On 19/7/25 04:08, Zdenek Kabelac wrote: > Especially in virtual world you often get already some 'provisioned' space > which nowhere near to anything physical. Try to stripe such drives makes > no practical sense.
Been following this discussion with interest. I think the above sentence sums up the entire discussion. A few more things I want to add. 1. Adding controllers, even in a physical machine, with individual disks and stripe across them is by far the worst thing you can do. As controllers, especially the high end ones, do have a huge amount of logic and knowledge of the underlying topology, it can, and will optimise read and write operations to align with head positions and other characteristics in case of HDD's. On SSD's it can/will align to disk firmware settings and instructions. These are in 99% of all cases not available in the OS as they will be masked by the controller. If you add more controllers these will all operate independently and have no insight in the IO pattern and can therefore only operate on a single io. As there are then dependencies on different controllers and different disks, all controls will need to come from the OS side. 2. Using guest IO controls. (Such as direct io parameters). These kind of parameters will almost always be ignored by the underlying hypervisor as that will by itself needs to determine the IO parameters in conjunction with other io activity of other guests on that hypervisor. 3. Unreal cache optimisations. Using dd is by far the worst option to use for performance tests as it will never (Ok, almost never) align with real workloads. If you use dd for performance test you will find that this will backfire in most cases when a normal workload is applied. The main reason is that dd will always have a sequential workload unless you start a large amount of dd instances to the same disk at once with different offsets. Even then you will see an obscure number coming back. 4. Use a tool that can record and replay io workloads such as SWAT and VDbench. There are a massive amount of dependencies when it comes to storage performance profile and optimisations. You may get really disappointing outcomes when systems are configured with settings that come out of an unrealistic profiling exercise and a real world workload will get deployed on them. cheers Erwin
