Comment 2 for bug 1198180

To clarify: we have about 200-300 containers running simultaneously, and the average lifetime of a single container is 2-10 mins; the network interfaces are also destroyed when the container goes down. So, in a specific point in time, we have about 200-300 network interfaces, and they're created and destroyed dynamicly (with lxc-start and lxc-stop).

strace'ing only mmap and unmap for init (-p 1) you can see, that memory is mapped for EVERY event, but only parts of it are unmapped:

[pid 3058] 08:43:07.444810 mmap(NULL, 677, PROT_READ, MAP_SHARED, 11, 0) = 0x7ffe40a7e000
[pid 3058] 08:43:07.444942 munmap(0x7ffe40a7e000, 677) = 0
[pid 3058] 08:43:07.446329 mmap(NULL, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f88f1b38000
[pid 3058] 08:43:07.446604 mmap(NULL, 26711, PROT_READ, MAP_PRIVATE, 3, 0) = 0x7f88f1b31000
[pid 3058] 08:43:07.446806 mmap(NULL, 3925208, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x7f88f1559000
[pid 3058] 08:43:07.446856 mmap(0x7f88f190d000, 24576, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x1b4000) = 0x7f88f190d000
[pid 3058] 08:43:07.446965 mmap(0x7f88f1913000, 17624, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7f88f1913000
[pid 3058] 08:43:07.447137 mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f88f1b30000
[pid 3058] 08:43:07.447200 mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f88f1b2f000
[pid 3058] 08:43:07.447250 mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f88f1b2e000
[pid 3058] 08:43:07.447490 munmap(0x7f88f1b31000, 26711) = 0

We do not have any custom upstart jobs, so only /etc/init/network-interface*.conf are run. We have also machines, where init uses more resources (after creating/destroying about 200K containers) :

    1 root 20 0 978m 951m 1360 S 77 0.7 3410:06 /sbin/init

it uses more, and more with every event served - and no, the memory is NOT freed after shutting down every container. Also, after some time, init uses almost one core (77% above), and it gets sluggish -i takes more time to serve new events. The only way for us right now, is to reboot the machine.