The mailing list for listmasters using Sympa

[Elijah Saxon <address@concealed>]

I am very impressed with sympa's performance processing and delivering
mail. I am less impressed with wwsympa's performance. Even with fastcgi, I
have always found wwsympa sluggish, even on a 2 ghz machine with 1 meg or
ram and no load. Regardless, it beats the pants off any alternatives and
is acceptably sluggish.

Lately, however, I have been wondering about wwsympa scalability with a
large number of lists. Sympa may work beautifully with lists of hundreds
of thousands of subscribers, but what about wwsympa with thousands of
lists?

In particular, because the configs for each list are kept in memory (is
this correct?), will wwsympa become memory bound? Currently, our wwsympa
processes are running at 150 megs each, but that is steadily climbing.
Even now, we would quickly run out of our 2 gig swap if we did not kill
off wwsympa processes nightly.

I have a couple questions:

- is anyone running a box with over 5000 lists? how?

- do wwsympa.fcgi processes share memory? could they?

- how much work would it be to port to mod_perl?

- what do people think about simply making swap really really big? does
wwsympa routinely access all the memory it keeps, or are there parts that
will safely get swapped out with disuse?

- if you have multiple virtual domains, will splitting the fastcgi
directives into each domain improve performance? i guess the trade off is
less memory per process, but smaller pool of processes which can handle
any given request. i could be confused as to how this works. obviously, it
would be better to just put the domains on a different machine.

- what can be done to improve wwsympa performance? (other than increasing
the number of processing or buying faster and dedicated hardware). i have
found that once over 1000 lists, a file system with b-tree directory
storage, like reiserfs, is a must.

thanks,
-elijah
http://lists.riseup.net