mclock disk scheduler not normalizing properly?

[knuter]

Hi guys,

I'm using limts on the vmdks to limit the disk performance, but something is not working correctly. I have the IOPS limit set to 3333 IOPS for a VMDK, and that part works fine. But the part of mclock normalizing IOs to 32KBs to account for large IOs is not working.
3333 IOPS x 32KB should be 106 656 KB/s, but it does well beyond that. It happens with VMDKs on several VMs, not just a single one, but not all it seems.
This VM is running hardware version 11 and the ESXi host is 6.5 build 7526125.

Anyone else using mclock and IOPS limits? Is it working properly for you?

Here you can see the average is actually 255 904 KB/s while IOPS remain within 3333.

Limits are set to 3333 IOPS for disk 0:2

IO size to normalize on is left to the default 32KB (32768 Bytes)

[knuter]

One of the other VMs, with a limit of 4000 IOPS for each VMDK, which should result in 128MB/s. It hovers around that most of the time, but suddenly it spikes through with 500MB/s as the same time as the IO size increases. Here it almost seems like mclock checks the average IO size a few minutes apart and between meassurements the VM is allowed to go crazy.