Documentation should cover the OS settings required to run mongodb at scale.

Discussion:

The Linux OS limits cover number of open files/sockets per process, number of threads per user, and various other per-process limits. If these are set too low, mongod/mongos will fail with various errors.

For mongod running standalone or in a replica set (non-sharded):

• One thread is allocated for each connection to the 'mongod' process
• One file descriptor is used by each incoming connection
• One file descriptor is used for each data file or journal file used by the database

For mongos (running in a sharded cluster):

• One thread is allocated for each incoming connection from a client
• One file descriptor is used by each incoming connection
• One file descriptor is used for a connection to each member of each shard

Here are the values I'd recommend for the 'ulimit' settings for the MongoDB processes:

• file size (-f) unlimited
• cpu time (-t) unlimited
• virtual memory (-v) unlimited
• open files (-n) 21000
• memory size (-m) unlimited
• processes/threads (-u) 32000

Diagnosing problems:

On Linux, users can determine the limits for the mongod/mongos processes by examining the contents of the file /proc/$PID/limits, where $PID is the process ID of the mongod/mongos process in question.