Re: slapd memory usage

[Zev Weiss]

Howard Chu wrote:
A. Schulze wrote:

Howard Chu:

Any idea why the memory usage is so different?

If the only difference is that you set the open file limit to 1024, 
then it sounds like whatever your default file limit is is much 
larger.

Hello Howard,

yes, it's unlimited by default. Tons of other daemon also run without 
this limits here.

But in contrast: all other daemons don't let explode the memory 
usage.
Maybe, it worth to find the difference?

That *is* the difference. slapd allocates an array of connection info, 
one slot per file descriptor. Running with "unlimited" files is clearly 
a bad idea here.

In general, running with larger limits than you actually need is a bad 
idea. This is elementary system administration.


Couldn't one s/running with larger limits/consuming more resources/ and 
s/system administration/software development/ and produce an equally 
valid argument though?  (If anything, larger-than-necessary limits seem 
the more justifiable of the two to me -- it allows for future growth, 
which can be hard to predict.)

More to the crux of the matter: why does slapd need to preallocate all 
OPEN_MAX possible connection info records at once instead of dynamically 
as connections are actually created?

I've actually been bitten by the inverse problem when slapd ran up 
against my distro's default FD limit (causing no small amount of grief 
to the various client systems on my network).  I of course remedied this 
by cranking up said limit a fair amount, but the number I chose was 
basically just a hand-wavy, seat-of-the-pants guess at something that 
would last a while before I had to tweak it again, and thus means that 
slapd's going to be sitting on significantly more memory than it really 
needs.

So as an administrator I'm left with the question of how to balance 
slapd's file descriptor requirements against my desire to not have it 
tying up a bunch of memory it's never actually going to use.  It seems 
like that balancing act would be a lot easier if slapd could dynamically 
allocate memory for connections.


Zev Weiss

(Apologies for any strangeness with the formatting/headers of this 
message; I wasn't subscribed to the list when the message quoted above 
was sent and hence have sort of manually synthesized this reply via 
copy/paste from the mailman web archive.)