Performance issue when indexing folder via SMB with a lot of files

[grepis]

I almost pulled trigger and ordered Unraid. But noticed issues with SMB share - slow performance with small files. I have tried to resolve it, but no success.

So configuration I am testing:
One cache pool with one drive (SSD), with only one share using this cache pool directly (without moving to array or writing to array).
I uploaded folder (just bunch of games -> tons of folders, tons of files). It is around 35k files, and 317GB. Speed is sort of ACCEPTABLE (113MB/s, sometime when a lot of small files in get down to 50MB/s) and match my 1Gbit network.
Now I am trying to delete this folder using Total Commander (Windows) over SMB share. As first step it calculate total size of all files in folder. And this operation (indexing) takes 40seconds. Before we jump into discussion of network problems, file system problem, OMV on the same machine with the same hardware can do this operation just in 8s (same disk, same BTRFS). My Synology with BTRFS gives 7s on the same content. So definitely something wrong with Unraid configuration.
I tried various settings in SMB - no improvement. Only improvement is when I share whole disk. Then it is around 11s.
Right now I do use OMV, but deployed Unraid in VM. And still getting same results. So before I commit to OMV, I still want, with help of gurus, try some options to resolve.

PS. I do like Unraid UI and real-time parity calculation (and ability to use cache) and move content to Array overnight, but this performance issue is showstopper for me.




 

[JorgeB]

Try using an exclusive share (or disk share), both will bypass FUSE, which is known to impact performance a lot with small files.

[grepis]

Yeh, it is significant improvement - I am getting 11s. Is there anything that can be tuned in FUSE layer? Because OMV with mergerfs don't induce any performance problems. Btw, what technic Unraid use to merge folders from various disks?

[JorgeB]

15 hours ago, grepis said:

Is there anything that can be tuned in FUSE layer?

Not really at the moment, but once we get to kernel 6.9, FUSE passthrough mode should be available, that's expected to increase performance by a lot, similar to native.