Re: striped LV and expected performance

Dne 21. 07. 25 v 18:42 matthew patton napsal(a):
> don't use LVM striping. Use MD/DM striping. Better yet, stripe at the hardware 

lvm handles both --type striped, type raid0

> controller level if you can. Generally speaking a 32kb or 64kb interleave or 
> "strip size" is ~optimal for any use case (yes, even SSD) that isn't LARGE 

Many/most SSDs do internal striping - thus they go with relatively large 
optimal io sizes (256K, 512K)

> streaming reads. If the hypervisor is VMware, it allocates in 1MB extents and 
> it will also interfere with I/O scheduling.
> 
> If you want to cut the hypervisor out of the loop, PCIe pass-thru is your only 
> recourse. And again, you can attach no less than 3 up to about 5 (SSD) drives 

This purely depends on individual hw capabilities.

> to a controller before you saturate the PCIe bus. Also consumer-grade (aka 
> junk) SSDs have miserable write endurance and write speed after a preliminary 
> (ie fake) boost. If you care about your data and speed that doesn't go to zero 
> ONLY buy enterprise grade SSDs and if you're doing a lot of writes, minimum 3x 
> DWPD class though if you're going to attempt parity-RAID, you'll want 10x DWPD 
> disks.

For boosting write performance there is probably a better option to simply use 
some nvme caching with  dm writecache or something like that.

But as said - trying to do any kind of storage optimization inside VM is 
largely destined to fail - as there is not good enough knowledge about disk 
topology - since VM fakes whatever it can...

Regards

Zdenek