Apache HTTP Server Version 2.4
|Description:||A variant of the |
event Multi-Processing Module (MPM) is
designed to allow more requests to be served simultaneously by
passing off some processing work to the listeners threads, freeing up
the worker threads to serve new requests.
To use the
event MPM, add
--with-mpm=event to the
script's arguments when building the
event is based on the
worker MPM, which implements a hybrid
multi-process multi-threaded server. A single control process (the parent) is responsible for launching
child processes. Each child process creates a fixed number of server
threads as specified in the
ThreadsPerChild directive, as well
as a listener thread which listens for connections and passes them to a worker thread for processing when they arrive.
Run-time configuration directives are identical to those provided by
worker, with the only addition
This MPM tries to fix the 'keep alive problem' in HTTP. After a client completes the first request, it can keep the connection open, sending further requests using the same socket and saving significant overhead in creating TCP connections. However, Apache HTTP Server traditionally keeps an entire child process/thread waiting for data from the client, which brings its own disadvantages. To solve this problem, this MPM uses a dedicated listener thread for each process to handle both the Listening sockets, all sockets that are in a Keep Alive state, sockets where the handler and protocol filters have done their work and the ones where the only remaining thing to do is send the data to the client.
This new architecture, leveraging non-blocking sockets and modern kernel
features exposed by APR (like Linux's epoll),
no longer requires the
configured to avoid the thundering herd problem.
The total amount of connections that a single process/threads block can handle is regulated
Async connections would need a fixed dedicated worker thread with the previous MPMs but not with event.
The status page of
mod_status shows new columns under the Async connections section:
write()to the socket returns
EAGAINto become writable again after an idle time. The worker holding the socket might be able to offload the waiting task to the listener thread, that in turn will re-assign it to the first idle worker thread available once an event will be raised for the socket (for example, "the socket is now writable"). Please check the Limitations section for more information.
KeepAliveTimeoutoccurs then the socket will be closed by the listener. In this way, the worker threads are not responsible for idle sockets, and they can be re-used to serve other requests.
These improvements are valid for both HTTP/HTTPS connections.
This mpm showed some scalability bottlenecks in the past, leading to the following
error: "scoreboard is full, not at MaxRequestWorkers".
limits the number of simultaneous requests that will be served at any given time
and also the number of allowed processes
the Scoreboard is a representation of all the running processes and
the status of their worker threads. If the scoreboard is full (so all the
threads have a state that is not idle) but the number of active requests
served is not
it means that some of them are blocking new requests that could be served
but that are queued instead (up to the limit imposed by
ListenBacklog). Most of the time,
the threads are stuck in the Graceful state, namely they are waiting to
finish their work with a TCP connection to safely terminate and free up a
scoreboard slot (for example, handling long-running requests, slow clients
or connections with keep-alive enabled). Two scenarios are very common:
MaxSpareThreads). This is particularly problematic because when the load increases again, httpd will try to start new processes. If the pattern repeats, the number of processes can rise quite a bit, ending up in a mixture of old processes trying to stop and new ones trying to do some work.
From 2.4.24 onward, mpm-event is smarter and it is able to handle graceful terminations in a much better way. Some of the improvements are:
ThreadsPerChildare used to limit the amount of active processes; meanwhile,
ServerLimittakes also into account the ones doing a graceful close to allow extra slots when needed. The idea is to use
ServerLimitto instruct httpd about how many overall processes are tolerated before impacting the system resources.
The behavior described in the last point is completely observable via
mod_status in the connection summary table through two new
columns: "Slot" and "Stopping". The former indicates the PID and
the latter if the process is stopping or not; the extra state "Yes (old gen)"
indicates a process still running after a graceful restart.
The improved connection handling may not work for certain connection
filters that have declared themselves as incompatible with event. In these
cases, this MPM will fall back to the behavior of the
worker MPM and reserve one worker thread per connection.
All modules shipped with the server are compatible with the event MPM.
A similar restriction is currently present for requests involving an
output filter that needs to read and/or modify the whole response body.
If the connection to the client blocks while the filter is processing the
data, and the amount of data produced by the filter is too big to be
buffered in memory, the thread used for the request is not freed while
httpd waits until the pending data is sent to the client.
To illustrate this point, we can think about the following two situations: serving a static asset (like a CSS file) versus serving content retrieved from FCGI/CGI or a proxied server. The former is predictable, namely the event MPM has full visibility on the end of the content and it can use events: the worker thread serving the response content can flush the first bytes until
EAGAIN is returned, delegating the rest to the listener. This one in turn
waits for an event on the socket and delegates the work to flush the rest of the content
to the first idle worker thread. Meanwhile in the latter example (FCGI/CGI/proxied content),
the MPM can't predict the end of the response and a worker thread has to finish its work
before returning the control to the listener. The only alternative is to buffer the
response in memory, but it wouldn't be the safest option for the sake of the
server's stability and memory footprint.
The event model was made possible by the introduction of new APIs into the supported operating systems:
Before these new APIs where made available, the traditional
poll APIs had to be used.
Those APIs get slow if used to handle many connections or if the set of connections rate of change is high.
The new APIs allow to monitor many more connections, and they perform way better when the set of connections to monitor changes frequently. So these APIs made it possible to write the event MPM, that scales much better with the typical HTTP pattern of many idle connections.
The MPM assumes that the underlying
implementation is reasonably threadsafe. This enables the MPM to
avoid excessive high level locking, or having to wake up the listener
thread in order to send it a keep-alive socket. This is currently
only compatible with KQueue and EPoll.
This MPM depends on APR's atomic
compare-and-swap operations for thread synchronization. If you are
compiling for an x86 target and you don't need to support 386s, or
you are compiling for a SPARC and you don't need to run on
pre-UltraSPARC chips, add
--enable-nonportable-atomics=yes to the
configure script's arguments. This will cause
APR to implement atomic operations using efficient opcodes not
available in older CPUs.
This MPM does not perform well on older platforms which lack good threading, but the requirement for EPoll or KQueue makes this moot.
glibchas been compiled with support for EPoll.
|Description:||Limit concurrent connections per process|
|Compatibility:||Available in version 2.3.13 and later|
The event MPM handles some connections in an asynchronous way, where request worker threads are only allocated for short periods of time as needed, and other connections with one request worker thread reserved per connection. This can lead to situations where all workers are tied up and no worker thread is available to handle new work on established async connections.
To mitigate this problem, the event MPM does two things:
This directive can be used to fine-tune the per-process connection limit. A process will only accept new connections if the current number of connections (not counting connections in the "closing" state) is lower than:
number of idle workers)
An estimation of the maximum concurrent connections across all the processes given an average value of idle worker threads can be calculated with:
number of idle workers)) *
ThreadsPerChild = 10 ServerLimit = 4 AsyncRequestWorkerFactor = 2 MaxRequestWorkers = 40 idle_workers = 4 (average for all the processes to keep it simple) max_connections = (ThreadsPerChild + (AsyncRequestWorkerFactor * idle_workers)) * ServerLimit = (10 + (2 * 4)) * 4 = 72
When all the worker threads are idle, then absolute maximum numbers of concurrent connections can be calculared in a simpler way:
AsyncRequestWorkerFactor + 1) *
ThreadsPerChild = 10 ServerLimit = 4 MaxRequestWorkers = 40 AsyncRequestWorkerFactor = 2
If all the processes have all threads idle then:
idle_workers = 10
We can calculate the absolute maximum numbers of concurrent connections in two ways:
max_connections = (ThreadsPerChild + (AsyncRequestWorkerFactor * idle_workers)) * ServerLimit = (10 + (2 * 10)) * 4 = 120 max_connections = (AsyncRequestWorkerFactor + 1) * MaxRequestWorkers = (2 + 1) * 40 = 120
AsyncRequestWorkerFactor requires knowledge about the traffic handled by httpd in each specific use case, so changing the default value requires extensive testing and data gathering from
MaxRequestWorkers was called
MaxClients prior to version 2.3.13. The above value
shows that the old name did not accurately describe its meaning for the event MPM.
AsyncRequestWorkerFactor can take non-integer
arguments, e.g "1.5".