gthread_aio runs out of fds after some time under load.
After setting network backend to ghtread_aio under load of ~150-200 concurrent connections with large file download site we run out of fds :(
I'm not so well skilled to debug it myself but after modifing network_gthread to TRACE srv->cur_fds I see that number incrases but decrase is very rare :( mainly it is decrased for small files and as time is going cur_fds is higher and higher and we have still 100-200 connections active.
After one day we moved to ~16000 'used' fds which is not true but tomorrow server will be out of free fds (according to counter) and will stop.
Some facts about our download site:
1. Mainly big files
2. unlimited connection ratio (mod_evasive not used)
3. mod_secure_download used.
4. users are downloading via download menagers multiple splitted files (as there is no connections per IP limit).
Problem also exists for other aio network backends which counts used file descriptors.
Problem is quite big because if we run out of FDS server stops and waits for restart (as used fds count can't decrase).
Of course I will provide more info on request and I'm ready to run some tests.
Also available in: Atom