Most operating systems limit an amount of TCP sockets that can be used 
simultaneously. It is possible for a busy Tor relay to run into these
limits, thus being unable to fully utilize the bandwidth resources it 
has at its disposal. Following system-specific tips might be helpful
to alleviate the aforementioned problem.

Linux
-----

Use 'ulimit -n' to raise an allowed number of file descriptors to be 
opened on your host at the same time.

FreeBSD
-------

Tune the followind sysctl(8) variables:
 * kern.maxfiles - maximum allowed file descriptors (for entire system)
 * kern.maxfilesperproc - maximum file descriptors one process is allowed
   to use
 * kern.ipc.maxsockets - overall maximum numbers of sockets for entire 
   system
 * kern.ipc.somaxconn - size of listen queue for incoming TCP connections
   for entire system

See also:
 * https://www.freebsd.org/doc/handbook/configtuning-kernel-limits.html
 * https://wiki.freebsd.org/NetworkPerformanceTuning

Mac OS X
--------

Since Mac OS X is BSD-based system, most of the above hold for OS X as well.
However, launchd(8) is known to modify kern.maxfiles and kern.maxfilesperproc
when it launches tor service (see launchd.plist(5) manpage). Also, 
kern.ipc.maxsockets is determined dynamically by the system and thus is 
read-only on OS X.

Disclaimer
----------

Do note that this document is a draft and above information may be
technically incorrect and/or incomplete. If so, please open a ticket
on https://trac.torproject.org or post to tor-relays mailing list.

Are you running a busy Tor relay? Let us know how you are solving
the out-of-sockets problem on your system.