[Samba] Performance regression of Windows clients?

[Giuseppe Lo Presti]

Hello Samba experts,

I'd like to revive an old thread from 2017, where a performance 
regression of the Windows client implementation of SMB was discussed 
(https://lists.samba.org/archive/samba/2017-August/210366.html).

We run a ctdb-managed cluster of Samba gateways to our distributed file 
system, and we have performed similar tests with the latest production 
Samba 4.12 release, using both Linux (cifs-utils 6.2-10) clients and 
Windows 10 native clients.

Our synthetic test simply "touches" and subsequently removes 100 files. 
The network captures show Linux clients just performing the required 
operations, including `SMB2_FILE_INTERNAL_INFO` and 
`SMB2_FILE_BASIC_INFO` requests appearing exactly 100 times each, 
whereas Windows clients perform in addition several expensive 
`SMB2_FIND_ID_BOTH_DIRECTORY_INFO Pattern: *` queries on the destination 
folder as well as on its parent (we counted 600+ such requests). The 
latter cause a remarkable performance loss, especially when the storage 
exposed via Samba is a network file system itself.

As the mentioned thread seems to have been moved to the samba-technical 
mailing list, was there any conclusion reached? In particular, we’re 
looking for suggestions on possible reconfigurations (if any!) of our 
Windows clients, as we don’t expect the Samba server can do anything to 
mitigate this behaviour. Or in alternative, is there an alternative SMB 
client for Windows (if it is at all possible to replace the native one)?

Thanks in advance for any suggestion on the topic and kind regards,

Giuseppe


--
Dr. Giuseppe Lo Presti
CERN IT/Storage
Geneva - Switzerland