Performance issue with linux kernel 5.4.0 & rclone mount

#1

I ran into a performance issue with rclone v1.50.2 & the new linux kernel 5.4.0

  • rclone mount (google drive, crypt)
  • measure transfer speed of a plain file copy operation (either cp/time, dd, pv)
  • linux kernel 5.3.8 or lower: speed well beyond 300mbit/s up to 900mbit/s (gigabit line)
  • linux kernel 5.4.0: speed max 20mbit/s

same mount command/params, just rebooted with another kernel several times, same result. using "rclone copy" instead of the mount always works as expected with up to 900mbit/s

Can anyone else confirm this behaviour?

Ubuntu 19.04 disco

rclone v1.50.2
- os/arch: linux/amd64
- go version: go1.13.4

rclone mount driveCrypt: /mnt/drive -vv --progress --read-only --allow-other --rc --umask 0222 --attr-timeout 1s --dir-cache-time 672h --vfs-cache-max-age 675h --vfs-read-chunk-size 8M --buffer-size 0 --poll-interval 1m0s --bind MY_IPv6_IP --vfs-read-chunk-size-limit 64M --tpslimit 7 --tpslimit-burst 10 --use-mmap

1 Like
#2

There is some discussion of a go and linux 5.4 kernel issue: https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.2-Go-Register-Corrupt

I'm not sure what you are seeing is related and I thought it only came into effect with go-tip, but you've compiled your rclone with go1.13.4

So assuming that that isn't related it looks like a FUSE problem since rclone copy does not show the problem.

Are there any errors in the log file? It might be worth running with -vv to see if there are any clues.

Also give the latest beta a go which has the latest fuse libraries and a number of fuse tweaks.

#3

I had some issues with the newer kernels and my goal was stability so I just moved back to Debian and ran the 4.9 kernel for now until everything gets fixed up.

There is a fun part of run bleeding edge, but the problem is finding bugs and I have no need for that. I saw a few posts here and there regarding TCP performance and perhaps that's what you are hitting.

#4

I'm seeing similar issues but haven't dug into it any deeper than my plex acting like it's starved for data. Spin for a long time, play a second, Spin for a long time, repeat.

I roll back to kernel 5.3.12 and it's fine.

Edit: I just wanted to let you know that other people that I work with are also seeing the same behavior...rolling back to 5.3.x also fixes it for them.