omg I expierenced kinda similiar issue first with my dedicated server and the rtorrent MergerFS Mount with an underlying Rclone mount and now also on my Homeserver using Plex directly on the Rclone mount and I just debugged it down to the same cause, unnecessary high IOwait. I already gave up until I just thought of checking here.
Linux deskmini 5.4.0-0.bpo.4-amd64 #1 SMP Debian 5.4.19-1~bpo10+1 (2020-03-09) x86_64 GNU/Linux
IOWait isn't a good measure in my opinion! If you use --async-read=true (the default) you are going to get IOWait, but you will get faster performance provided you don't see those failed to wait for in-sequence read messages. Those are what really kill the performance. If you don't want to see IOWait then set --async-read=false and it will all disappear along with some performance.
is i possible to find an option with --async-read=true without failed to wait for in-sequence read messages? Should I raise --vfs-read-wait higher than 50ms? Any Ideas?
I raised it till --vfs-read-wait 1000ms and still got the failed to wait for in-sequence read Message in debug log. (40 times in 1 minute under high load)
Hello everyone.
Running 1.51 on an Ubuntu server and have the exact same issue.
Few stupid question:
Why do I see IO Wait on Netdata but nothing on iotop -o ?
Is rclone writing constantly on the disk with this bug ? --> I have expensive NVMe, should I better downgrade or is it safe to continue run this version until there is a patch ? I mean I don't want this bug causing an infinite write on my NVMe until the patched version.
iotop is showing active disk utilization and what is consuming active disk IO.
netdata is showing IO Wait, which is a separate measure of a process waiting for disk IO to complete. Anything in IO would eventually show in iotop as being IO if that makes sense.
Why I don't have issues without messing with this async stuff? Sometimes I get 200 or more files open in the mount, and I haven't noticed any issues at all.
The issue was with stock 1.51 as that defaulted to turning on async reads and there were fixes that went into the beta to 'smooth' it out. It can also be the settings you have are reading such small increments, you are not seeing the issue.
You'd probably want to compare IOWAIT on 1.50.2 and 1.51 and see if you have any changes in it.
Oh I see. So now if I'm using rclone 1.51, besides using the beta, are there any other flags I can add to the rclone and mergerfs file to mitigate this issue, or is rolling back to 1.50 the only way?