Virtualiser vos Terminal Servers
Suite à la publication des résultats de la phase 2 du projet VRC, qui concerne la virtualisation des serveurs TS et XenApp, sur des machines Intel “Nehalem” Xeon 5500, voici quelques conclusions fort intéressantes :
- Virtualizing 32-bit Terminal Server and Citrix XenApp workloads remains highly recommended on all hypervisors (when using the latest generation AMD or Intel processors): running x86 TS workload on a hypervisor in comparison to bare-metal configuration is far more scalable.
- Virtualizing 64-bit Terminal Server and Citrix XenApp workloads requires a little more thought. Windows Server 2008×64 Terminal services has been designed from the ground up to scale well on the latest hardware generation. In contrast to x86, x64 Terminal Server is not automatically more scalable when virtualized. For example, the maximum amount of vCPU’s possible today per VM is 8, as a result it is necessary to setup 2 VM’s to utilize all 16 logical processors.
- While it is possible to setup many relatively small x64 TS VM’s, this contradicts with the concept of running x64 TS, which has been build and designed from the ground up to host many users on a single Windows instance. It still makes a lot of sense to virtualize x64 TS: but not specifically for performance reasons. The management and consolidation benefits virtualization brings are also valid for x64 Terminal server workloads.
- The Intel Nehalem processor architecture brings impressive performance increases for Terminal server workloads on all hypervisors. The move from HP’s AMD (G5) to Intel (G6), with similar configurations, almost doubled the amount of users on every platform. Both Hyper-Threading and EPT-D introduced by Intel Nehalem have an impressive positive impact on Terminal Server workloads. This is also logical, because Terminal Server workloads are typified by the rather extreme amount of processes, threads and context switches within the VM. Having more logical processors available and the performance optimization addressing memory page tables is specifically beneficial to terminal server workloads.
- It is also safe to conclude that these spectacular performance increases are not caused by specific developments on a hypervisor level. Intel’s innovations in the Nehalem architecture can be almost solely accredited for the performance improvements seen with TS workloads. It is up to the hypervisors to support/enable the hardware innovations.
- When comparing hypervisors, performance is close to equal throughout when no Hyper-threading is used by the VM’s. In all test the hypervisors perform with a 5% range with the Terminal Server workloads, with a slight edge for vSphere 4.0. Utilizing Hyper-Threading, on all platforms a performance increase in seen, but vSphere 4.0 trailing slightly by 15% in comparison to XenServer 5.5 and Hyper-V 2.0. These differences are only visible under full load, under normal operating conditions all hypervisors perform practically equal. When VMware releases specific updates to improve the performance of Terminal Server workloads, Project VRC will review and update the published results.
- Strikingly, XenServer 5.5 and Hyper-V 2.0 perform almost identical in all tests. The only differentiator between these two is that XenServer 5.5 (and vSphere 4.0) support 8vCPU’s, where Hyper-V 2.0 has a maximum of 4vCPU per VM.
Attention, ces tests sont réalisés sur des hyperviseurs Hyper-V 2008 R2, vSphere 4, XenServer 5.5.
Je vous invite d’ailleurs à lire plus amplement le WhitePaper, qui détaille les scénarii de test ainsi que les résultats, en fonction des hyperviseurs.