Extremely slow initial sync with no obvious bottleneck

Support
#1

I’m running syncthing on two machines: an Ubuntu 16.04 VM and a Macbook. There’s about 5 GB to sync in 110k files. It’s been going on for hours now and is 8% complete. rsync of the same data takes a couple of minutes, tops; didn’t time it exactly but it’s orders of magnitude faster.

Neither of the machines is starved on CPU (single-digit % used on the target Mac, about 30% on the Linux machine) or disk I/O (mostly zero on both sides) but still the network I/O is ~100 bytes per second. The speed seems to be about one file per second.

FWIW, calling /rest/db/status?folder=my-folder-id takes ~4 seconds every time. GUI is closed on both sides. I’m using syncthing v0.14.49

Anything I can do do find the bottleneck? (perf top on the Linux side shows the Go GC and LevelDB having lots of fun but it’s still far away from taking a single core, let alone all 12 threads).

Edit: the hosts see each other over local IPs, without a relay.

#2

Are the files in a single folder? I suspect the database management overhead on small files is causing this.

#3

No, there are about 12k directories (that I could also watch get synced one by one before file sync started).

Can I speed up the DB somehow? I can live e.g. without fsync() after every write if that makes the sync fast (when the DB gets corrupted, I can just nuke and resync, but not really if it takes two days for 5GB).

#4

I mean syncthing folders, not actual folders.

#5

You can’t disable fsync, but that could also contribute to the slowness, I guess you could check with strace which syscalls it spends most time in.

#6

Oh, in that case yes, it’s a single folder.

It’s mostly futex() and pread64(), both are pretty fast though:

https://pastebin.com/qmmMjFPt

#7

Can you run with STTRACE=model env var to see what its doing? Is it actually making progress just slowly?

#8

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.