porter

porter is a simple Erlang library that is makes creating many client tcp connections easier in Linux. Its primary function is to keep track of source ports used per remote address and safely allocate new ones for new connections.

The following blog post does a great job explaining why you need to manage your own local ports if you're creating many outgoing TCP connections from a single Linux server.

https://idea.popcount.org/2014-04-03-bind-before-connect/

In short, you must use SO_REUSEADDR, and the kernel's port allocation algorithm does not allocate ports efficiently when using SO_REUSEADDR.

Creating a connection with porter

The porter module provides porter:connect/5 for creating tcp and ssl connections, with arguments:

• Module: gen_tcp or ssl
• IP: ipv4 remote address
• Port: remote port
• Opts: options passed into Module:connect
• Timeout: timeout passed into Module:connect

Its return is {ok, Socket, PorterState} or {error, Reason}. The caller must keep track of PorterState during the lifetime of the socket. If the caller intends to initiate the connection close, you must use porter:close/3. If the connection is closed by the remote, you must call porter:on_closed/1. These functions ensure the allocated local port is released back into the pool.

Local port release and tcp_tw_reuse

When a local port is released (connection is closed), it enters an aging period before it's fully available for a new porter connection. We want to give the socket some period in the TIME_WAIT state, if necessary, to cleanly shutdown. However, the aging period is shorter than the TIME_WAIT interval, 60 seconds. Therefore, it's strongly suggested that you use tcp_tw_reuse on your server so that porter's short aging of the local port does not create collisions with existing TIME_WAIT sockets in the kernel.

Configuration

You must define the following environment variables for porter to operate correctly:

• ip_local_port_range: porter does not attempt to read the ip_local_port_range from your Linux system configuration, so you must provide it explicitly. Default is {1024, 65535}
• retry_policy: If a connection attempt fails with eaddrnotavail or eaddrinuse, porter will optionally try again with a new local port allocation. This is to meant to protect against collisions with other local ports in use on the system that may be unrelated to porters management of local ports. An example retry_policy is [{attempts, 10}], which will try 10 different local ports before giving up. The default retry_policy is [], for which porter takes no action on a failed connection.

Internals

Local ports are tracked in several ets tables. The master table is public and named, called porter_addr_map, so it can be dumped at any time with ets:tab2list(porter_addr_map).

The porter_addr_map contains permanent pool entries and transient aging entries, one entry per remote addr, remote port combination. Each entry in the table is a 2-tuple, where the second value is a handle to another ets table. For example: the record {{pool,{10,1,4,148},8883},1017980} describes a local port pool for 10.1.4.148:8883 and the pool contents are located in ets table 1017980, viewable with ets:tab2list(1017980)..