> I wonder if someone can tell me about the capacity of a SUN
> box with Solaris 2.4 that would be needed to manage many lists
> where a bunch might have thousands of subscribers.
>
The number of lists and the number of subscribers doesn't matter.
What matters is traffic volume.  For a first order approximation
you need to estimate the average number of postings per time period
and the average number of recipients per posting.  Why?  The resource
consumption will be almost entirely that of sending the messages.
The LISTSERV resource consumption in generating the messages will
be down in the noise.
 
A better estimate can be obtained from the average number of postings
per time period and the average number of unique destination host
names per smtp message generated by LISTSERV.
 
Things start to get complicated if you have time constraints.  I.e.
90% of all recipients *must* receive the message within x minutes.
If you are using sendmail the only easy way to force short delivery times
is by reducing the average number of recipient domain names in each
smtp message generated by LISTSERV.  That results in an increase in
the number of sendmail processes running and therefor an increase in
the hardware resources required to keep it all going.
 
Our experience leads me to the estimate that you can generate and
send about 250K messages per day with a 60MHz Sparc 20 with 128MB
of RAM.  As you go past 350K messages per day you'll need to upgrade
to a dual processor Sparc 20 with 256MB.  You're probably safe in
translating the CPU capacity needed to other chips by using SpecInt92
numbers.  This is for large recipient count smtp messages.  Again
as a first order approximation halving the average recipient count
per message will double the number of sendmail processes and double
the hardware resource requirements.  I've been using MAXBSMTP=1000
to get the numbers mentioned above.  The 90th percentile delivery
time for messages (not recipients) is usually just short of two hours.
By using MAXBSMTP=64 as an experiment I was able to pull that figure
down to just short of 30 minutes, but the system was grossly overloaded
with LISTSERV developing *large* input queues.  Those large input
queues defeated the whole process as what matters to the users is total
message turnaround time not just sendmail turnaround time.
 
It is possible to do somewhat better if you want to take the time and
make the effort to become far more of a sendmail expert than I am or
ever want to be.  Oh, and I'm not a sendmail hater, it is very good
for very many things.  High volume bulk mail delivery just isn't one
of them.
 
Syracuse is in the process of moving our LISTSERV operation off of the
Sun box to Windows NT.  Why?  L-Soft's LSMTP mail transfer agent will
just blow the doors off of sendmail.  Back in February we beta tested
LSMTP on a P90 and it was delivering to 400K recipients per day with
a 90th percently delivery time of less than 10 minutes with MAXBSMTP=1000
instead of the 120 minutes which is normal with sendmail.  On top of that
a 96MB P133 will run both LISTSERV and LSMTP and won't cost you anywhere
near the price of a dual processor Sparc 20.  It probably won't cost as
much as the 256MB of RAM that's on our Sparc 20.
--
Glenn A. Malling <[log in to unmask]>
Syracuse University Computing Services   +1 (315) 443-4111
220 Machinery Hall  Syracuse, New York 13244-1260
 
Postmaster for SUnix and LISTSERV.