We have been able to virtualize Traverse in both XenServer and VMWare environments at Cisco with great performance and even higher availability than running on bare metal by taking advantage of both motion and workload distribution technologies provided by each of the platforms. Although there were some noticeable differences between IO and CPU performance, both of these hypervisors are very capable of running a large Traverse installation which is something that I was originally was a little skeptical about.
My group and I needed to prove to ourselves that we could virtualize Traverse reliably for our projected workloads over next 2-3 years. To do so we setup a test environment consisting of a BVE and a DGE and created an test suite to simulate a heavy workload using EDF. One hundred threads sent test results for 500 devices containing 250 tests each as fast as possbile. That is a total of 125,000 tests receiving randomly generated result data and with the thresholds we had defined it meant that at any point in time about 20% of the tests we had were changing state.
Throughout this exercise IO performance was our bottleneck, all of it low throughput but high in IOPS. In the end with our fastest configuration we were able send 54,000 samples per minute to our test DGE with just 50% CPU utilization (using 8 vCPUs) and between 2-4k IOPS. If your using 5 minute sample times this would equate to 270k tests. In the end, the configuration we were using consisted of Cisco UCS B200s with 2.53Ghz Nahalem chips, 10G iSCSI, and 9 SSDs in a RAID 5.
I realize this isn’t real world Traverse usage since there would most likely be much more CPU and network usage doing ping, snmp, and other tests but I think it shows that Traverse can do some pretty high work loads in a virtual environment if it’s given the necessary resources like CPU, network, and storage (especially when it is shared with other hosts). Not cutting corner’s here will keep your system slowing down or from being constantly in a vMotion/XenMotion and paging you because all of your test results are skewed.
I think if you have small install and run the your DGEs with the recommended defaults it really isn't that hard to meet Traverse's hardware requirements in a virtual environment. Our initial install ran on a 1G network and used FibreChannel SAN to a pretty slow LUN made up of 10 1TB SATA disks in a RAID 10 which was shared storage for 32 VMs. We never had performance or scalability issues that were not caused by human error.