This section describes the internal message-processing model of syslog-ng, as well as the flow-control feature that can prevent message losses.
The syslog-ng application monitors (polls) the sources defined in its configuration file, periodically checking each source for messages. When a log message is found in one of the sources, syslog-ng polls every source and reads the available messages. These messages are processed and put into the output buffer of syslog-ng (also called fifo). From the output buffer, the operating system sends the messages to the appropriate destinations.
In large-traffic environments many messages can arrive during a single poll loop, therefore syslog-ng reads only a fixed number of messages from each source. The log-fetch-limit() option specifies the number of messages read during a poll loop from a single source.
Figure 14: Managing log messages in syslog-ng
TCP and unix-stream sources can receive the logs from several incoming connections (for example, many different clients or applications). For such sources, syslog-ng reads messages from every connection, thus the log-fetch-limit() parameter applies individually to every connection of the source.
Figure 15: Managing log messages of TCP sources in syslog-ng
Log paths without flow-control
Every destination has its own output buffer. The output buffer is needed because the destination might not be able to accept all messages immediately. The log-fifo-size() parameter sets the size of the output buffer. The output buffer must be larger than the log-fetch-limit() of the sources, to ensure that every message read during the poll loop fits into the output buffer. If the log path sends messages to a destination from multiple sources, the output buffer must be large enough to store the incoming messages of every source.
Log paths with flow-control
The syslog-ng application uses flow-control in the following cases:
NOTE: The way flow-control works has changed significantly in version syslog-ng OSE3.22. If you are using an older version of syslog-ng OSE, consult the documentation of the version you are using for details about flow-control.
The flow-control of syslog-ng introduces a control window to the source that tracks how many messages can syslog-ng accept from the source. Every message that syslog-ng reads from the source lowers the window size by one, every message that syslog-ng successfully sends from the output buffer increases the window size by one. If the window is full (that is, its size decreases to zero), syslog-ng stops reading messages from the source. The initial size of the control window is by default 100. If a source accepts messages from multiple connections, all messages use the same control window.
When using flow-control, syslog-ng automatically sets the size of the output buffer so that it matches the size of the control window of the sources. Note that starting with syslog-ng OSE3.22, log-fifo-size() only affects log paths that are not flow-controlled.
NOTE: If the source can handle multiple connections (for example, network() and syslog()), the size of the control window is divided by the value of the max-connections() parameter and this smaller control window is applied to each connection of the source.
Dynamic flow-control
In addition to the static control window set using the log-iw-size() option, you can also allocate a dynamic window to the source. The syslog-ng application uses this window to dynamically increase the static window of the active connections. The dynamic window is distributed evenly among the active connections of the source. The syslog-ng application periodically checks which connections of the source are active, and redistributes the dynamic window. If only one of the connections is active, it receives the entire dynamic window, while other connections receive only their share of the static window.
Using dynamic flow-control on your syslog-ng server is useful when the source has lots of connections, but only a small subset of the active clients send messages at high rate, and the memory of the syslog-ng server is limited. In other cases, it is currently not recommended, because it can result in higher memory usage and fluctuating performance compared to using only the static window.
When flow-control is used, every source has its own control window. As a worst-case situation, memory of the host must be greater than the total size of the messages of every control window, plus the size of the dynamic window, that is, the log-iw-size()+dynamic-window-size(). This applies to every source that sends logs to the particular destination. Thus if two sources having several connections and heavy traffic send logs to the same destination, the control window of both sources must fit into the memory of the host. Otherwise, some messages might not fit in the memory, and messages may be lost.
If dynamic flow-control is disabled (which is the default behavior), the value of the log-iw-size() option cannot be lower than 100. If dynamic flow-control is enabled, you can decrease the value of the log-iw-size() option (to the minimum of 1).
In case of soft flow-control there is no message lost if the destination can accept messages. It is possible to lose messages if it cannot accept messages (for example, the file destination is not writable, or the disk becomes full), and all buffers are full. Soft flow-control cannot be configured, it is automatically available for file destinations.
Hard flow-control: In case of hard flow-control there is no message lost. To use hard flow-control, enable the flow-control flag in the log path. Hard flow-control is available for all destinations.
Example: Soft flow-control
source s_file {
file("/tmp/input_file.log");
};
destination d_file {
file("/tmp/output_file.log");
};
destination d_tcp {
network("127.0.0.1"
port(2222)
);
};
log {
source(s_file);
destination(d_file);
destination(d_tcp);
};
|
|
Caution:
Hazard of data loss! For destinations other than file, soft flow-control is not available. Thus, it is possible to lose log messages on those destinations. To avoid data loss on those destinations, use hard flow-control. |
Example: Hard flow-control
source s_file {
file("/tmp/input_file.log");
};
destination d_file {
file("/tmp/output_file.log");
};
destination d_tcp {
network("127.0.0.1"
port(2222)
);
};
log {
source(s_file);
destination(d_file);
destination(d_tcp);
flags(flow-control);
};
Handling outgoing messages
The syslog-ng application handles outgoing messages the following way:
Figure 16: Handling outgoing messages in syslog-ng OSE
-
Output queue: Messages from the output queue are sent to the target syslog-ng server. The syslog-ng application puts the outgoing messages directly into the output queue, unless the output queue is full. The output queue can hold 64 messages, this is a fixed value and cannot be modified.
-
Disk buffer: If the output queue is full and disk-buffering is enabled, syslog-ng puts the outgoing messages into the disk buffer of the destination.
-
Overflow queue: If the output queue is full and the disk buffer is disabled or full, syslog-ng puts the outgoing messages into the overflow queue of the destination. (The overflow queue is identical to the output buffer used by other destinations.) The log-fifo-size() parameter specifies the number of messages stored in the overflow queue, unless flow-control is enabled. When dynamic flow-control is enabled, syslog-ng sets the size of the overflow queue automatically. For details on sizing the log-fifo-size() parameter, see Configuring flow-control.
Using flow-control on a source has an important side-effect if the messages of the source are sent to multiple destinations. If flow-control is in use and one of the destinations cannot accept the messages, the other destinations do not receive any messages either, because syslog-ng stops reading the source. For example, if messages from a source are sent to a remote server and also stored locally in a file, and the network connection to the server becomes unavailable, neither the remote server nor the local file will receive any messages.
NOTE: Creating separate log paths for the destinations that use the same flow-controlled source does not avoid the problem.
If you use flow-control and reliable disk-based buffering together with multiple destinations, the flow-control starts slowing down the source only when:
-
one destination is down, and
-
the number of messages stored in the disk buffer of the destination reaches (disk-buf-size() minus mem-buf-size()).
For details on how flow-control works, see Managing incoming and outgoing messages with flow-control. The summary of the main points is as follows:
-
The syslog-ng application normally reads a maximum of log-fetch-limit() number of messages from a source.
-
From TCP and unix-stream sources, syslog-ng reads a maximum of log-fetch-limit() from every connection of the source. The number of connections to the source is set using the max-connections() parameter.
-
Every destination has an output buffer. The size of this buffer is set automatically for log paths that use flow-control, and can be set using the log-fifo-size() option for other log paths.
-
Flow-control uses a control window to determine if there is free space in the output buffer for new messages. Every source has its own control window, the log-iw-size() option sets the size of the static control window. Optionally, you can enable a dynamic control window for the source using the dynamic-window-size() option.
-
When a source accepts multiple connections, the size of the control window is divided by the value of the max-connections() parameter and this smaller control window is applied to each connection of the source.
The dynamic control window is automatically distributed among the active connections of the source.
-
If the control window is full, syslog-ng stops reading messages from the source until some messages are successfully sent to the destination.
-
If the output buffer becomes full, and neither disk-buffering nor flow-control is used, messages may be lost.
|
|
Caution:
If you modify the max-connections() or the log-fetch-limit() parameter, do not forget to adjust the log-iw-size() and dynamic-window-size() parameters accordingly. |
Example: Sizing parameters for flow-control
Suppose that syslog-ng has a source that must accept up to 300 parallel connections. Such situation can arise when a network source receives connections from many clients, or if many applications log to the same socket.
Set the max-connections() parameter of the source to 300. However, the log-fetch-limit() (default value: 10) parameter applies to every connection of the source individually, while the log-iw-size() (default value: 1000) parameter applies to the source. In a worst-case scenario, the destination does not accept any messages, while all 300 connections send at least log-fetch-limit() number of messages to the source during every poll loop. Therefore, the control window must accommodate at least max-connections()*log-fetch-limit() messages to be able to read every incoming message of a poll loop. In the current example this means that log-iw-size() should be greater than 300*10=3000. If the control window is smaller than this value, the control window might fill up with messages from the first connections — causing syslog-ng to read only one message of the last connections in every poll loop.
The output buffer of the destination must accommodate at least log-iw-size() messages, but use a greater value: in the current example 3000*10=30000 messages. That way all incoming messages of ten poll loops fit in the output buffer. If the output buffer is full, syslog-ng does not read any messages from the source until some messages are successfully sent to the destination.
source s_localhost {
network(
ip(127.0.0.1)
port(1999)
max-connections(300)
);
};
destination d_tcp {
network("10.1.2.3"
port(1999)
localport(999)
log-fifo-size(30000)
);
};
log {
source(s_localhost);
destination(d_tcp);
flags(flow-control);
};
If other sources send messages to this destination, then the output buffer must be further increased. For example, if a network host with maximum 100 connections also logs into the destination, then increase the log-fifo-size() by 10000.
source s_localhost {
network(
ip(127.0.0.1)
port(1999)
max-connections(300)
);
};
source s_tcp {
network(
ip(192.168.1.5)
port(1999)
max-connections(100)
);
};
destination d_tcp {
network("10.1.2.3"
port(1999)
localport(999)
log-fifo-size(40000)
);
};
log {
source(s_localhost);
destination(d_tcp);
flags(flow-control);
};
The syslog-ng Open Source Edition application can store messages on the local hard disk if the destination (for example, the central log server) or the network connection to the destination becomes unavailable. The syslog-ng OSE application automatically sends the stored messages to the destination when the connection is reestablished. The disk buffer is used as a queue: when the connection to the destination is reestablished, syslog-ng OSE sends the messages to the destination in the order they were received.
Every such destination uses a separate disk buffer (similarly to the output buffers controlled by log-fifo-size()). The hard disk space is not pre-allocated, so ensure that there is always enough free space to store the disk buffers even when the disk buffers are full.
If syslog-ng OSE is restarted (using the /etc/init.d/syslog-ng restart command, or another appropriate command on your platform), it automatically saves any unsent messages from the disk buffer and in-memory queues. After the restart, syslog-ng OSE sends the saved messages to the destination. In other words, the disk buffer is persistent. The disk buffer is also resistant to syslog-ng OSE crashes.
The syslog-ng OSE application supports two types of disk buffering: reliable and normal. For details, see Enabling reliable disk-based buffering and Enabling normal disk-based buffering, respectively.
Message handling and normal disk-based buffering
When you use disk-based buffering, and the reliable() option is set to no, syslog-ng OSE handles outgoing messages the following way:
Figure 17: Handling outgoing messages in syslog-ng OSE
-
Output queue: In-memory queue. If there is space left in it, syslog-ng OSE puts the message into this queue first . Messages stored here are processed faster, because syslog-ng OSE can skip writing to, and reading from the disk, as well as serializing or deserializing the message, saving I/O and processor time as a result. The contents of the in-memory output queue are persisted to the disk-buffer file during syslog-ng OSE reload, restart or stop, but they cannot be persisted if in the event of power failures, or if syslog-ng OSE crashes. By default, the output queue can hold 1000 messages (you can adjust this number using the quot-size() option).
-
Disk-buffer file: Disk queue. If there is no space left in the output queue, the message is stored on the disk-buffer file. Messages stored here are persisted on the disk, even in case of power failures or if syslog-ng OSE crashes. Using the disk-buffer file takes considerable amount of disk I/O and processor time. The size of this queue can be set with the disk-buf-size() option.
-
Overflow queue: In-memory queue. This queue is used to trigger flow-control if it is set. The contents of the in-memory overflow queue are persisted to the disk-buffer file in case of syslog-ng OSE reload, restart or stop, but they are not persisted in case of power failures or if syslog-ng OSE crashes. Setting the size of the overflow queue can be done with the mem-buf-length() option.
|
|
Caution:
Hazard of data loss!
In case of normal disk-buffers, the messages stored in the output queue and the overflow queue can be lost in case of power failures or if syslog-ng OSE crashes. |
NOTE: Using disk buffer can significantly decrease performance.
Message handling and reliable disk-based buffering
When you use disk-based buffering, and the reliable() option is set to yes, syslog-ng OSE handles outgoing messages the following way.
The mem-buf-size() option determines when flow-control is triggered. After the size of the disk-buffer file reaches (disk-buf-size() minus mem-buf-size()), messages are written into both the disk-buffer file and the overflow queue, indicating that flow-control needs to slow down the message source. These messages are not taken out from the control window (governed by log-iw-size()), causing the control window to fill up.
If the control window is full, the flow-control completely stops reading incoming messages from the source. (As a result, mem-buf-size() must be at least as large as log-iw-size() times the average message size.)
Figure 18: Handling outgoing messages in syslog-ng OSE with the reliable disk-buffer option
-
Output queue: In-memory and disk queue. If there is space left in it, syslog-ng OSE puts the message into this queue first. In case of reliable disk-buffer, in addition to storing the message in memory, it is stored directly in the disk-buffer file as well for safety reasons (see the next point). Messages stored here are processed faster, because syslog-ng OSE can skip reading from the disk, and deserializing the message, saving I/O and processor time. By default, the output queue can hold 1000 messages (you can adjust it using the quot-size() option).
-
Disk-buffer file: Disk queue. If there is no space left in the output queue, the message is stored on the disk-buffer file. Messages stored here are persisted on the disk, and survive syslog-ng OSE crash or power failure. Using the disk-buffer file takes considerable amount of disk I/O and processor time. The size of this queue can be set with the disk-buf-size() option.
-
Overflow queue: In-memory and disk queue. This queue is used to trigger flow-control if it is set. Similarly to the output queue, in case of reliable disk-buffer in addition to storing the message in memory, it is stored directly in the disk-buffer file as well for safety reasons. Setting the size of the overflow queue can be done with the mem-buf-size() option.
