Re: Mail delivery software [was Best List Software for SCO]

[Matti Aarnio <mea @ utu . fi>]

People appear to be learning again, that the problem on lists is not
that much of what MLM we run, rather what kind of mailer we have..

> From: Keith Moore <moore@cs.utk.edu>
> Date: Thu, 28 Sep 1995 15:20:53 -0400
> Subject: Re: Best List Software for SCO 
> 
> > Majordomo has severe scalability problems. I don't know how many
> > lists you intend to run, but Majordomo, in combination with
> > bulk_mailer, periodically sends our dedicated email server into the
> > stratosphere with a load average of 40 to 50 when four or five
> > messages to large lists (400+ users) arrive within moments of each
> > other.
> 
> I'm not surprised.  Even without the Perl-related problems, the
> current version of bulk_mailer will cause the creation of several
> processes (instead of just one) each time a new message arrives.  It
> does get the message out faster, but you pay a price for this in
> having all of those processes fired up.

	Ack, and the process startup is often heavier, than having
	a set of processes waiting to be activated.
	(See how NCSA HTTPD works - a bunch of processes with one
	 master who does accept() connections, and then sends the
	 newly arrived file-descriptor to a free worker.)

> A big part of the problem is that sendmail isn't really designed for
> high-volume mail.  So for instance, it doesn't manage a limited number
> of 'worker bee' processes to send outgoing mail; it fires up a
> separate process every time you want to send a message out, and leaves
> it to a single queue process (worker bee) to periodically "clean up"
> any messages that weren't delivered on the first pass.  For a
> high-volume site, that's too many processes on the front-end, and not
> enough processes on the back end.

	Right.   I have been hacking at an UNIX-mailer called
	ZMailer, which in its new form runs transport channels
	(back end) in process pools, which pools can contain
	1..N programs.

	The front-end in it is a set of router processes (1 to N)
	that can be run in parallel, and that are started at the
	systen startup.

	For the interested people, see:
		ftp://ftp.funet.fi/pub/unix/mail/zmailer/
	(Look for zmailer-2.99.18 or latter)
...
> ------------------------------
> 
> From: Info-LabVIEW List Maintainer <info-labview-request@pica.army.mil>
> Date: Thu, 28 Sep 95 16:09:20 EDT
> Subject: Re:  Best List Software for SCO
> 
> >bulk_mailer is really just a band-aid for sendmail, and it can only
> >help so much.  If you want to do lots of volume on a single machine,
> >you might want to consider a different MTA.  Check out Innosoft's
> >PMDF, for example.
> >
> >Keith Moore
> 
> Any comments on MMDF in this vein?

	It is one of the programs that PMDF used as a model,
	and not a bad one at that.

	ZMailer used it as a model too.

> We run it here and I'm quite pleased with how it works. Of course, my lists
> aren't really high volume in the sense that folks are talking here, with
> about 30k/day going out - 350 users getting it as individual msgs and 1500
> getting a daily digest.
> 
>                   Tom Coradeschi, Info-LabVIEW List Maintainer
>                       <info-labview-request@pica.army.mil>
>           http://k-whiner.pica.army.mil/info-labview/info-labview.html
> 
> ------------------------------
> 
> From: Eric Thomas <ERIC@SEARN.SUNET.SE>
> Date: Thu, 28 Sep 1995 23:14:13 +0100
> Subject: Re: Mail delivery software [was Best List Software for SCO]
> 
> For what it's  worth, here are some  figures for LSMTP. The  data is from
> yesterday's run (24h).
> 
> Machine #1: 275MHz Alpha, 128M
> 
> Total: 1,474,627 deliveries
> 
> Average delivery time:            1557 sec (25:57)
> 50% delivered in less than:         50 sec (00:50)
> 90% delivered in less than:        230 sec (03:50)
> 95% delivered in less than:        320 sec (05:20)
> 96% delivered in less than:        354 sec (05:54)
> 97% delivered in less than:        401 sec (06:41)
> 98% delivered in less than:        462 sec (07:42)
> 99% delivered in less than:        587 sec (09:47)

	Eric is showing impressive figures, but what do they mean ?
	That is, assuming a list where there are multiple recipients
	at some host, are all those recipients at that host counted
	as one, or as individuals ?

	There is a big difference in handling 1000k messages with
	1 recipient each, and 10k messages with 100 recipients each..

...
>   Eric

	/Matti Aarnio <mea@utu.fi>