Table of Contents
List of Examples
timer_interval parameterqueue parameterpipe parameterrl_check usagerl_check_pipe usageTable of Contents
This module implements rate limiting for SIP requests. In contrast to the PIKE module this limits the flow based on a per SIP request type basis and not per source IP. The RPC interface can be used to change tunables while running Kamailio.
The module implements the pipe/queue policy from BSD's ipfw manual, with some simplifications. In principle, each specified method is associated with its own queue and a number of queues are connected to a certain pipe (see the queue and pipe params).
Please also take a look at the “pipelimit” module, that implements the pipe policy with database support. Note that it doesn't implement the queues that exist in this module.
Limiting the rate messages are processed on a system directly influences the load. The ratelimit module can be used to protect a single host or to protect a Kamailio cluster when run on the dispatching box in front.
A sample configuration snippet might look like this:
...
if (is_method("INVITE|REGISTER|SUBSCRIBE") {
if (!rl_check()) {
append_to_reply("Retry-After: 5\r\n");
sl_send_reply("503","Limiting");
exit;
};
};
...
Upon every incoming request listed above rl_check
is invoked. It returns an OK code if the current per request load is below the
configured threshold. If the load is exceeded the function returns an
error and an administrator can discard requests with a stateless
response.
The ratelimit module supports two different statc algorithms to be used by rl_check to determine whether a message should be blocked or not.
This is a trivial algorithm that imposes some risks when used in conjunction with long timer intervals. At the start of each interval an internal counter is reset and incremented for each incoming message. Once the counter hits the configured limit rl_check returns an error.
The downside of this algorithm is that it can lead to SIP client synchronization. During a relatively long interval only the first requests (i.e. REGISTERs) would make it through. Following messages (i.e. RE-REGISTERs) will all hit the SIP proxy at the same time when a common Expire timer expired. Other requests will be retransmitted after a given time, the same on all devices with the same firmware/by the same vendor.
The Random Early Detection Algorithm tries to circumvent the synchronization problem imposed by the tail drop algorithm by measuring the average load and adapting the drop rate dynamically. When running with the RED algorithm (enabled by default) Kamailio will return errors to the Kamailio routing engine every n'th packet trying to evenly spread the measured load of the last timer interval onto the current interval. As a negative side effect Kamailio might drop messages although the limit might not be reached within the interval. Decrease the timer interval if you encounter this.
This algorithm relies on information provided by network interfaces. The total amount of bytes waiting to be consumed on all the network interfaces is retrieved once every timer_interval seconds. If the returned amount exceeds the limit specified in the modparam, rl_check returns an error.
When running Kamailio on different machines, one has to adjust the drop rates for the static algorithms to maintain a sub 100% load average or packets will start getting dropped in the network stack. While this is not in itself difficult, it isn't neither accurate nor trivial: another server taking a notable fraction of the CPU time will require re-tuning the parameters.
While tuning the drop rates from the outside based on a certain factor is possible, having the algorithm run inside ratelimit permits tuning the rates based on internal server parameters and is somewhat more flexible (or it will be when support for external load factors - as opposed to cpu load - is added).
Using the PID Controller model (see Wikipedia page), the drop rate is adjusted dynamically based on the load factor so that the load factor always drifts towards the specified limit (or setpoint, in PID terms).
As reading the CPU load average is relatively expensive (opening /proc/stat, parsing it, etc), this only happens once every timer_interval seconds and consequently the FEEDBACK value is only at these intervals recomputed. This in turn makes it difficult for the drop rate to adjust quickly. Worst case scenarios are request rates going up/down instantly by thousands - it takes up to 20 seconds for the controller to adapt to the new request rate.
Generally though, as real life request rates drift by less, adapting should happen much faster.
The following modules must be loaded before this module:
SL: Stateless request handling.
The initial length of a timer interval in seconds. All amounts of messages have to be divided by this timer to get a messages per second value.
IMPORTANT: A too small value may lead to performance penalties due to timer process overloading.
Default value is 10.
The format of the queue parameter is "pipe_no:method". For each defined method, the algorithm defined by pipe number "pipe_no" will be used.
To specify a queue that accepts all methods, use “*” instead of METHOD. As queues are matched against request methods, you will usually want to have this as the last queue added or other queues with specific methods will never match. At this time, glob or regexp patterns are not supported.
Example 1.2. Set queue parameter
...
# assign pipe no 0 to method REGISTER
# assign pipe no 3 to method INVITE
# assign pipe no 2 to all other methods
modparam("ratelimit", "queue", "0:REGISTER")
modparam("ratelimit", "queue", "3:INVITE")
modparam("ratelimit", "queue", "2:*")
...
The format of the pipe param is "pipe_no:algorithm:limit". For each defined pipe, the given algorithm with the given limit will be used.
A pipe is characterised by its algorithm and limit (bandwidth, in ipfw terms). When specifying a limit, the unit depends on the algorithm used and doesn't need to be specified also (eg, for TAILDROP or RED, limit means packets/sec, whereas with the FEEDBACK algorithm, it means [CPU] load factor).
Example 1.3. Set pipe parameter
...
# define pipe 0 with a limit of 200 pkts/sec using TAILDROP algorithm
# define pipe 1 with a limit of 100 pkts/sec using RED algorithm
# define pipe 2 with a limit of 50 pkts/sec using TAILDROP algorithm
# define pipe 3 with a limit of load factor 80 using FEEDBACK algorithm
# define pipe 4 with a limit of 10000 pending bytes in the rx_queue
# using NETWORK algorithm
modparam("ratelimit", "pipe", "0:TAILDROP:200")
modparam("ratelimit", "pipe", "1:RED:100")
modparam("ratelimit", "pipe", "2:TAILDROP:50")
modparam("ratelimit", "pipe", "3:FEEDBACK:80")
modparam("ratelimit", "pipe", "4:NETWORK:10000")
...
Check the current request against the matched ratelimit algorithm. If no parameter is provided, the queue will be matched based on method type, and then the pipe will be identified based on the matched queue. If a pipe number is provided as a parameter, then the given pipe number will be used for identifying the ratelimit algorithm. The pipe number must be provided as number or via a pseudovariable.
The method will return an error code if the limit for the matched algorithm is reached.
Meaning of the parameters is as follows:
pvar - the pseudovariable holding the pipe id to be used by ratelimit.
This function can be used from ANY_ROUTE.
Example 1.4. rl_check usage
...
# perform queue/pipe match for current method
if (!rl_check()) {
append_to_reply("Retry-After: 5\r\n");
sl_send_reply("503","Limiting");
exit;
}
...
# use pipe no 1 for the current method
# set int pvar to 1
$var(p) = 1;
if (!rl_check("$var(p)")) {
append_to_reply("Retry-After: 5\r\n");
sl_send_reply("503","Limiting");
exit;
}
...
Check the current request against the matched ratelimit algorithm of the pipe provided as parameter. The parameter can be provided as number or variable.
The method will return an error code if the limit for the matched algorithm is reached.
Meaning of the parameters is as follows:
pipe_no - the pipe id to be used by ratelimit.
This function can be used from REQUEST_ROUTE.
Example 1.5. rl_check_pipe usage
...
# use pipe no 1 for the current method
if (!rl_check_pipe("1") {
append_to_reply("Retry-After: 5\r\n");
sl_send_reply("503","Limiting");
exit;
}
...
Lists the parameters and variables in the ratelimit module.
Name: rl.stats
Parameters: none
RPC Command Format:
kamcmd rl.stats
Sets the pipe parameters for the given pipe id.
Name: rl.set_pipe
Parameters:
pipe_id - pipe id.
pipe_algorithm - the algorithm assigned to the given pipe id.
pipe_limit - the limit assigned to the given pipe id.
RPC Command Format:
kamcmd rl.set_pipe 2 RED 10
Gets the list of in use pipes.
Name: rl.get_pipes
Parameters:none
RPC Command Format:
kamcmd rl.get_pipes
Sets the queue parameters for the given queue id.
Name: rl.set_queue
Parameters:
queue_id - queue id.
queue_method - the method assigned to the given queue id.
pipe_id - the pipe id assigned to the given queue id.
RPC Command Format:
kamcmd rl.set_queue 3 INVITE 2
Gets the list of in use queues.
Name: rl.get_queues
Parameters: none
RPC Command Format:
kamcmd rl.get_queues
Sets the PID Controller parameters for the Feedback Algorithm.
Name: rl.set_pid
Parameters:
ki - the integral parameter.
kp - the proportional parameter.
kd - the derivative parameter.
RPC Command Format:
kamcmd rl.set_pid 0.5 0.5 0.5
Gets the list of in use PID Controller parameters.
Name: rl.get_pid
Parameters: none
RPC Command Format:
kamcmd rl.get_pid
Force the value of the load parameter. This method is useful for testing the Feedback algorithm.
Name: rl.push_load
Parameters:
load - the forced value of load (it must be greater then 0.0 and smaller then 1.0).
RPC Command Format:
kamcmd rl.push_load 0.85