Microsoft, Red Hat and Intel have published their performance evaluation of the impact Meltdown and Spectre mitigation has on various systems.
There are a total of three variants for Spectre and Meltdown, two for Spectre (Variant 1 and 2) and one for Meltdown (Variant 3). In their benchmark tests Microsoft has concluded that “Variant 1 and Variant 3 mitigations have minimal performance impact, while Variant 2 remediation, including OS and microcode, has a performance impact.” Without providing numbers, they further detailed their finding as follows:
- There is insignificant performance impact on Windows 10 for desktop users with Intel Skylake, Kabylake or newer CPU
- There is a noticeable performance decrease for Windows 7/8/10 for desktop users with Intel Haswell or older CPU
- There is a “significant performance impact“ for Windows Server applications on any processor, especially I/O intensive ones
Regarding the 45 editions of Windows that Microsoft is supporting at this time, the process of patching has been started for 41 one of them through the Windows Update channel, and the other four are to be updated soon. Among the latter are a few editions of Windows Server 2008/2012 and Windows Embedded 8.
Red Hat has conducted their own set of tests on Linux machines and have concluded the following:
- <2% reduction in performance for applications using Linux accelerator technologies that do not go through the kernel but perform direct access to devices
- 2-5% for CPU intensive jobs that usually run in user space
- 3-7% for applications with “significant sequential disk or network traffic”. Applications running on the Java VM fall in this category
- 8-19% for applications with high kernel-to-user space transitions. An example are OLTP workloads
Intel has been criticized for not providing data on how the Meltdown and Spectre fixes are affecting the performance of their CPUs. Now they have published their benchmark results, some of the most comprehensive so far. For a complete description of the findings, the hardware, OS and benchmarks used for testing, we recommend reading the benchmark table (PDF) made available. We summarize here their findings:
- 2-14% performance hit for SYSMark 2014 on Windows 10/Coffee Lake/Kaby Lake processors. SYSMark 2014 benchmarks usage patterns of business office, media creation and financial analysis applications
- Up to 21% on SYSMark 2014 on Windows 7-10 on Skylake
- There are better results for PCMark 10 across all systems tested, around 1-5%. PCMark simulates home workloads
- Even better results for 3DMark Sky Diver with negligible performance impact of 0-1%. 3DMark Sky Diver is a graphics benchmark
- 5-10% performance hit for web applications running in Edge or IE
Results vary much based on the benchmark, the processor and the mitigation used. It is advisable for companies to benchmark their own applications to see if performance is affected much and decide if something needs to be done about it.
AMD said only Variant 1 of the vulnerability can affect their processors with minimal performance impact. Regarding ARM, Apple patched iOS in December and said they did not notice a measurable impact in performance across multiple benchmarks.