Just a few days ago, Microsoft formally announced that SQLIO has now been fully deprecated, in favor of DiskSpd.
While that tool was great in the days of spinning rust (read: hard drives), it’s time to move on. Let’s remember that SQLIO really had nothing to do with SQL Server. It was a storage benchmarking tool, much like IOmeter, CrystalDiskMark, ATTO, HDTune, and probably 20 or 30 more that I won’t bother mentioning.
Nowadays, Microsoft is advocating for the use of DiskSpd, which happens to have a much more appropriate name, in my humble opinion – not that I like the “Disk” part either 🙂
In this post, I would like to remind you that both SQLIO and DiskSpd are great for performing basic plumbing tests, where you are trying to find the maximum data throughput you will obtain from your system. My colleague and friend Lou Lydiksen has blogged about this previously, and has expressed his fondness of fio for performing these sort of testing on Linux/Unix systems.
For a more realistic simulation of your production environment, though, you will need to use something else. Mimicking your actual workload with DiskSpd will be a very, very time consuming task, if at all possible. The reason is that SQL Server never actually uses a fixed IO block size. It can issue IOs as tiny as a few hundred bytes (typically your sector size for transaction log flushes), or a large as several megabytes in size (the case of backups and restores). You can certainly try to kick off multiple DiskSpd threads at different IO block sizes and see what kind of behavior you’d get from your system, but in the end this continues to be just a synthetic test.
In data reducing arrays like the Pure Storage FlashArray, there’s nothing like testing with your actual workload, and looking at how your actual production IO patterns behave on your target system.
So, what’s a great tool for this? You might wonder. My favorite, by far, is Distributed Replay. It might not be the friendliest tool in the world (take note, Microsoft!) but it certainly is one of the best (if not _the_ best) tools for determining whether a target system is suitable for a given workload. Note how I didn’t specifically say that it’s great for testing storage subsystems; it is, but that’s one of the many scenarios where you would use Distributed Replay. You should use every time you do things like:
In my next blog post, I’ll provide some plumbing test examples using DiskSpd, and then we will move onto how to work with Distributed Replay. Stay tuned!