Testing Meltdown and Spectre patches: VMware ESXI v1
In my previous Blogpost about testing Meltdown and Spectre patches, the results focused on XenServer. In this blogpost we are going to talk about ESXI and the performance impact we measured in our labs.
After all the news and hectic around Meltdown and Spectre it seems things are calming down a bit. Recent talks are focusing more on next generation exploits rather than the performance impact. The Login VSI Meltdown and Spectre emergency edition has been a great success and helped many of our customers, vendors and community friends. I’d like to share some results from our internal performance validation labs where I’ve conducted tests on a large number of operating systems measuring the impact of patches on the OS, Hypervisor and BIOS level. Please keep in mind that this is a lab environment, results will vary on your own systems.
I focused on Windows Server 2012R2 and Windows Server 2016 for these tests and started with a clean image that was optimized for performance with VMware OSOT. With every test I added a new layer of patches, adding the layers of mitigation one at a time.
|Clean||Windows 2012R2 feb. 2018 updates / Windows Server 2016 (Update feb 2018)|
|Microsoft Patch (MS)||KB4056898 for 2012R2, KB4056890 for 2016|
|Microsoft & Hypervisor Patch (MS-HV)||ESXi650-201803001 (52456)|
|Microsoft, Hypervisor and Microcode Patch (MS-HV-MCU)||Dell R730 Bios ver. 2.7.1|
When we look at the results for Microsoft Server 2016 we see that the results are like the 2012R2 tests. But the impact is just slightly less (about 15-20%).
Of course as always, these tests are done in our lab and results may vary upon testing your own hardware and software combination.
VSImax Results, higher is better
VSIBase results, lower is better
2016 % Impact closest to 0 is best
We’ve learned that Hypervisors react different to Meltdown and Spectre, versions, vendors and the order of patching will have an impact on how performance is affected. Its therefore recommended to run simulations in your own environment as mileage will vary. In the blogpost written by my colleague Blair we do see that the impact on Server 2016 is less than we see on Server 2012R2.
This raises again a number of new questions:
- Which level of the Meltdown and Spectre patches impact the ESXI hypervisor the most?
- How does optimization of the VM influence the test results?
- How do other hypervisors behave after implementing the Meltdown and Spectre patches?
These tests were performed with Windows Server operating systems. I am currently switching focus to Client operating systems namely Windows 10 and Windows 7 to see what kind of impact we can deduce here.
About the Lab used for these tests
Automate, automate, automate. In our VDILIKEAPRO test lab we have minimized human interaction (and thus error) by making sure all processes are executed fully automatic. For example the base images are created using the Microsoft Deployment Toolkit where as a next step PowerShell magic (thanks Henk & Sonny!) takes over and automatically clones the VM’s 6 times on a single host, prepares them for use, and runs 10 automated test cycles. Results are then averaged ignoring the first run.
All the connections to the target machines are done using the RDP protocol. And the Login VSI default workload for this type of testing, the Knowledge Worker workload, was used for all tests.