On Thu, 2007-10-25 at 10:59 +0200, Petter Reinholdtsen wrote: > In a simple test to see how libnss-ldapd handled several lookups at > once (trying to figure out how many connections I would get to the > LDAP server, I ran several of these commands: > > (ls -lL /home/&); ls -l /proc/$(cat /var/run/nslcd/nslcd.pid)/fd/
nslcd currently starts 5 threads in which each thread opens it's own connection to the LDAP server. The communication between the NSS part and nslcd part uses an I/O timeout value of of upto a couple of seconds (both NSS part and nslcd have separate timeouts for reading and writing). These values are currently hard-coded in the source. Any improvements to the values or mechanism are appreciated. > This resulted in the kernel killing nslcd because the machine ran out > of memory. I am testing on a xen client with 128 MiB RAM. The reason > seem to be that nslcd is using a lot of memory while processing the > request from this command. The /home/ directory contain 305 users (or > symlinks to users) home directories. > > The nslcd process started at 47/2 MiB (VIRT/RES) and increases to > 78/31 after one invocation of the command listed above. Here is a > table of how much is increases: > > Runs VIRT RES > 0 47 2 > 1 78 31 > 2 110 65 > 3 140 95 > 4 (killed) > > Is there anything more debug information I can provide? There are a number of known memory leaks in the code. I'm working on a better memory management model for the LDAP results (you have to be very careful about free-ing memory you got from the LDAP library). I have already identified and fixed a couple of leaks and I'm still looking for new ones. I will probably make a new release soon that contains these fixes. -- -- arthur - [EMAIL PROTECTED] - http://people.debian.org/~adejong --
signature.asc
Description: This is a digitally signed message part
