imudp: UDP Syslog Input Module¶
|Author:||Rainer Gerhards <email@example.com>|
Provides the ability to receive syslog messages via UDP.
Multiple receivers may be configured by specifying multiple input statements.
Note that in order to enable UDP reception, Firewall rules probably need to be modified as well. Also, SELinux may need additional rules.
Parameter names are case-insensitive.
This is a performance optimization. Getting the system time is very costly. With this setting, imudp can be instructed to obtain the precise time only once every n-times. This logic is only activated if messages come in at a very fast rate, so doing less frequent time calls should usually be acceptable. The default value is two, because we have seen that even without optimization the kernel often returns twice the identical time. You can set this value as high as you like, but do so at your own risk. The higher the value, the less precise the timestamp.
Note: the timeRequery is done based on executed system calls (not messages received). So when batch sizes are used, multiple messages are received with one system call. All of these messages always receive the same timestamp, as they are effectively received at the same time. When there is very high traffic and successive system calls immediately return the next batch of messages, the time requery logic kicks in, which means that by default time is only queried for every second batch. Again, this should not cause a too-much deviation as it requires messages to come in very rapidly. However, we advise not to set the “timeRequery” parameter to a large value (larger than 10) if input batches are used.
Can be used the set the scheduler priority, if the necessary functionality is provided by the platform. Most useful to select “fifo” for real-time processing under Linux (and thus reduce chance of packet loss). Other options are “rr” and “other”.
Scheduling priority to use.
This parameter is only meaningful if the system support recvmmsg() (newer Linux OSs do this). The parameter is silently ignored if the system does not support it. If supported, it sets the maximum number of UDP messages that can be obtained with a single OS call. For systems with high UDP traffic, a relatively high batch size can reduce system overhead and improve performance. However, this parameter should not be overdone. For each buffer, max message size bytes are statically required. Also, a too-high number leads to reduced efficiency, as some structures need to be completely initialized before the OS call is done. We would suggest to not set it above a value of 128, except if experimental results show that this is useful.
New in version 7.5.5.
Number of worker threads to process incoming messages. These threads are utilized to pull data off the network. On a busy system, additional threads (but not more than there are CPUs/Cores) can help improving performance and avoiding message loss. Note that with too many threads, performance can suffer. There is a hard upper limit on the number of threads that can be defined. Currently, this limit is set to 32. It may increase in the future when massive multicore processors become available.
New in version 8.37.0.
This parameter is for controlling the case in fromhost. If preservecase is set to “on”, the case in fromhost is preserved. E.g., ‘Host1.Example.Org’ when the message was received from ‘Host1.Example.Org’. Default to “off” for the backword compatibility.
Local IP address (or name) the UDP server should bind to. Use “*” to bind to all of the machine’s addresses.
Specifies the port the server shall listen to.. Either a single port can be specified or an array of ports. If multiple ports are specified, a listener will be automatically started for each port. Thus, no additional inputs need to be configured.
Single port: Port=”514”
Array of ports: Port=[“514”,”515”,”10514”,”…”]
New in version 8.18.0.
Manages the IP_FREEBIND option on the UDP socket, which allows binding it to an IP address that is nonlocal or not (yet) associated to any network interface.
The parameter accepts the following values:
- 0 - does not enable the IP_FREEBIND option on the UDP socket. If the bind() call fails because of EADDRNOTAVAIL error, socket initialization fails.
- 1 - silently enables the IP_FREEBIND socket option if it is required to successfully bind the socket to a nonlocal address.
- 2 - enables the IP_FREEBIND socket option and warns when it is used to successfully bind the socket to a nonlocal address.
Bind socket to given device (e.g., eth0)
For Linux with VRF support, the Device option can be used to specify the VRF for the Address.
Binds the listener to a specific ruleset.
New in version 7.3.1.
The rate-limiting interval in seconds. Value 0 turns off rate limiting. Set it to a number of seconds (5 recommended) to activate rate-limiting.
New in version 7.3.1.
Specifies the rate-limiting burst in number of messages.
New in version 8.3.3.
Specifies the value of the inputname property. In older versions, this was always “imudp” for all listeners, which still is the default. Starting with 7.3.9 it can be set to different values for each listener. Note that when a single input statement defines multiple listener ports, the inputname will be the same for all of them. If you want to differentiate in that case, use “name.appendPort” to make them unique. Note that the “name” parameter can be an empty string. In that case, the corresponding inputname property will obviously also be the empty string. This is primarily meant to be used together with “name.appendPort” to set the inputname equal to the port.
New in version 7.3.9.
Appends the port the inputname property. Note that when no “name” is specified, the default of “imudp” is used and the port is appended to that default. So, for example, a listener port of 514 in that case will lead to an inputname of “imudp514”. The ability to append a port is most useful when multiple ports are defined for a single input and each of the inputnames shall be unique. Note that there currently is no differentiation between IPv4/v6 listeners on the same port.
This is an experimental parameter; details may change at any time and it may also be discontinued without any early warning. Permits to set a default timezone for this listener. This is useful when working with legacy syslog (RFC3164 et al) residing in different timezones. If set it will be used as timezone for all messages that do not contain timezone info. Currently, the format must be “+/-hh:mm”, e.g. “-05:00”, “+01:30”. Other formats, including TZ names (like EST) are NOT yet supported. Note that consequently no daylight saving settings are evaluated when working with timezones. If an invalid format is used, “interesting” things can happen, among them malformed timestamps and rsyslogd segfaults. This will obviously be changed at the time this feature becomes non-experimental.
New in version 7.3.9.
This request a socket receive buffer of specific size from the operating system. It is an expert parameter, which should only be changed for a good reason. Note that setting this parameter disables Linux auto-tuning, which usually works pretty well. The default value is 0, which means “keep the OS buffer size unchanged”. This is a size value. So in addition to pure integer values, sizes like “256k”, “1m” and the like can be specified. Note that setting very large sizes may require root or other special privileges. Also note that the OS may slightly adjust the value or shrink it to a system-set max value if the user is not sufficiently privileged. Technically, this parameter will result in a setsockopt() call with SO_RCVBUF (and SO_RCVBUFFORCE if it is available). (Maximum Value: 1G)
This plugin maintains statistics for each listener and for each worker thread.
The listener statistic is named starting with “imudp”, followed followed by the listener IP, a colon and port in parenthesis. For example, the counter for a listener on port 514 (on all IPs) with no set name is called “imudp(*:514)”.
If an “inputname” is defined for a listener, that inputname is used instead of “imudp” as statistic name. For example, if the inputname is set to “myudpinut”, that corresponding statistic name in above case would be “myudpinput(*:514)”. This has been introduced in 7.5.3.
The following properties are maintained for each listener:
- submitted - total number of messages submitted for processing since startup
The worker thread (in short: worker) statistic is named “imudp(wX)” where “X” is the worker thread ID, which is an monotonically increasing integer starting at 0. This means the first worker will have the name “imudp(w0″), the second “imudp(w1)” and so on. Note that workers are all equal. It doesn’t really matter which worker processes which messages, so the actual worker ID is not of much concern. More interesting is to check how the load is spread between the worker. Also note that there is no fixed worker-to-listener relationship: all workers process messages from all listeners.
Note: worker thread statistics are available starting with rsyslog 7.5.5.
- disallowed - total number of messages discarded due to disallowed sender
This counts the number of messages that have been discarded because they have been received by an disallowed sender. Note that if no allowed senders are configured (the default), this counter will always be zero.
This counter was introduced by rsyslog 8.35.0.
The following properties are maintained for each worker thread:
- called.recvmmsg - number of recvmmsg() OS calls done
- called.recvmsg - number of recvmsg() OS calls done
- msgs.received - number of actual messages received
- Scheduling parameters are set after privileges have been dropped. In most cases, this means that setting them will not be possible after privilege drop. This may be worked around by using a sufficiently-privileged user account.
This sets up an UDP server on port 514:
module(load="imudp") # needs to be done just once input(type="imudp" port="514")
This sets up a UDP server on port 514 bound to device eth0:
module(load="imudp") # needs to be done just once input(type="imudp" port="514" device="eth0")
The following sample is mostly equivalent to the first one, but request a larger rcvuf size. Note that 1m most probably will not be honored by the OS until the user is sufficiently privileged.
module(load="imudp") # needs to be done just once input(type="imudp" port="514" rcvbufSize="1m")
In the next example, we set up three listeners at ports 10514, 10515 and 10516 and assign a listener name of “udp” to it, followed by the port number:
module(load="imudp") input(type="imudp" port=["10514","10515","10516"] inputname="udp" inputname.appendPort="on")
The next example is almost equal to the previous one, but now the inputname property will just be set to the port number. So if a message was received on port 10515, the input name will be “10515” in this example whereas it was “udp10515” in the previous one. Note that to do that we set the inputname to the empty string.
module(load="imudp") input(type="imudp" port=["10514","10515","10516"] inputname="" inputname.appendPort="on")
Additional Information on Performance Tuning¶
Threads and Ports¶
The maximum number of threads is a module parameter. Thus there is no direct relation to the number of ports.
Every worker thread processes all inbound ports in parallel. To do so, it adds all listen ports to an epoll() set and waits for packets to arrive. If the system supports the recvmmsg() call, it tries to receive up to batchSize messages at once. This reduces the number of transitions between user and kernel space and as such overhead.
After the packages have been received, imudp processes each message and creates input batches which are then submitted according to the config file’s queue definition. After that the a new cycle beings and imudp return to wait for new packets to arrive.
When multiple threads are defined, each thread performs the processing described above. All worker threads are created when imudp is started. Each of them will individually awoken from epoll as data is present. Each one reads as much available data as possible. With a low incoming volume this can be inefficient in that the threads compete against inbound data. At sufficiently high volumes this is not a problem because multiple workers permit to read data from the operating system buffers while other workers process the data they have read. It must be noted that “sufficiently high volume” is not a precise concept. A single thread can be very efficient. As such it is recommended to run impstats inside a performance testing lab to find out a good number of worker threads. If in doubt, start with a low number and increase only if performance actually increases by adding threads.
A word of caution: just looking at thread CPU use is not a proper way to monitor imudp processing capabilities. With too many threads the overhead can increase, even strongly. This can result in a much higher CPU utlization but still overall less processing capability.
Please also keep in your mind that additional input worker threads may cause more mutex contention when adding data to processing queues.
Too many threads may also reduce the number of messages received via a single recvmmsg() call, which in turn increases kernel/user space switching and thus system overhead.
If real time priority is used it must be ensured that not all operating system cores are used by imudp threads. The reason is that otherwise for heavy workloads there is no ability to actually process messages. While this may be desirable in some cases where queue settings permit for large bursts, it in general can lead to pushback from the queues.
For lower volumes, real time priority can increase the operating system overhead by awaking imudp more often than strictly necessary and thus reducing the effectiveness of recvmmsg().
imudp threads and queue worker threads¶
There is no direct relationship between these two entities. Imudp submits messages to the configured rulesets and places them into the respective queues. It is then up the the queue config, and outside of the scope or knowledge of imudp, how many queue worker threads will be spawned by the queue in question.
Note, however, that queue worker threads and imudp input worker threads compete for system ressources. As such the combined overall value should not overload the system. There is no strict rule to follow when sizing overall worker numbers: for queue workers it strongly depends on how compute-intense the workload is. For example, omfile actions need few worker threads as they are fast. On the contrary, omelasticsearch often waits for server replies and as such more worker threads can be beneficial. The queue subsystem auto-tuning of worker threads should handle the different needs in a useful way.
- rsyslog video tutorial on how to store remote messages in a separate file.
- Description of rsyslog statistic counters. This also describes all imudp counters.
Help with configuring/using