We recently conducted a series of tests to determine the limits of VM density for a single XenServer 5.5 host running XenDesktop VMs (XenServer Single Server Scalability with XenDesktop). Previously, our VM-per-host limit for typical server-based workloads had been only 50. We knew, as did our customers and partners from their own experience running XenDesktop workloads on XenServer, that the real number of desktop VMs possible on a single host was considerably higher. The results of our tests were better than expected—130 Windows XP desktops on a 72GB, dual socket, quad core (Xeon x5570) Intel Nehalem host.
Commenting in his blog Desktop Virtualization Performance Testing, Simon Bramfitt said of these results “This is a significant improvement on Citrix’s previously published results, and certainly enough to bring it back into line with results VMware previously published”. No argument there. It’s significant in that it puts to rest the question whether or not XenDesktop together with XenServer can stand up to its competitors and achieve similar ROI based on real-world scenarios. He also noted that as important as the numbers are themselves, our decision to use a third party test program from Login Consultants (Login VSI 2.0) as the basis for these tests was the right one. We agree. Login VSI is the benchmarking tool behind Project Virtual Reality Check (http://www.virtualrealitycheck.net/) and is poised to become the standard for virtual desktop performance benchmarking (if it hasn’t already done so). Using Login VSI, we were able to establish not just the number of XenDesktop VMs we could run on a XenServer host, but more importantly, how many of these VMs could run under a considerable workload while meeting the standard for acceptable performance at the same time. We knew that by choosing to use Login VSI vs. an internal test program, these results would be better able to stand up to scrutiny in the market, the truest measure of performance.
Incidentally, 130 Windows XP VMs on a dual socket, quad core Nehalem server translates to 16.25 VMs per physical CPU core. VMW made it known just a few days ago that they had set their own sites on achieving 16 VMs per core VMware to increase consolidation ratio to 16 VMs/core?, but has yet to actually demonstrate that it can be done. Citrix, on the other hand, already has.