Linux 5.5 Performance Overall Is Comparable To Older Kernels For Most Workloads

Since the stable release of Linux 5.5 this weekend I have been carrying out benchmarks for looking at how the performance of this newly-minted kernel compares to older releases. Here are benchmark results of Linux 5.3 vs. 5.4 vs. 5.5 with an AMD Ryzen Threadripper 3970X but the results are similar to other HEDT and lower-end systems we've tested thus far.

Outside of some AMD and Intel graphics performance improvements, when it comes to overall CPU/system performance the Linux 5.5 performance is comparable to 5.4. Early on the Linux 5.5 performance was looking bleak with a sizable performance hit that turned out to be an AppArmor regression. There were some other oddities but some noise made it difficult to narrow down. But fortunately with Linux 5.5 final, the performance seems to be in good standing and I haven't noticed any big performance hits in my Linux 5.5 stable testing from desktop to server platforms.

While the Linux 5.5 performance isn't having any dramatic changes on a macro level, Linux 5.5 does have many new features. Aside from "features", there are important 5.5 changes as well like with the Threadripper 3970X and friends having the important MCE fix.

While no big changes overall, for some particular workloads there are some Linux 5.5 performance changes.


SQLite on NVMe solid-state storage at least with EXT4 has been seeing better performance on the systems tested so far. Slower storage devices don't appear to have the same benefits.


For the heavy CPU workloads, I haven't encountered any Intel/AMD systems with really noticeable differences now that the regressions have been addressed.



With some of the video encode tests and other highly threaded workloads on the 3970X, there were some slight pullbacks.


The Hackbench kernel scheduler benchmark saw a very visible hit during the time of the AppArmor kernel code regression but now is showing a slight improvement on the Ubuntu systems we've tried.

Stress-NG's forking test was up on Linux 5.5, also on Intel Core i9 systems.

But message passing performance measured by Stress-NG did drop.

Context switching performance on AMD was also surprisingly lower.

But if taking the geometric mean of nearly 100 results, the Linux 5.3 through 5.5 performance was basically flat for this Threadripper 3970X system and jiving with what we're seeing on other x86_64 tests. More benchmarks within this OpenBenchmarking.org result file. Now to begin looking at Linux 5.6 Git...