Sizing Horizon Enviornments

[sjesse]

Does anyone have any good resources on sizing horizon enviornments, outside of the maximums guide?

I don't have a big environment , but its getting bigger. Basically we are trying to provide a desktop as a service to our users. The come to us with there use case, and we create a desktop pool based on there needs. These are very small pools, 15 - 20 desktops/ users at a time, and right now we have about 15 pools. The mistake I think I made in sizing this is with instant clones, I wasn't including the instant clones in the calculations, and I'm starting to see some sluggishness. We have vRealize operations manager running with the horizon adapter setup, and everything there look ok, but I manually looked at a few of the instant clones having issues and I saw cpu ready values around 10%. My understanding in most cases you want this below 5%, so I'm thinking that may be a side effect of all the extra parents running. With 7 hosts running that 30 extra desktops, and our images are between 2 vcpu and 4 vcpu, so thats any where between 60 and 120 extra vcpus running. The total cpu cores we have right now are 104, and the total vcpus running are 714, which is around a 7:1 ratio.

Any suggestions or resources is appreciated.

[BenFB]

First off I'll say that the ideal vCPU to pCPU ratio varies based on the hardware, VM configuration, OS image/what applications are running and the end user. For our environment I've found that we can run at 7:1 without any user complaints. It's important to differentiate between physical cores and logical processors (hyperthreads) when doing the calculation.

Can you tell me more about the hardware configuration of your hosts and what versions you are running.

[sjesse]

Hi Ben

Thanks for taking the time. We are running horizon 7.4, appvolumes 2.13.3, vsphere 6.5, and most of our desktops are windows 7 desktops.I'm working with re purposed hardware currently, which are ibm hs23  blades with xeon e5-2665 processors, which provide 16 cores between both processors. We are in the area where new hardware may be possible, which is one of the reasons I'm really focusing on this right now.

I've been reviewing our metrics, and we are overcommiting alot of cpus, mainly because we initially ported our standard desktop images as the image size we are using in virtual desktops. I'm currently at a 5:1 ratio now, with no image using more the 4 vCPUs. We had our RDS servers at 8 vCPUs, but due to concurrent logins, I've had to deploy more RDS servers to handle those better, so I'm not using the 8 on any of them. Our IT department uses these and I was using 4 cpus for the, but I can't really see them using more then 2, so I'm trying a mix of 2 vCPU and 4gb of ram and 2 vCPU and 8GB of ram. From what I saw at vmworld and other research I think windows 10 should work similar but may need more memory, which we are almost don't testing with.

[BenFB · Accepted solution]

The E5-2665 has 8 physical cores and with 2 sockets populated you would have 16 cores like you mentioned. Assuming a 5:1 vCPU to pCPU ratio you should be able to get 80 vCPU per host. This would work out to 20 x 4 vCPU VMs. I would recommend right sizing as much as possible since with 16 physical cores you can only simultaneously schedule 4 x 4 vCPU VMs and the rest are waiting for their turn. With the same number of machines you might see better performance if they were all 2 vCPU instead of 4 vCPU. Regardless, if all of the virtual machines are heavily used at once I could still see you having contention and high ready times.

Have you looked at esxtop? This is a great guide. http://www.running-system.com/vsphere-6-esxtop-quick-overview-for-troubleshooting/

[sjesse]

I have, and used it before, and I've planning running in batch mode next week, during the busy time to double check if we are having the issues any more. Your point about keeping the vm cpu size down is the important key I was missing. We had way too many 4 vCPU  running at the same time, and I did see a decrease in cpu ready after changing this. Its been a few days and no one really noticed, and I'm not getting any alarms about cpus at 100% so I think we are in better shape.

Thanks I appreciate the feedback.