[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Processed: destruction of round-robin functionality is fucking up our mirrors and making Debian suck for many people, hence fixing this is a release-critical "wish"



On Sun, Dec 16, 2007 at 03:45:37AM +0100, Josip Rodin wrote:
> After around 11 hours, we've had:
> * villa 4.29 MB/s
> * lobos 3.91 MB/s
> * steffani 14.86 MB/s

The rule9 prediction was:

A: 000.000.000.000-127.255.255.255: steffani, villa, lobos
B: 128.000.000.000-191.255.255.255: steffani
C: 192.000.000.000-212.211.131.255: villa, lobos
D: 212.211.132.000-212.211.132.127: villa
E: 212.211.132.128-212.211.132.255: lobos
F: 212.211.133.000-255.255.255.255: villa, lobos

Class A is pure round-robin, so we can ignore rule9 and assign 1/3 of its
traffic to each host.

The difference between villa and lobos is aiui due not only to the difference
between the D and E IP ranges, but also because the round-robin ordering of
the hosts is [lobos, steffani, villa], which means that since rule9 happens
after round-robin, you get orderings:

	lobos, steffani, villa -> [lobos, villa], steffani
	steffani, villa, lobos -> [villa, lobos], steffani
	villa, lobos, steffani -> [villa, lobos], steffani

A/3 + B               = 14.86 MB/s (steffani)
A/3 + 2C/3 + 2F/3 + D = 4.29 MB/s  (villa)
A/3 + C/3 + F/3 + E   = 3.91 MB/s  (lobos)

If you assumethe 212.211.132.0/24 traffic is negligible, and thus D = E = 0,
then subtracting lobos from villa gives:

   C/3 + F/3 = 4.29 MB/s - 3.91 MB/s = 0.38 MB/s

And thus filling in for lobos, we get A/3 = 3.91 MB/s - 0.38 MB/s =
3.53 MB/s.

Going back to steffani, that gives B = 14.86 MB/s - 3.53 MB/s = 11.33 MB/s,
and we thus have:

    A     = 10.59 MB/s
    B     = 11.33 MB/s
    C + F =  1.14 MB/s
    D     =  0    MB/s (by assumption)
    E     =  0    MB/s (by assumption)

Which gives us 23.06 MB/s which was the total of what we started with, yay.

Note that 192.168.*.* addresses are in class C, so can only possibly make
up just under 5% of our traffic, which seems pretty negligible. 10.*.*.*
addresses are in class A, and 172.16... class addresses are in class B. I
would've thought there wouldn't be significantly more of those than for
the 192.168.*.* private addresses though.

Anyway, hope that's of some use.

Cheers,
aj

Attachment: signature.asc
Description: Digital signature


Reply to: