{1 Introduction into Netplex}
{b Contents}
- {!Netplex_intro.terminology}
- {!Netplex_intro.webserver}
- {!Netplex_intro.runwebserver}
- {!Netplex_intro.procmod}
- {!Netplex_intro.crsock}
- {!Netplex_intro.servproc}
- {!Netplex_intro.defproc}
- {!Netplex_intro.workload}
- {!Netplex_intro.messaging}
- {!Netplex_intro.logging}
- {!Netplex_intro.debug}
- {!Netplex_intro.mt}
- {!Netplex_intro.rpc_netplex}
Netplex is a generic (stream) server framework. This means, Netplex
does a lot of things for a network service that are always the same,
regardless of the kind of service:
- Creation of server sockets
- Accepting new network connections
- Organizing multiple threads of execution - either by multiple
processes, multiple POSIX threads, or multiplexing
- Workload management
- Writing log files
- Broadcasting messages to all server components
- Support for a configuration file format
Netplex currently only supports stream sockets (TCP or Unix Domain).
Ocamlnet already includes Netplex adapters for Nethttpd (the HTTP daemon),
and RPC servers. It is likely that more adapters for other network
protocols will follow.
Netplex can bundle several network services into a single system of
components. For example, you could have an RPC service that can be
managed over a web interface provided by Nethttpd. Actually, Netplex
focuses on such systems of interconnected components. RPC plays a
special role in such systems because this is the network protocol
the components use to talk to each other. It is also internally
used by Netplex for its administrative tasks.
{2:terminology Terminology}
In the Netplex world the following words are preferred to refer to
the parts of a Netplex system:
- The {b Netplex controller} is the core component of Netplex.
The controller opens sockets and manages how new connections
are accepted. For every socket, the controller determines
which {b Netplex container} will accept the next connection
that is tried to be established. Furthermore, the controller
manages the startup and shutdown of the Netplex system.
- The {b Netplex services} are the user-defined components
of a Netplex system. Every service runs in its own process(es)
(if multi-processing is selected) or in its own thread(s)
(for POSIX multi-threading). It is up to the user to define
what a service is.
- The {b Netplex protocols} are the languages spoken by the
services. A protocol is bound to one or more sockets.
This means that a service is implemented by a number of
protocols.
- The {b Netplex containers} are processes or threads that may
execute a certain service. Every container is bound to a
specific service. It is possible that only one container
is used for a particular service, but one can also configure
that containers are dynamically started and stopped as the
workload of the system changes.
{2:webserver Example - A Simple Web Server}
In order to create a web server, this main program and the following
configuration file are sufficient. (You find an extended example in
the "examples/nethttpd" directory of the Ocamlnet tarball.)
{[
let main() =
(* Create a parser for the standard Netplex command-line arguments: *)
let (opt_list, cmdline_cfg) = Netplex_main.args() in
(* Parse the command-line arguments: *)
Arg.parse
opt_list
(fun s -> raise (Arg.Bad ("Don't know what to do with: " ^ s)))
"usage: netplex [options]";
(* Select multi-processing: *)
let parallelizer = Netplex_mp.mp() in
(* Start the Netplex system: *)
Netplex_main.startup
parallelizer
Netplex_log.logger_factories
Netplex_workload.workload_manager_factories
[ Nethttpd_plex.nethttpd_factory() ]
cmdline_cfg
;;
Sys.set_signal Sys.sigpipe Sys.Signal_ignore;
main();;
]}
The configuration file:
{[
netplex {
controller {
max_level = "debug"; (* Log level *)
logging {
type = "stderr"; (* Log to stderr *)
}
};
service {
name = "My HTTP file service";
protocol {
(* This section creates the socket *)
name = "http";
address {
type = "internet";
bind = "0.0.0.0:80"; (* Port 80 on all interfaces *)
};
};
processor {
(* This section specifies how to process data of the socket *)
type = "nethttpd";
host {
(* Think of Apache's "virtual hosts" *)
pref_name = "localhost";
pref_port = 80;
names = "*:0"; (* Which requests are matched here: all *)
uri {
path = "/";
service {
type = "file";
docroot = "/usr";
media_types_file = "/etc/mime.types";
enable_listings = true;
}
};
};
};
workload_manager {
type = "dynamic";
max_jobs_per_thread = 1; (* Everything else is senseless *)
min_free_jobs_capacity = 1;
max_free_jobs_capacity = 1;
max_threads = 20;
};
}
}
]}
As you can see, the main program is extremely simple. Netplex includes
support for command-line parsing, and the rest deals with the question
which Netplex modules are made accessible for the configuration file.
Note that detailed information about the config file is available in
the {!Netplex_admin.main}.
Here, we have:
- The multi-execution provider is {!Netplex_mp.mp} which implements
multi-processing. (Btw, multi-processing is the preferred
parallelizing technique in Netplex.) Replace it with
{!Netplex_mt.mt} to get multi-threading.
- The {!Netplex_log.logger_factories} are the list of all predefined
logging mechanisms. The configuration file can select one of these
mechanisms.
- The {!Netplex_workload.workload_manager_factories} are the list of
all predefined worload management mechanisms. The configuration
file can select one of these mechanisms per service.
- Finally, we pass {!Nethttpd_plex.nethttpd_factory} as the only
service processor.
The configuration file consists of nested sections whose extents are
denoted by curly braces. The sections are partly defined by Netplex
itself (e.g. the controller section and the workload manager section),
and partly by the service provider (almost everything inside
"processor"). That means that the components of a Netplex system
pick "their" part from the configuration file, and that, depending
on which components are linked into this system, the config files
may look very different.
Here, we have:
- The [controller] section sets the log level and the logging method.
The latter is done by naming one of the logger factories as
logging [type]. If the factory needs more parameters to create
the logger, these can be set inside the [logging] section.
- For every [service] there is a [name] (can be freely chosen), one
or several [protocol]s, a [processor], and a [workload_manager].
The [protocol] section declare which protocols are available and
to which sockets they are bound. Here, the "http" protocol (name can again
be freely chosen) is reachable over TCP port 80 on all network
interfaces. By having multiple [address] sections, one can bind
the same protocol to multiple sockets.
- The [processor] section specifies the [type] and optionally a
lot of parameters (which may be structured into several sections).
By setting [type] to "nethttpd" we select the
{!Nethttpd_plex.nethttpd_factory} to create the processor
(because "nethttpd" is the default name for this factory).
This factory now interprets the other parameters of the [processor]
section. Here, a static HTTP server is defined that uses /usr
as document root.
- Finally, the [workload_manager] section says how to deal with
parallely arriving requests. The [type] selects the dynamic
workload manager which is configured by the other parameters.
Roughly said, one container (i.e. process) is created in advance
for the next network connection ("pre-fork"), and the upper limit
of containers is 20.
{2:runwebserver Running This Example}
If you start this program without any arguments, it will immediately
fail because it wants to open [/etc/netplex.conf] - this is the
default name for the configuration file. Use [-conf] to pass the
real name of the above file.
Netplex creates a directory for its internal processing, and this is
by default [/tmp/netplex-<id>] (where id is a hex sequence derived from
the file location of the configuration). You can change this directory by
setting the [socket_directory] parameter in the [controller] section.
In this directory, you can find:
- A directory [netplex.controller] which refers to the controller
component.
- For every service another directory containing local run-time files.
The directory has the same name as the service.
Netplex comes with a generic administration command called
[netplex-admin]. You can use it to send control messages to Netplex
systems. For example,
{[ netplex-admin -list ]}
outputs the list of services. A more detailed list can be obtained with
{[ netplex-admin -containers ]}
The command
{[ netplex-admin -shutdown ]}
shuts the system (gracefully) down. It is also possible to broadcast
messages to all components:
{[ netplex-admin name arg1 arg2 ... ]}
It is up to the components to interpret these messages.
{2:procmod The Process Model}
Netplex uses a generalized pre-fork process model. Let me explain this
model a bit, as it is important to know it in order to understand
Netplex fully.
The most trivial form of a multi-process Unix server is the post-fork
model. Although it is not used by Netplex, it is the model explained
in many books, and it is what many people think a Unix server has to
look like. Actually, the post-fork model has lots of limitations, and
is not suited for high-performance servers.
In the post-fork model, the master process accepts new network
connections in an endless loop, and whenever a new connection is
established, a sub process (container process) is spawned that processes
the network traffic. There is a serious logical limitation, and a
performance limitation:
- In the post-fork model every container process can only deal with
one connection at a time. The reason is that at the time of
spawning the container there is only one connection, and one cannot
assign the container another connection later.
- The post-fork model spawns the container processes at a bad moment.
Spawning is a very expensive operation, and doing it just after
connection establishment is bad because this means that the
client has to wait longer for a response. Furthermore, spawning
for every connection wastes system resources.
In the pre-fork model, these disadvantages can be avoided. Here, one
or several processes are spawned in advance. Furthermore, these
processes cannot only manage one connection but any number, and
this can happen sequentially (one connection is processed after the
other) or in parallel (using multiplexing).
This is achieved by letting the containers themselves accept the new
connections instead of the master process. In the Unix process model
it is possible that server sockets are shared by several processes,
and every container is allowed to accept the next connection. However,
the containers should cooperate, and avoid that several containers call
[Unix.accept] at the same time (as this leads to performance problems
when a container must be able to watch several ports for new
connections - a problem we do not discuss here). There are many ways to
organize
such cooperation, and for simplicity, Netplex implements this by
exchanging RPC messages with the master process, the controller.
Effectively, the controller has the task of scheduling which of the
containers accepts the next arriving network connection.
What actually happens is the following. We assume here that we have
a number of idle container processes that could accept the next connection.
- The controller selects one of the containers as the one that will get
the next connection.
- The selected container watches the ports for incoming connections.
Note that this is done in a Unixqueue, so that if there are
already connections to be processed, this can be done in a multiplexed
way in parallel with watching for new connections.
- When the next connection arrives, the container accepts it, and
invokes the service component to process it.
- Immediately after connection establishment, the container tells the
controller what happened, so the controller can watch out for another
container to take over the role of accepting further connections.
- When the connection is fully processed, another control message
is sent to the controller because the controller must know at all
times how many connections are being processed by which containers.
This is simply an important load parameter.
The details of this mechanism are not very interesting for using
it. However, one must know that
- connections are accepted by the sub processes and not by the
master process,
- the sub processes can accept as many connections as they want to,
either one after the other, or even several at once,
- the controller schedules tasks and determines which connection
is accepted by which container,
- there is a certain protocol between controller and container, and
although the details are hidden from the user, this has consequences
for the user interface. In particular, the reason why the
[when_done] function must be called upon connection termination
is that a control message must be sent to the controller.
Another implication of the pre-fork model is that one needs workload
management. As processes are created in advance, the question arises
how many are created, and when the processes are terminated to free
resources. Netplex comes with two workload managers: One manager
simply creates a fixed number of processes which are never terminated,
and the other manager tries to adapt the number to the load by
dynamically starting and stopping processes. This is discussed below in
detail.
{2:crsock Creating Sockets}
The server sockets are always created by the controller at program
startup. This is a strict requirement because only this ensures that
the created container processes share the same sockets.
The sockets are descibed in the [protocol] section of the
configuration file. For an Internet socket this section looks like
{[
protocol {
name = "<name>";
address {
type = "internet";
bind = "<address>:<port>";
};
};
]}
The [<name>] is only important when there are several protocols in order
to distinguish between them. The [<address>] can be:
- An IP address of a network interface to bind the socket to this
particular interface. Both IPv4 and IPv6 addresses are supported.
IPv4 addresses are simply given in "dotted quad" notation
(e.g. [192.168.76.23]), and IPv6 addresses must be enclosed in brackets
(e.g. [[fe80::250:56ff:fec0:1]]).
- The special IPv4 address [0.0.0.0] to bind the socket to all IPv4
network interfaces, or the special IPv6 address [[::0]] to bind it
to all IPv6 network interfaces.
- A resolvable host name which is the same as using the corresponding
IP address.
The [<port>] must be the port number or 0 to use an anonymous port.
For a local (Unix domain) socket, the [protocol] section looks like
{[
protocol {
name = "<name>";
address {
type = "local";
path = "<path>";
};
};
]}
where the [<path>] is the filename of the socket.
One can have several [address] sections to create several sockets for the
same protocol.
For detailed documentation, see {!Netplex_admin.protocol}.
{2:servproc Services And Processors}
A Netplex system consists of exactly the services that are enumerated
in the config file. This means it is not sufficient to build in support
for a service into the program, one must also activate it in the config
file. This gives the end user of the program a lot of flexibility when
running the system: By simply changing the config file one can enable
or disable services. It is also possible to run the same program binary
several times with different config files.
The services are implemented by processors, which are user-defined
objects that handle the network connection after it is accepted by
the component. The [processor] section of the service selects the
processor by name, and optionally passes configuration parameters to
it:
{[
processor {
type = "<name>";
... further parameters allowed ...
}
]}
The mentioned name of the processor type is used to find the so-called
factory for the processor (an object with a [create_processor] method).
All factories must be available at Netplex startup so the library
knows which factories exist when the config file is interpreted
(the factories are an argument of {!Netplex_main.startup}).
Processor objects are somewhat strange in so far as they exist both in
the controller and in the container processes. In particular, these
objects are created by the controller, and they are duplicated once
for all container processes when these are actually created.
The processor objects (of type {!Netplex_types.processor}) consist of
a number of methods. We have already seen one of them, [process],
which is called in the container process when a new connection is
accepted. The other methods are called at other points of interest
(see {!Netplex_types.processor_hooks} for more details):
{b Methods called on the controller instance of the processor}
- [post_add_hook] is immediately called after the addtion of the
service to the controller.
- [post_rm_hook] is immediately called after the removal of the
service from the controller.
- [pre_start_hook] is called just before the next container process
is spawned.
- [post_finish_hook] is called after termination of the container.
{b Methods called on the container instance of the processor}
- [post_start_hook] is called just after the container process
has been created, but now for the copy of the processor object
that lives in the container process. This is a very useful hook
method, because one can initialize the container process
(e.g. prepare database accesses etc.).
- [pre_finish_hook] is called just before the container process
will (regularly) terminated.
- [receive_message] is called when a message from another
container arrives.
- [receive_admin_message] is called when a message from the
administrator arrives.
- [shutdown] is called when the shutdown notification arrives.
The shutdown will lead to the termination of the process
when all network connections managed by Unixqueue are finished.
This method must terminate such connections if they have been
created in addition to those Netplex manages. The [shutdown]
notification is generated whenever a container needs to be
stopped, for example when it has been idle for too long and is
probably not needed right now (workload-induced shutdown), or
when the whole system is stopped (administrative shutdown).
- [system_shutdown] is another shutdown-related notification.
It is only emitted if the whole Netplex system is going to be
stopped. In this case, all containers first receive the
[system_shutdown] notifications, so they can prepare the real
shutdown that will happen soon. At the time the [system_shutdown]
is emitted, the whole system is still up and running, and so every
action is still possible. Only after all containers have finished
their [system_shutdown] callbacks, the real shutdown begins, i.e.
[shutdown] notifications are sent out.
- [global_exception_handler] is called for exceptions falling through
to the container, and is the last chance to catch them.
Because of the instances in the controller and the containers it is
usually a bad idea to store state in the processor object.
If multi-threading is used instead of multi-processing, there is only
one instance of the processor that is used in the controller and
all containers.
{2:defproc Defining Custom Processors}
Using predefined processor factories like {!Nethttpd_plex.nethttpd_factory}
is very easy. Fortunately, it is not very complicated to define a
custom adapter that makes an arbitrary network service available as
Netplex processor.
In principle, you must define a class for the type
{!Netplex_types.processor} and the corresponding factory implementing
the type {!Netplex_types.processor_factory}. To do the first,
simply inherit from {!Netplex_kit.processor_base} and override the
methods that should do something instead of nothing. For example,
to define a service that outputs the line "Hello world" on the
TCP connection, define:
{[
class hello_world_processor : processor =
let empty_hooks = new Netplex_kit.empty_processor_hooks() in
object(self)
inherit Netplex_kit.processor_base empty_hooks
method process ~when_done container fd proto_name =
Unix.clear_nonblock fd;
let ch = Unix.out_channel_of_descr fd in
output_string ch "Hello world\n";
close_out ch;
when_done()
method supported_ptypes = [ `Multi_processing; `Multi_threading ]
end
]}
The method [process] is called whenever a new connection is made.
The [container] is the object representing the container where the
execution happens ([process] is always called from the container).
In [fd] the file descriptor is passed that is the (already accepted)
connection. In [proto_name] the protocol name is passed - here it is
unused, but it is possible to process the connection in a way that
depends on the name of the protocol.
Note that the file descriptors created by Netplex are in non-blocking
mode. It is, however, possible to switch to blocking mode when this is
more appropriate ([Unix.clear_nonblock]).
The argument [when_done] is very important. It {b must} be called by
[process]! For a synchronous processor like this one it is simply called
before [process] returns to the caller.
For an asynchronous processor (i.e. a processor that handles several
connections in parallel in the same process/thread), [when_done] must
be called when the connection is fully processed. This may be at any
time in the future.
The class [hello_world_processor] can now be turned into a factory:
{[
class hello_world_factory : processor_factory =
object(self)
method name = "hello_world"
method create_processor ctrl_cfg cfg_file cfg_addr =
new hello_world_processor
end
]}
As you see, one can simply choose a [name]. This is the type of
the [processor] section in the configuration file, i.e. you need
{[
...
service {
name = "hello world sample";
...
processor {
type = "hello_world"
};
...
}
...
]}
to activate this factory for a certain service definition. Of course,
the instantiated [hello_world_factory] must also be passed to
{!Netplex_main.startup} in order to be available at runtime.
The [create_processor] method simply creates an object of your class. The
argument [ctrl_cfg] is the configuration of the controller (e.g.
you find there the name of the socket directory). In [cfg_file]
the object is passed that accesses the configuration file as
tree of parameters. In [cfg_addr] the address of the [processor]
section is made available, so you can look for additional
configuration parameters.
You may wonder why it is necessary to first create [empty_hooks].
The hook methods are often overridden by the user of processor
classes. In order to simplify this, it is common to allow the
user to pass a hook object to the processor object:
{[
class hello_world_processor hooks : processor =
object(self)
inherit Netplex_kit.processor_base hooks
method process ~when_done container fd proto_name = ...
method supported_ptypes = ...
end
]}
Now, the user can simply define hooks as in
{[
class my_hooks =
object(self)
inherit Netplex_kit.empty_processor_hooks()
method post_start_hook container = ...
end
]}
and pass such a hook object into the factory.
{2:workload Workload Management}
Workload managers decide when to start new containers and when to stop
useless ones. The simplest manager is created by the
{!Netplex_workload.constant_workload_manager_factory}. The user
simply defines how many containers are to be started. In the
config file this is written as
{[
workload_manager {
type = "constant";
threads = <n>;
}
]}
where [<n>] is the number of containers > 0. Often this manager is used
to achieve n=1, i.e. to have exactly one container. An example would
be a stateful RPC server where it is important that all network connections
are handled by the same process. (N.B. n=1 for RPC servers does not
enforce that the connections are serialized because Ocamlnet RPC servers
can handle multiple connections in parallel, but of course it is enforced
that the remote procedures are invoked in a strictly sequential way.)
If n>1, it is tried to achieve that all containers get approximately
the same load.
If processes die unexpectedly, the constant workload manager starts
new components until the configured number of processes is again reached.
The dynamic workload manager (created by
{!Netplex_workload.dynamic_workload_manager_factory}) is able to start
and stop containers dynamically. There are a few parameters that control
the manager. A "thread" is here another word for a started container.
A "job" is an established network connection. Using this terms, the
task of the workload manager is to decide how many threads are needed
to do a varying number of jobs. The parameters now set how many jobs
every thread may execute, and how quickly new threads are created or
destroyed to adapt the available thread capacity to the current job
load.
If the service processor can only accept one network connection after
the other (like Nethttpd_plex), the only reasonable setting is that
there is at most one job per thread. If one configures a higher number
in this case, unaccepted network connections will queue up resulting
in poor performance.
If the service processor can handle several connections in parallel it
is possible to allow more than one job per thread. There is no general
rule how many jobs per thread are reasonable, one has to experiment to
find it out. In this mode of having more than one job per thread,
Netplex even allows two service qualities, "normal" and "overload". If
possible, Netplex tries to achieve that all containers deliver normal
quality, but if the load goes beyond that, it is allowed that
containers accept more connections than that. This is called an
overload situation. Often it is better to allow overload than to
refuse new connections.
The dynamic workload manager is enabled by the section
{[
workload_manager {
type = "dynamic";
... parameters, see below ...
}
]}
The required parameters are:
- [max_threads]: How many containers can be created at maximum for
this service.
- [max_jobs_per_thread]: How many jobs every container can execute
at maximum. The upper limit for the number of jobs is thus
[max_threads * max_jobs_per_thread].
- [min_free_job_capacity]: This parameter controls how quickly new
containers are started when the load goes up. It is tried to
ensure that there are as many containers so this number of jobs
can be additionally performed. This parameter must
be at least 1.
- [max_free_job_capacity]: This parameter controls how quickly
containers are stopped when the load goes down. It is tried to
ensure that unused containers are stopped so the capacity for
additional jobs is not higher than this parameter.
This parameter must be greater or equal than
[min_free_job_capacity].
In order to configure the overload mode:
- [recommended_jobs_per_thread]: The number of jobs a container
can do with normal service quality. A higher number is considered
as overload.
The effect of this parameter is that it is avoided that a container
gets more jobs than recommended as long as possible.
Another parameter is:
- [inactivity_timeout]: If a container idles longer than this number
of seconds and is not needed to ensure [min_free_job_capacity] it is
shut down. Defaults to 15 seconds.
More documentation is available in {!Netplex_admin.workload}.
{2:messaging Messaging}
There are two kinds of messages one can send to Netplex containers:
normal messages come from another Netplex container, and admin messages
are sent using the [netplex-admin] command.
Messages have a name and a (possibly empty) list of string parameters.
They can be sent to an individual receiver container, or to a number
of containers, even to all. The sender does not get an acknowledgment
when the messages are delivered.
Messages can e.g. be used
- to signal that internal state is output to log files in order to debug
a special situation
- to enable or disable special features of the running system
- to flush caches
and for other comparably simple communication needs.
In order to receive a normal message, one must define the
[receive_message] method in the processor object, and to receive an
admin message, one must define the [receive_admin_message] method.
A normal message is sent by the container method [send_message].
The receiver is identified by the service name, i.e. all containers
with the passed name get the message. The name may even contain
the wildcard [*] to select the containers by a name pattern.
An admin message is sent using the [netplex-admin] command.
There are a few predefined messages understood by all containers.
See {!Netplex_admin.admin} for a list.
In general, messages starting with "netplex." are reserved for
Netplex itself.
{2:logging Logging}
Log messages can be written in the containers. The messages are first
sent to the controller where they are written to stderr, to files, or
to any object of the type {!Netplex_types.logger}. That the messages
are first sent to the controller has a lot of advantages: The messages
are implicitly serialized, no locking is needed, and it is easy to
support log file rotation.
In order to write a log message, one needs the container object.
The module [Netplex_cenv] always knows the container object of the
caller, to get it:
{[
let cont = Netplex_cenv.self_cont()
]}
If you call [self_conf] outside a container, the exception
{!Netplex_cenv.Not_in_container_thread} is raised. This is e.g. the
case if you call it from the [pre_start] or [post_finish] callbacks.
Logging is now done by
{[
let cont = Netplex_cenv.self_cont() in
cont # log level message
]}
where [level] is one of [`Debug], [`Info], [`Notice], [`Warning], [`Err],
[`Crit], [`Alert], [`Emerg], and [message] is a string. The levels are
the same as for syslog.
You can also call {!Netplex_cenv.log} and {!Netplex_cenv.logf},
which simply use [self_cont] to get the container and call its [log]
method to write the message.
The config file controls what to do with the log messages. The easiest
way is to send all messages to stderr:
{[
controller {
max_level = "debug"; (* Log level *)
logging {
type = "stderr"; (* Log to stderr *)
}
};
]}
Further types of logging are documented in {!Netplex_admin.controller}.
{2:debug Debug logging}
There are various built-in debug logging streams:
- {!Netplex_container.Debug.enable}: Logs the perspective of the
container. Logged events are e.g. when connections are accepted,
and when user-defined hook functions are invoked. These messages
are quite interesting for debugging user programs.
- {!Netplex_controller.Debug.enable}: Logs the perspective of the
controller. The events are e.g. state changes, when containers
are started, and scheduling decisions. This is less interesting
to users, but might nevertheless worth activating it.
- {!Netplex_workload.Debug.enable}: Outputs messages from the
workload manager.
For these messages, the mechanism of {!Netlog.Debug} is used - which has
the advantage that messages can also be generated when no Netplex logger
is available.
{2:mt Compatibility with multi-threading}
As already noted above, it is possible to use Netplex with two
concurrency models, namely multi-threading and multi-processing.
In the first case, for every container a new thread is started, and
in the second case, a new process.
You should know that multi-threading in OCaml has some limitations. In
particular, it is not possible to exploit more than one CPU core of
the system (although the threads are real kernel threads). Because of
this, multi-processing is also supported by Netplex. It does not have
this deficiency, because every process executes an independent OCaml
runtime.
Nevertheless, it can be quite useful to just start helper threads, even
if the main concurrency model is multi-processing. The question is how
to do, and where the traps are.
First, one warning ahead: Once you have started threads, it is no longer
safe to spawn new processes. This is a general difficulty of the POSIX
system API, and cannot be worked around in OCaml programs (to some extent
it is possible in C). The immediate consequence is that you must not start
threads in the controller process (assumed you have multi-processing). The
controller process continuously forks new container processes, and having
threads could be deadly for it. For the API exposed by Netplex this means
that doing such is not supported at all - all controller-specific APIs
assume that they are only called from the same thread.
Ok, so you just don't this. However, is it allowed to have additional
threads in containers? This is in deed not problematic. Just be sure that
you start these threads after the container process has been created, e.g.
from [post_start_hook]. In order to make life simpler, a number of APIs
have now been made thread-safe (since Ocamlnet-3.5):
- All methods of container objects: {!Netplex_types.container}
- All functions in {!Netplex_cenv}
- All synchronization helpers in {!Netplex_semaphore}, {!Netplex_mutex},
and {!Netplex_sharedvar}
This means you can call these methods and functions without any
additional synchronization effort.
Note that all extra threads are implicitly killed when the
container process finishes. If you don't like this and want to
terminate them in a more controlled manner, you can do this from the
[pre_finish_hook]. If your main concurrency model is multi-threading,
though, no such killing will occur - your helper threads will simply
continue running when the container is finished. (Unfortunately, this
difference between the two models is unavoidable.)
{1:rpc_netplex Netplex RPC systems}
A short description how to build systems of RPC services is given
in {!Rpc_intro.rpc_netplex}.
