Re: where to sync from?
On Wed, Mar 21, 2007 at 07:36:40PM -0300, Carlos Carvalho wrote:
> >A common test that I used to run is to run mtr to the destination site and
> >see if there is anything sticking out; and then run something like
>
> I tried mtr for the first time. Nice but not reliable(?!?!). From my
> machine (not in the same site as the mirror) to our gateway it reports
> 16-22% of packet loss!! This is nonsense, real packet loss is zero
> (tried ping -f to confirm).
>
> >wget -O /dev/null http://<site>/debian/pool/main/k/kde-i18n/kde-i18n_3.5.6.orig.tar.gz
> >And leave it running until a pattern is established. That particular
> >file is used simply because it's huge (~300 MB) so you won't run out
> >of bytes before being able to see a pattern; you can interrupt it
> >once it stops fluctuating, obviously.
>
> There's no pattern unfortunately. The rate varies between 250K/s to
> about 2M/s with the US. Over the full file our connection to
> ftp.us.debian.org (128.101.240.212) is faster: 1.31M/s, versus 1.04M/s
> to Europe. There are also less hops. I don't have access to
> ftp-master, of course.
See, that's where mtr came in handy. You may not have been getting actual
loss when directing ping packets via ICMP, but the UDP traceroute packets
with lower priority are getting lost. Packets get lost when the intervening
routers drop them in order to slow you down. mtr usually helps determine
which routers do that.
> >BTW, what's your /proc/sys/net/ipv4/tcp_wmem like?
>
> Now it's the default, 4096 65536 16777216. Some time ago we
> fiddled a lot with it and other parameters but after we put 8GB of ram
> we didn't see any difference any more.
With some network equipment, it seems to only make sense to increase the
default value (the one in the middle), they won't care for the maximum, so
you might wish to test that. For example, syncproxy.eu.d.o has been running
4096 3145728 8388480 for a while now, and this was set manually after
a bunch of test downloads (it was also before the new 2.6.17 defaults).
At one point I remember I pushed it too far, something like 14 times larger
than that old default, so the rate started to fluctuate differently - it
grew up until one point, and then dropped sharply. The reason was that my
massive throughput overflowed the buffers of an intervening network device.
And that device turned out to be - the switch my mirror was plugged into!
--
2. That which causes joy or happiness.
Reply to: