Network issues after upgrading to 11.1-U6

[rungekutta]

Upgraded to 11.1-U6 and ran into symptoms today when trying to transfer large files from MacOS (High Sierra) - highly erratic speeds, from bursts close to gigabit to running to a crawl, eventually stalling and timing out. I first thought SMB was the issue but had the same erratic behaviour also via scp. From a Windows 10 client, SMB was stable near gigabit (as I'm used to) but scp performance erratic.

Only obvious difference is the upgrade, and indeed after reboot into 11.1-U5 all problems went away and I'm more consistent near gigabit speed again irrespective of protocol and client, as I'm used to.

Setup: Supermicro X11SSM-F, 32GB ECC RAM, Intel i3, 6x 3TB in RAIDZ2. 2x onboard Intel NICs, one on the DMZ and one on LAN. FreeNAS running on bare metal. Using "Cubic" TCP congestion control, which in previous versions has given a good performance boost over default newreno, particularly for clients on WiFi.

Anyone seen anything similar and/or have any ideas? I don't really know how to troubleshoot this.

 

[aristotlejones]

I know this doesn't help, but that is the default I have experienced with FreeNAS when using any RAIDZ setup. I switched to a ZPool of mirrors a while back and at least is made it consistent write speeds even though it crushed my total disk space. BTW, what is "Cubic" TCP congestion control?

 

[rungekutta]

I don’t think the problem is RAIDZ2 unless something in the U6 upgrade has crushed performance. Never had problems keeping up with gigabit speeds (~80MB/s read/write) before, including in U5 that I’ve now rolled back to.

TCP congestion control controls the number of unacknowledged packets allowed in-transit. Default in FreeBSD is the newremo algorithm which is quite old and works best with homogenous networks with low packet loss. Cubic is more recent and more aggressively adaptive to different network characteristics (e.g. clients on both Ethernet and WiFi), and the default in Linux since quite a few years back.

Again this shouldn’t be the issue but mentioned for completeness in case U6 contains some bug related to this.

 

[dlavigne]

It's worth reproducing on U6, creating a debug, then creating a ticket at bugs.freenas.org so a dev can investigate.

 

[rungekutta]

Ok, I’ll do that as soon as I can find the time.

 

[rungekutta]

There is another report now on what looks like the same issue. Bit more information in that thread. Suggest we keep the discussion there.
https://forums.freenas.org/index.ph...ransfer-aborts-on-smb-share-mac-client.69553/