Environment:
- TrueNAS Fangtooth 25.04.2.6
- Debian-based, EPYC 9124, 256GB RAM, 10x NVMe RAIDZ2
- ~60 SMB clients (Windows), AD joined
- Multi-protocol shares (SMB + NFS) — oplocks disabled by design
Problem:
DCC applications (Maya, Houdini, Blender) issue small sequential reads (~16KB per SMB2 Read Request) when opening large scene files (5–11GB).
Wireshark confirms ~700,000 individual read requests for an 11GB Maya
ASCII file, with effective network utilisation of 250–300 Mbps on 1GbE
despite wire speed being available.
- QNAP QTS NAS loaded the same file in ~233s.
- TrueNAS Fangtooth loads it in ~384–422s.
- SMB throughput test confirms both deliver identical wire speed (~110 MB/s).
- The difference is QNAP’s default server-side read-ahead (use sendfile, socket options SO_RCVBUF/SO_SNDBUF).
What we tried on Fangtooth:
- socket options via share auxsmbconf — BLACKLISTED
- use sendfile via share auxsmbconf — BLACKLISTED
- socket options via global smb_options — BLACKLISTED
- use sendfile via global smb_options — BLACKLISTED
- readahead:offset / readahead:length via auxsmbconf — ACCEPTED
- vfs objects (to load readahead.so) via auxsmbconf — BLACKLISTED
readahead.so exists at /usr/lib/x86_64-linux-gnu/samba/vfs/readahead.so
The params are accepted but the module cannot be loaded because vfs objects is blacklisted.
Current smb4.conf [global] has zero performance parameters set — confirmed by grep for sendfile|socket options|read raw|aio|receivefile returning empty.
Questions:
- Is there a supported method in Fangtooth to load
vfs_readaheador configure socket options for VFX/CGI read workloads? - Is the blacklisting of these parameters intentional with an equivalent replacement planned, or is this a regression?
- For multi-protocol NFS+SMB shares specifically, is there any supported read-ahead mechanism available?
Not looking to bypass security hardening — looking for the correct supported path to tune SMB read-ahead on Fangtooth for high file-count, large sequential read VFX workloads.
Thanks,
Bhavik