Context Of Testing :
Worldwide company. A corp Archtect team that is creating a private cloud product w/ VMware cloud suite and Nutanix / DELL bloc to support them. A 3 year planning w/ 20+ cluster deployment 'til 2018.
Giving a way of testing :
- performance at deployment.
- Performance w/ NOS / ESX /Other products changes.
on a 10 nodes XC630 cluster, using DEREK's and other internet resources, executing some performances testing :
- 8 workers + 8 unformatted drives by master,
- 1 master by node,
- ESX 6.02 / NOS 4.6.4 and more,
- 8 vCPU + 32 Go RAM on CVM,
- RF3 + dedup and compression inline (real configuration).
- 4k Random read,
- 16k 70%random 70%Read
4k random read, since it targets only RAM cache on CVM is really handsome w/ 185k IOPS by XC630 at 16 outstanding IOs. Scaleout is achieved w/ 1,850 M IOPS on 10 blocs
70% read / 70% random and 16k IO. Totally irelevant w/ a lot of errors / no throughput.
What am I doing wrong ? Do I need to format HDD for Write IOs ?thanks for your help,Thomas CHARLES
Best answer by Jon
Two suggestions, in general:
1) If you're performance testing pre-production, you might as well get to a newer NOS version that 4.6.4.
5.0 is out, available on portal.nutanix.com, and has quite a few enhancements compared to 4.6.4 across the board (lots of new features, as well as a few performance enhancements). If you dont want to go right to the latest release (i.e 5.0), 4.7.3.x has been out for a while.
Dont think that will fix this specific issue as gut feeling is that its an IOmeter setup issue, but its a good idea nonetheless
2) Check out the newer blog on IOMeter + Nutanix here: https://next.nutanix.com/t5/Nutanix-Connect-Blog/Configuring-Iometer-Performance-Tests-on-Nutanix-the-Right-Way/ba-p/13255
Now, with respect to your IOMeter write errors, when I've seen that in the past, thats because something was wrong with the disks IO meter is writing too. Usually that means they are formatted or otherwise in use, which is strange since you noted they were actually unformatted.
Also, do you see this error on ALL the IOMeter instances? or just some? Did you set them all up individually? or from some cloned template?