dovecot - Feb 2013 - Imap-login service count/limits

Hello Timo,

I'm trying to configure dovecot-2.1.15 master correctly regarding imap-login
limits. I want to use service_count=0 and I'm starting with something like
this
:

#default_process_limit = 100
#default_client_limit = 1000

service imap-login {

[...]

  service_count = 0
  process_min_avail = 8
  vsz_limit = 1024M

}


-> Does it make sense to set a process_limit if service_count = 0 ? why ?
Doesn't service_count = 0 mean that no additionnal imap-login process would
be
created (except for process_min_avail) ?


-> Does it make sense to set process_min_avail <> 0 if service_count =
0 ? why
? and if yes, what's the rule of thumb to find the correct value according
to
the number of connections estimation ?


-> Assuming every client is configured to use STARTTLS and opens by default 5
   imap connections, each client would consume 5 of the 1000
default_client_limit,
   right ?

-> is 1024M too high for vsz_limit and may expose to memory leaks ?

In other words, what limit should be raised to scale a lot a connections ?

Thanks

-- 
Thomas Hummel 	    | Institut Pasteur
<hummel at pasteur.fr> | Groupe Exploitation et Infrastructure

On Tue, Feb 26, 2013 at 11:08:18AM +0100, Thomas Hummel wrote:
> -> Does it make sense
Sorry, I was unclear. Let me rephrase :

from my understanding of the doc, process_min_avail seems to have a slightly
different behavior depending of if service_count is 1 or 0. Is it correct ?

The doc states that process_min_avail in the case of service_count = 1 is used
to avoid latency and in the case of service_count = 0 to make sure not only the
one imap-login is used (if below the client_limit threshold).

Why would it make no sense to set (service_count=0) process_min_avail >
number
of cpu ? Because there is no fork/exec of another imap-login latency ?

Besides, why would process_limit be reached since service_count=0 means that
this one imap-login process handles an unlimited number of connections ?

Thanks

-- 
Thomas Hummel 	    | Institut Pasteur
<hummel at pasteur.fr> | Groupe Exploitation et Infrastructure

On Tue, Feb 26, 2013 at 02:04:07PM +0100, Thomas Hummel wrote:
> Besides, why would [...]
I think I get it, I might have been confused by the term "unlimited"
in "the
only useful values are 0 (unlimited)" in the sample config file comment.

As I see it now :

- service_count = 0 :

 . number of connections is not "unlimited" but limited by
client_limit (or
   default_client_limit) for each imap-login process first and then by
   process_limit (or default_process_limit)

 . since only when client_limit is reached a new imap-login is forked (in the
   limit of process_limit), pre-forking process_min_avail load-balance the work
   across all CPUs instead of staying on a single CPU until client_limit is
reached.

- service_count = 1

  . client_limit has no effect since 1 connection <-> 1 process
  . pre-forking process_min_avail avoid the fork-exec overhead/latency

Is this correct now ?

Thanks

-- 
Thomas Hummel 	    | Institut Pasteur
<hummel at pasteur.fr> | Groupe Exploitation et Infrastructure