On 2022-08-05 22:46, Scott Miller wrote:
And to answer Jeffs question, we've not done any CVO testing in GCP, so I can't answer the throughput questions.
Neither have I/we. And won't do either, no time soon.
Scott Miller wrote:
We have found that, in general, the VM Instance type that hosts the CVO instances matters, since the Instance type also determines networking connectivity.
Yep, the VM Instance type does matter very much. Bigger = better, basically... in AWS for instance, there are some types that have much more bandwidth to the disk layer in there. And ones that have more bandwidth to the [NFS] clients if needed ("n"). Of course the CVO (ONTAP) must be able to actually drive that back end bandwidth and for especially write this is not easy for it. CVO is a dog for W, pretty much -- compared to the real thing, a FAS appliance. 7.2K rpm or 10K rpm or SSD, it's the same
Scott Miller wrote:
Disk type (in this case GCP PDs) matters, too; I'd test with Local SSD to see where that gets you.
I agree. You'll get the spinning disks our of the equation first, then see what you can get out of a CVO instance. Then try it with spinnind PD (7.2K rpm or whatever)
Scott Miller wrote:
Lastly, Throughput is dependent on workload, which determines the IO pattern. Maybe create an FIO config that generates traffic close to what you expect, and use that to drive various CVO configurations.
This sounds like something very hard to do. Unless your IRL workload really is very simplistic. I'd be hard pressed to even consider trying fio or iozone etc for any (aggregated) workloads we have here. It won't lead to better knowledge with anything a bit more complex. The only way is to get knowledge is to run the "real thing". Purchase the biggest supported CVO instance there is, configure it with as much "virtual gear" you can and then throw the real workload at it and see what happens. Does it work OK? Good. No? You're out of luck pretty much. The only thing you can do is create more CVO instances. If you need several dozens to handle your workload, then it will be a nightmare to manage, probably. You can't scale up CVO more than a little (and it becomes rather expensive!) and you can't scale out like you can with on-prem AFF/FAS appliances. So, the scale out tactic becomes to have one (hopefully!) CVO for each one of your applications. If you have 100 apps, then you'll end up with 100 CVOs. A.s.o.
/M