[Mailman-Users] High usage -- making Mailman "nice"

[John Dennis]

On Thu, 2005-12-08 at 11:10 -0600, Caylan Van Larson wrote: 
> Morning,
> 
> We're running MM v2.1.4-4, we're experiencing such high usage on our  
> dual xeon server when mailman sends to 4000 users that port 25  
> becomes unresponsive.  is there a way to "nice" the mailman delivery?

It sounds like your MTA is inefficient in some manner, is not a huge
load for a dual server backed with a high bandwith connection. How many
client connections can it simultaneously handle? Any chance you've got
DNS delays (especially reverse lookup)?

There are ways to control how many recipients are sent in each SMTP
transaction. SMTP_MAX_RCPT = 500 is the default in mailman, defined in
Defaults.py, overridden in mm_cfg.py

If your MTA can only handle one, or a small number of client connections
then each connection will be busy handling SMTP_MAX_RCPTS and other
client connections will queue up. If you have 4,000 recipients then you
have the potential to tie up 8 SMTP client connections. If you raise
SMTP_MAX_RCPTS you may exceed your MTA's limit on number of recipients,
but you'll consume few connections increasing the pool of available
connections. If you lower it you'll increase the total connections
needed, but your MTA should take connections round robin allowing other
to jump in. By all means verify how many client connections your SMTP
can handle.

If you have VERP or personalization turned on mailman will have to
initiate one SMTP transaction per recipient (are these features turned
on?). This is inefficient because there is one recipient per SMTP
connection. It will take the SMTP much longer to process because of the
higher overhead, but on the other hand there should be plenty of
opportunity for other clients to obtain a connection from the SMTP
connection pool because mailman will cycle through connections in a
serial manner leaving the rest of the SMTP client connection pool
available for other SMTP clients.

HTH,
-- 
John Dennis <jdennis at redhat.com>