Thanks for the feedback, this is from a few years ago when we were still new to creating content. You should find our content a little better now, we hope 😆

Hey Danilo, Mitch here! Thanks for the question. The traditional view of benchmarking different file systems has always been to remove the cache in whatever way possible before attempting benchmarks. We believe this is a little misguided for a few reasons. The reason it has been done that way traditionally is because the assumption is that all file systems will cache equally well (or uselessly on the larger scale) and so it's better to get a real picture of what the file system is capable of in most scenarios when cache is not going to be helpful. This is assuming however that all file systems use the same caching methods and algorithms. While this may be true for many file systems that use simple LRU (Least Recently Used) cache, ZFS on the other hand uses an algorithm called ARC (Adaptive Replacement Cache). This is a much more complex caching system and allows for the ability for higher efficiency and better performance for many workloads. Now all of that being said, for ZFS On Linux direct IO was not possible until version 0.8 but it is now available. So in the end, what we typically will recommend is to look at the workloads you are planning on using your ZFS pool for. If your source requires a Direct IO workload, then it will make sense to benchmark in that way, but if your workload is able to take advantage of the ZFS ARC, then we believe benchmarking with it disabled is like intentionally handicapping yourself. Hope that helps!