I've been asked this question quite a lot lately, and to be honest I've been a bit stumped, other than a  gut feeling that monitoring from the OS as well as the VMware layer is a good idea, I had little else to offer the argument. So I figured I should take a more serious look at the monitoring available within VMware vCenter and come up with a more compelling argument. Perhaps the more seasoned VMware admins out their could comment to offer their thoughts on the topic?

vCenter for VMware provides an excellent selection of performance graphs enabling you to track the CPU, memory, network and I/O consumption for each VM as well as the ESX host itself. You can also set "alarms" that trigger e-mail notifications. So why would I ever need an additional agent based OS monitoring package? What more could you possibly need? Do the OS metrics become less important in a VM world?

Why would we be interested in performance stats in the first place?

Capacity planning

When your entire estate is on VMware controlled by vCenter you are well placed to make judgments on future capacity. All the information is available in a common format regardless of the underlying client OS. you can make judgments based on current resource utilization and look for trends over time. This does however assume that each of your client OS's are behaving themselves, and unless you have a handle on OS metrics you could be building capacity for poorly performing OS or Applications. You may find you are chasing your tail until you drill down to the OS and app to understand more about the loads you are planning for.

You may have already consolidated your applications onto single physical servers before virtualising them, so it might be important for you to identify the load per instance. It might make more sense for these instances to live in different VM's so that you can load balance between ESX servers, but unless you are monitoring at the OS/app layer these metrics will be invisible.

It is also likely that you are not fully virtualised. Perhaps the two week study that the consulting company did suggested a handful of quick wins for VMware, leaving some systems behind on aging physical servers. If you use a single OS monitoring solution for your physical and virtual servers you will be better placed to consolidate performance stats to assess "what if" scenarios for selecting the next physical servers for virtualisation. (and you are less likely to need another study!)

Root cause analysis following user complaint

In an ideal world, some time in the bright future we will all have user experience monitoring. (I hope that there are some out there who have and I would love to talk to you if you do!) In the mean time most of us have to answer politely to disgruntled users who are frustrated with response times, screen freezes and timeouts. These moans, whilst are not always pleasantly delivered, are a valuable piece of performance feedback. With VMware vCenter you will hopefully be able to quickly identify the bottleneck on the appropriate OS image during the period of contention. You may be able to throw a few GB at it, or a few more cores to prevent it recurring, But how will you decide that its time to split the app from the DB, or push some the batch based workload somewhere else? You are likely to be restricted to eaising symptoms rather than treating the cause, when you can dig no lower than the basic metrics

Memory management

Memory management in VMware uses "ballooning" to prompt the OS to page out lesser used pages according to the OS's own algorithm when it needs space to satisfy memory demand. You need to be sure the OS has sufficient swap space to cope with the level of ballooning VMware is likely to try. You will also be interested to see the performance impact within the OS from this operation. Without OS based performance tools you are not going to get the detail on paging that you need to create the full picture. Without good information from VMware and the OS it will be like landing a helicopter blindfolded, and I suspect most will give up and just buy more memory!