Re: Exim 0.54 feature requests

Página superior
Eliminar este mensaje
Responder a este mensaje
Autor: D. J. Bernstein
Fecha:  
A: exim-users
Asunto: Re: Exim 0.54 feature requests
> If you route all the remote addresses you will normally find a
> *large* number will have common MX hosts


Unfortunately, this is not the case.

For example, here are the figures for one mailing list with 231
recipients. There are 547 relevant MX IP addresses. A greedy algorithm
for doing what you propose (I wasn't able to improve on this with
simulated annealing) would send

8 through relay2.uu.net
6 through sun3.nsfnet-relay.ac.uk
4 through punt.demon.co.uk
3 through each of 3 hosts: mail1.best.com, lena.s.bawue.de, deneb.dfn.de
2 through each of 9 hosts: davinci.pixi.com, eunet.eu.net, etc.
1 through each of 186 hosts

for a total of 201 copies of the message. The subsequent relaying
includes more than 30 copies across wide-area links---I stopped counting
after that.

To summarize, your proposal loses bandwidth, even if we ignore the extra
RCPT commands and Received lines and TCP setup-teardown and other
``minor'' costs. It also worsens latency. It also adds some completely
unnecessary load to the secondaries. In short, it's not a good idea.

> } (Some people use bulk_mailer, which achieves parallelism at the
> } expense of disk space by sending separate copies of the message.)
> I have to add that this is a very similar approach to qmail, which
> acheives parallelism at the expense of bandwidth.


That's not what I said; you're replacing a serious concern with a
frivolous concern.

bulk_mailer drastically increases the _disk space_ used by mailing-list
messages. Instead of one copy of the message for a week, you have (say)
twenty copies for a few hours, and ten copies for more than a day, and
five copies that last the whole week. Mailing-list messages, in turn,
represent a big part of the standing queue at typical sites---not
because there are so many messages, but because each one lasts so long.
Disks are cheap, but not cheap enough for a massive increase in the size
of the mail queue.

In contrast, the _bandwidth_ increase that you're talking about, where
more than one copy of a message is sent to one host, is almost
completely unnoticeable. It's under 1% of the mail traffic at most
sites, which means under 0.07% of the Internet traffic. Go to sleep for
six hours and your bandwidth use will increase by more than that.

---Dan