Manager and increasingly through things like System Manager and others.

And basically what it does is it tells our system, it tells our end users, and more importantly, it tells some of these new things like balance placement that we'll talk about.

This is how much capacity is left on the system.

You can put this much more on the system if you want to.