I can add that this problem happen after "too many open file descriptors" condition reached but it persists even after the sockets are cleaned.
On Fri, Jul 19, 2013 at 5:29 AM, Oleg Moskalenko <mom040...@gmail.com>wrote: > Hi > > I am the author and developer of the rfc5766-turn-server project: > > https://code.google.com/p/rfc5766-turn-server/ > > I have a strange problem with libevent2 consuming 100% of 100% of CPU in > Linux (any linux) on high load condition, even after the load is back to > normal. Basically, the system is doing nothing, just spinning inside > libevent loop, indefinitely. The system is responsitive, but CPU is > consumed 100%. When I connect to it with gdb, I always find myself > somewhere inside timing functions in epoll loop, for example: > > ================================== > > #0 0x00007f33e913a0d3 in epoll_wait () from > /lib/x86_64-linux-gnu/libc.so.6 > #1 0x00007f33ea3b2f37 in ?? () from > /usr/lib/x86_64-linux-gnu/libevent_core-2.0.so.5 > #2 0x00007f33ea39ede4 in event_base_loop () from > /usr/lib/x86_64-linux-gnu/libevent_core-2.0.so.5 > #3 0x000000000040e1ec in run_events (eb=0x14ae060) at > src/apps/relay/mainrelay.c:1006 > #4 0x000000000040e2dd in run_listener_server (eb=0x14ae060) at > src/apps/relay/mainrelay.c:1039 > #5 0x0000000000410ebd in main (argc=0, argv=0x7fffffe549d8) at > src/apps/relay/mainrelay.c:2457 > > ================================= > > everything works fine in other OS's (BSD, Solaris). But all Linuxes have > this problem. > > I was searching the Internet and I found the thread where people are > having the exact same problem: > > http://lists.pgfoundry.org/pipermail/pgbouncer-general/2011-May/000805.html > > They resolved the problem by not using the libevent. Unfortunately, my > system is built around libevent2 and I cannot afford that solution. > > is there a known bug, and how do I deal with this problem ? Too many users > are using Linux, I cannot tell everybody to switch to BSD. > > Thanks > Oleg > >
