Looking At The Linux Performance Two Years After Spectre / Meltdown Mitigations
Last week marked the two year anniversary since the formal public disclosure of the Spectre and Meltdown disclosures. To commemorate that anniversary, I was running some fresh benchmarks of various Intel desktop and server processors with the in-development Ubuntu 20.04 LTS to look at the performance impact today with the default CPU vulnerability mitigations and then again with the mitigations disabled at run-time.
A daily snapshot of Ubuntu 20.04 LTS was used as of last week for offering the very latest look at the Linux performance two years after Spectre/Meltdown entered the public spotlight. Ubuntu 20.04 LTS right now is running on a Linux 5.4 based kernel, GCC 9.2.1, and the other latest stable packages.
Tests were run out-of-the-box with the defaults mitigations for ITLB, L1TF, MDS, Meltdown, Speculative Store Bypass (Spectre V2), Spectre V1, Spectre V2, and TSX Async Abort (TAA). Tests were then run when disabling the mitigations via the "mitigations=off" kernel command-line option that can be used for disabling the various mitigations. Tests were not done disabling Hyper Threading or going to other extremes but simply looking at how the default Linux performance has changed two years after these vulnerabilities became public.
Desktop CPUs tested were old and new through the recently released Core i9 10980XE Cascadelake-X processor that sports many hardware-based mitigations:
- Intel Core i7 5960X
- Intel Core i7 8700K
- Intel Core i5 9400F
- Intel Core i9 9900K
- Intel Core i9 10980XE
Meanwhile on the server CPU side were various Intel Xeon parts through the latest Cascadelake processors and their hardware mitigations:
- Intel Xeon E3-1280 v5
- Intel Xeon E3-1275 v6
- 2 x Intel Xeon Platinum 8280
Via the Phoronix Test Suite a variety of mostly real-world workloads were used for showing in relevant workloads how the performance of these mitigations continued to shift the Linux out-of-the-box performance. With the latest Cascadelake processors the impact is much less thanks to the hardware mitigations but still not entirely eliminated.
First up in this article are the desktop tests followed by server.