[mirror-admin] Rsync/SSH/Tar file transfers
Matthew Galgoci
mgalgoci at redhat.com
Wed Dec 7 21:14:56 EST 2011
> Using dstat while during the transfer I see an average ~100Mb/s of
> traffic at any given time. A FAR cry from the full 1000Mb/s they're
> connected at. The speed did burst to ~170Mb/s on single large files.
>
> rsync -avP server.com:/huge/web/2011-12-06 /tmp
> sent 3101025 bytes received 15370824277 bytes 13732849.76 bytes/sec
>
> Rsync reports ~109Mb/s average transfer throughput
>
> #2) Raw Rsync
>
> Right away I can see a difference in the transfer speed. Average is
> sitting around ~200Mb/s to ~250Mb/s. It bursts up to 400Mb/s on single
> large files. Disk is definitely a limiting factor, the transfer is more
> bursty.
>
> rsync -avP rsync://backup@server.com/incremental/web/2011-12-06 /tmp
>
> sent 3072973 bytes received 13688001486 bytes 18291348.64 bytes/sec
>
> ~146Mb/s doing raw Rsync. Probably because it's so bursty.
>
> #3) Tar + Netcat
>
> Started out SMOKING fast. The disk write actually peaked ABOVE 1000Mb/s
> at the beginning! Transfer isn't as bursty, and it has averaged out
> around 90Mb/s. Transferred 15409176735 bytes in 1010 seconds, which is
> ~122Mb/s.
>
> time tar cf - 2011-12-06/ | nc client.com 7000
>
> In short, rsync using the daemon was the fastest method to transfer
> files. Almost 20% faster in fact.
Scott, I've seen rsync abuse the tcp protocol in unspeakable ways for
performance gain. It uses zero windows as a busy wait, and I'm sure it
does other tricks to squeeze every single bit of performance out of a
connection.
--
Matthew Galgoci
Network Operations
Red Hat, Inc
919.754.3700 x44155
------------------------------
"It's not whether you get knocked down, it's whether you get up." - Vince Lombardi
--
More information about the Mirror-admin
mailing list