Fri, 17 Nov 1995 00:48:02 +0200
|
On Thu, 16 Nov 1995 17:17:16 -0500 Paul Graham <[log in to unmask]>
said:
>We're one of the (apparently) few sites running listserv (1.8b) with
>sendmail (8.7.1) on sun equipment (under solaris 2.4). I'm interested in
>any experience tuning and or driving sendmail for the ~200k
>connection/day range.
Perhaps the following chart might help:
********************************************************
* Top 20 sites by volume delivered daily (NOV95, 1-17) *
********************************************************
Pos Recipients Site
--- ---------- ----
1 1301411 PEACH.EASE.LSOFT.COM (Windows NT 3.50 + LSMTP V1.0a)
2 281181 LISTSERV.SYR.EDU (Solaris 2.4 + LSMTP V1.0a)
3 267072 UGA (VM/ESA 2.2)
4 243617 UKCC (VM/ESA 2.2)
5 223687 UBVM (VM/ESA 2.2)
6 208609 LISTSERV.ACSU.BUFFALO.EDU (Solaris 2.4)
7 193839 NDSUVM1 (VM/ESA 2.2)
8 183964 AMERICAN.EDU (AIX 3.2.4)
9 176027 MAELSTROM.STJOHNS.EDU (OpenVMS V6.2 + LSMTP V0.1a)
10 174842 CUNYVM (VM/ESA 2.1.0)
(etc)
Syracuse used to run sendmail and may be able to help you. I also believe
AOL has a number of non-LISTSERV machines doing about 200k/day (although
it's mostly incoming mail so it might be a totally different situation).
>I end up with rather peaky use curves and it seems if I could keep the
>load at the average value I'd actually be able to deliver to 1,000,000
>recipients daily.
It's not that simple unfortunately. If one could keep the load flat
constant over a 24h cycle, even a small machine would do wonders. But
most of the mail comes during business hours and in general if your 24h
average is X deliveries per hour, your most busy hour will show about 3X
deliveries. Again roughly speaking, if your machine can do Y daily
deliveries on a flat 24h input flow model, you can process a Y/2 workload
with mediocre performance and a Y/3 workload with good performance. For
instance, back when I was running SUNET's traffic using sendmail, I had a
PC that could handle 360k/day provided that the mail came in smoothly
enough and there was no outage, etc. I was getting good performance with
my workload of ~100k/day (at the time) and my peak hourly rates were
12-15k/hour depending on the day. Sometimes I got 120k/day and this was
still working ok. I knew I could do 150-180k/day and this would still be
ok, but there's no way I'd be able to run a real workload of 360k/day. It
would have meant overnight delivery in the *best* case and the users
would have killed me :-) And then there's the queue build effect which
slows down everything, and of course the fact that the higher input load
during peak hours causes paging or otherwise decreases the absolute
efficiency of the delivery system during the day.
Anyway, 1M/day with a flat input model means about 333k/day usable. So
I'm not sure there's any major conflict between the theoretical peak and
the ~250k/day you are actually observing.
Eric
|
|
|