Ready-to-Use Patch Sets

Manual kernel configuration is a good way to learn about the kernel. However, if you want to play it a little safer and don't have time for a deep dive, several performance-optimized custom kernels are available for download. Two popular patch sets are XanMod and Liquorix.

The overall benefit of using a custom Linux kernel will vary depending on the workload. The average boost in synthetic tests may be around five percent, but the difference in experience (for example, the start time for Firefox while playing a high-resolution video in the background) might be considerably more significant. Keep in mind that many of the described optimization techniques that involve different scheduling and changing inner kernel timers only show up when the system has a significant load. The best way to conduct a simple home-grown benchmark is to set up a resource-hogging process, such as video encoding or a huge file compression, and then try to do some normal browsing along with it. Custom-patched kernels will likely deliver a smoother experience and maintain responsiveness even if CPU usage is nearly 100 percent.

As for real-world benchmarks, I conducted a series of tests using the sysbench tool. Sysbench can stress test Linux systems and test CPU, memory, threads, and I/O performance. The tests were run against three flavors of the 5.12 kernel: the mainline Ubuntu kernel, the XanMod kernel, and the Liquorix kernel. Table 1 shows the sysbench results for my test system running on Intel Xeon E5450 with 4GB of memory and with Pop!_OS 20.04 installed on a budget SSD drive. Keep in mind that these results are from my tests run on my hardware. The developers of these projects will likely have their own tests for specific scenarios in which their kernels excel.

It is clear that all three kernels run very close to each other when measured for CPU performance, with XanMod only 1.3 percent faster than the generic kernel. In the threads test, the contenders showed different results as long as each used its own process/thread scheduler. The vanilla Linux kernel comes with the completely fair scheduler (CFS), which turned out to be on par with XanMod's CacULE (enhanced ULE scheduler originally derived from FreeBSD), but both lost the race to the Multiple Queue Skiplist Scheduler (MuQSS) enabled in Liquorix. It turns out that in a densely threaded environment, MuQSS scores up to 40 percent better than its rivals.

When it comes to memory latencies, you can clearly see that both kernel flavors performance-wise try their best to reduce the maximum latency value. In that regard, the record belongs to XanMod, but the average latency is still lower in Liquorix. The I/O test for sequential writing then showed that XanMod was actually more balanced, and it delivered the best throughput figures, whereas Liquorix yielded 10 percent less write speed. By the way, both XanMod and Liquorix were using Budget Fair Queues (BFQs) for handling disk loads, whereas the generic kernel stuck with mq-deadline. So, in the end, the choice comes down to the acceptable throughput trade-off for the sake of better responsiveness. Regardless of your choice, both the customized kernels will run faster than the generic kernel.

Conclusion

Linux was created for hacking and tinkering, and users still have many options for tweaking the Linux kernel. This article touched on a few important options, and you will find many more if you spend some time with the Linux kernel configuration menu and browse the kernel documentation [6]. Just remember that it is better to test these techniques first on a non-critical system before you roll them out on your production network.