This afternoon I attended the #VSP2384 session – Distributed Datacenters with Multiple vCenters Deployments Best Practises. I didn’t take notes as I quickly got absorbed by this excellent presentation. I’m already looking forward to downloading the video of this as soon as its uploaded to the VMworld website and watching it again.
However, as a brief abstract here is some of what they covered:
Ravi Soundararajan, who has 7 years of experience looking at customers vCenter performance, explained the components of the vCenter software. He then looked at how different factors such as network latency and bandwidth over WAN links could affect your design. Ravi explained the needs of each of the software components in vCenter and how to right-size the server hardware.
Ratnadeep Bhattacharjee then took over, and jumped into different vCenter deployments for spilt sites, and covered the effects of using Linked Mode in the design.
Some of the useful bits I gleaned were:
- Hosts separated over WAN links from their vCenters are more sensitive to network latency than they are to bandwidth.
- vCenter network traffic is regularly very bursty (up to 10x).
- vCenter 5 is more resource hungry than previous versions due the additional services they provide. So you should consider upping the hardware as you move to version 5.
- JVM resources should be sized appropriately (can be bad to oversize).
- vCenter minimum hardware requirements are minimums. You may need more for your environment.
- There is a hard limit of 500 connection sessions on vCenters and this is the same for Linked Mode vCenters (collectively, all Linked Mode vCenters only get 500) . Clients use one and each VM console uses one. Other 3rd party software can use these as well. So this limit can hit the larger deployments.
A couple of things I got from questions I posed at the end was this:
- The official vSphere 5.0 documentation (PDFs) state you can’t keep vCenter 5.0 in Linked Mode with 4.x vCenters. However the vSphere 5.0 Release Notes state you can. The former is true and the latter is a documentation mistake. As you upgrade your Linked Mode vCenters, you disconnect them from 4.x hosts and can only join them back to the other vCenters after they are upgraded.
- When you upgrade your vCenter from 4.1 to 4.1U1, your client doesn’t automatically ask you to upgrade (like it would if you were connecting with a 4.0 client to a vCenter 4.1). If you are running a 4.1U1 vCenter, you should manually uninstall and install the latest client. This resolves a number of client side issues.
If you are interested in vCenter designs for multiple sites, or just want to understand more about the internals of vCenter, then I highly recommend you check it out once it’s available. I heard from some folks this week that they felt the super technical deep dive stuff was lacking this year. If you wanted more of that stuff then I’d say you should download this session.
4 thoughts on “VMworld 2012: #VSP2384 Distributed Datacenters with Multiple vCenters Deployments BP”
I’m curious about this statement: “•Hosts separated over WAN links from their vCenters are more sensitive to network latency than they are to bandwidth”. Is there any documentation or official sources that go into this in more detail?
Our organization currently has one VCenter supporting host clusters in remote datacenters across continents, and I’m now wondering if this is going to burn us someday.
This configuration won’t burn you. I was merely saying that our experiments seemed to suggest that when we increased the latency, the impact on overall operations was a bit more than when we decreased the available bandwidth.
ps. Forbes, thank you so much for the kind remarks on the talk.
Thanks for responding to this Ravi.
Terry, I’ve run split vCenter links over less than normal connections such as:
– Cross-continent links of 100Mbps but latency over 200ms.
– Local narrow links where latency is not an issue, but bandwidth is thin DSL or ISDN type connections.
– Satellite links where bandwidth is obviously bad, but latency is terrible.
I wouldn’t worry unless you are talking extremes, as Ravi was really explaining worst case stuff. My personal experience is that bandwidth has to be very low to effect general work (although console access is noticeable quickly). However poor latency conditions has forced us to install remote vCenters at some sites, which connect back to the mother-ship with linked mode. Although administering multiple vCenters isn’t ideal, it does make for a workable solution.
I realize VMware would like us to aggregate all of our vCenters through vCD eventually, but that’s overkill for many organizations. An enhanced Linked Mode, or better yet a vCenter for vCenters would very welcome. Hopefully as they work to build out the web client, we’ll see more of this type of functionality.
Thanks for the follow-up. I originally interpreted your comment to mean the hosts would start throwing errors over a certain latency threshold. In fact you were reinforcing my current experience of accessing vcenter over a 10Mb wan link, where I can count off the seconds when performing certain operations on my hosts, which are local to me.