.. include:: replace.txt
.. highlight:: cpp


.. heading hierarchy:
   ------------- Chapter
   ************* Section (#.#)
   ============= Subsection (#.#.#)
   ############# Paragraph (no number)

Events and Simulator
--------------------

|ns3| is a discrete-event network simulator.  Conceptually, the simulator
keeps track of a number of events that are scheduled to execute at a
specified simulation time.  The job of the simulator is to execute the
events in sequential time order.  Once the completion of an event occurs,
the simulator will move to the next event (or will exit if there are no
more events in the event queue).  If, for example, an event scheduled
for simulation time "100 seconds" is executed, and the next event is not
scheduled until "200 seconds", the simulator will immediately jump from
100 seconds to 200 seconds (of simulation time) to execute the next event.
This is what is meant by "discrete-event" simulator.

To make this all happen, the simulator needs a few things:

1) a simulator object that can access an event queue where events are
   stored and that can manage the execution of events
2) a scheduler responsible for inserting and removing events from the queue
3) a way to represent simulation time
4) the events themselves

This chapter of the manual describes these fundamental objects
(simulator, scheduler, time, event) and how they are used.

Event
*****

An event represents something that changes the simulation status, i.e.,
between two events the simulation status does not change, and the event
will likely change it (it could also not change anything).

Note that another way to understand an event is to consider it as a delayed
function call. With the due differences, a discrete event simulation is not
much different from a "normal" program where the functions are not called
immediately, but are marked with a "time", and the time is used to decide
the order of the functions execution.

The time, of course, is a simulated time, and is quite different from the
"real" time. Depending on the simulation complexity the simulated time
can advance faster or slower then the "real" time, but like a "real" time
can only go forward.

An example of an event is the reception of a packet, or the expiration
of a timer.

An event is represented by:

* The time at which the event will happen
* A pointer to the function that will "handle" the event,
* The parameters of the function that will handle the event (if any),
* Other internal structures.

An event is scheduled through a call to ``Simulator::Schedule``, and once
scheduled, it can be canceled or removed.
Removal implies removal from the scheduler data structure, while cancel
keeps them in the data structure but sets a boolean flag that suppresses
calling the bound event function at the scheduled time.  When an event is
scheduled by the Simulator, an ``EventId`` is returned.  The client may use
this event ID to later cancel or remove the event; see the example program
``src/core/examples/sample-simulator.{cc,py}`` for example usage.
Cancelling an event is typically less computationally expensive than
removing it, but cancelled events consumes more memory in the scheduler
data structure, which might impact its performances.

Events are stored by the simulator in a scheduler data
structure.  Events are handled in increasing order of
simulator time, and in the case of two events with the same
scheduled time, the event with the lowest unique ID (a
monotonically increasing counter) will be handled first.
In other words tied events are handled in FIFO order.

Note that concurrent events (events that happen at the very same time)
are unlikely in a real system - not to say impossible. In |ns3|
concurrent events are common for a number of reasons, one of them
being the time representation. While developing a model this must
be carefully taken into account.

During the event execution, the simulation time will not advance, i.e., each
event is executed in zero time. This is a common assumption in
discrete event simulations, and holds when the computational complexity of
the operations executed in the event is negligible.
When this assumption does not hold, it is necessary to schedule a second event
to mimic the end of the computationally intensive task.

As an example, suppose to have a device that receives a packet and has to
perform a complex analysis on it (e.g., an image processing task). The
sequence of events will be:

* T(t) - Packet reception and processing, save the result somewhere, and
  schedule an event in (t+d) marking the end of the data processing.
* T(t+d) - Retrieve the data, and do other stuff based them.

So, even if the data processing actually did return a result in the
execution of the first event, the data is considered valid only after
the second event.

The image below can be useful to clarify the idea.

.. image:: figures/time-consuming-event-handling.png


Simulator
*********

The Simulator class is the public entry point to access event scheduling
facilities. Once a couple of events have been scheduled to start the
simulation, the user can start to execute them by entering the simulator
main loop (call ``Simulator::Run``). Once the main loop starts running, it
will sequentially execute all scheduled events in order from oldest to
most recent until there are either no more events left in the event
queue or Simulator::Stop has been called.

To schedule events for execution by the simulator main loop, the
Simulator class provides the Simulator::Schedule* family of functions.

1) Handling event handlers with different signatures

These functions are declared and implemented as C++ templates to handle
automatically the wide variety of C++ event handler signatures used in
the wild. For example, to schedule an event to execute 10 seconds in the
future, and invoke a C++ method or function with specific arguments, you
might write this:

::

   void handler(int arg0, int arg1)
   {
     std::cout << "handler called with argument arg0=" << arg0 << " and
        arg1=" << arg1 << std::endl;
   }

   Simulator::Schedule(Seconds(10), &handler, 10, 5);

Which will output:

.. sourcecode:: text

   handler called with argument arg0=10 and arg1=5

Of course, these C++ templates can also handle transparently member
methods on C++ objects:

*To be completed:  member method example*

Notes:

* the |ns3| Schedule methods recognize automatically functions and
  methods only if they take less than 5 arguments. If you need them to
  support more arguments, please, file a bug report.
* Readers familiar with the term 'fully-bound functors' will recognize
  the Simulator::Schedule methods as a way to automatically construct such
  objects.

2) Common scheduling operations

The Simulator API was designed to make it really simple to schedule most
events. It provides three variants to do so (ordered from most commonly
used to least commonly used):

* Schedule methods which allow you to schedule an event in the future
  by providing the delay between the current simulation time and the
  expiration date of the target event.
* ScheduleNow methods which allow you to schedule an event for the
  current simulation time: they will execute _after_ the current event is
  finished executing but _before_ the simulation time is changed for the
  next event.
* ScheduleDestroy methods which allow you to hook in the shutdown
  process of the Simulator to cleanup simulation resources: every
  'destroy' event is executed when the user calls the Simulator::Destroy
  method.

3) Maintaining the simulation context

There are two basic ways to schedule events, with and without *context*.
What does this mean?

::

  Simulator::Schedule(Time const &time, MEM mem_ptr, OBJ obj);

vs.

::

  Simulator::ScheduleWithContext(uint32_t context, Time const &time, MEM mem_ptr, OBJ obj);

Readers who invest time and effort in developing or using a non-trivial
simulation model will know the value of the |ns3| logging framework to
debug simple and complex simulations alike. One of the important
features that is provided by this logging framework is the automatic
display of the network node id associated with the 'currently' running
event.

The node id of the currently executing network node is in fact tracked
by the Simulator class. It can be accessed with the
Simulator::GetContext method which returns the 'context' (a 32-bit
integer) associated and stored in the currently-executing event. In some
rare cases, when an event is not associated with a specific network
node, its 'context' is set to 0xffffffff.

To associate a context to each event, the Schedule, and ScheduleNow
methods automatically reuse the context of the currently-executing event
as the context of the event scheduled for execution later.

In some cases, most notably when simulating the transmission of a packet
from a node to another, this behavior is undesirable since the expected
context of the reception event is that of the receiving node, not the
sending node. To avoid this problem, the Simulator class provides a
specific schedule method: ScheduleWithContext which allows one to
provide explicitly the node id of the receiving node associated with
the receive event.

*XXX: code example*

In some very rare cases, developers might need to modify or understand
how the context (node id) of the first event is set to that of its
associated node. This is accomplished by the NodeList class: whenever a
new node is created, the NodeList class uses ScheduleWithContext to
schedule a 'initialize' event for this node. The 'initialize' event thus executes
with a context set to that of the node id and can use the normal variety
of Schedule methods. It invokes the Node::Initialize method which propagates
the 'initialize' event by calling the DoInitialize method for each object
associated with the node. The DoInitialize method overridden in some of these
objects (most notably in the Application base class) will schedule some
events (most notably Application::StartApplication) which will in turn
scheduling traffic generation events which will in turn schedule
network-level events.

Notes:

* Users need to be careful to propagate DoInitialize methods across objects
  by calling Initialize explicitly on their member objects
* The context id associated with each ScheduleWithContext method has
  other uses beyond logging: it is used by an experimental branch of |ns3|
  to perform parallel simulation on multicore systems using
  multithreading.

The Simulator::* functions do not know what the context is: they
merely make sure that whatever context you specify with
ScheduleWithContext is available when the corresponding event executes
with ::GetContext.

It is up to the models implemented on top of Simulator::* to interpret
the context value. In |ns3|, the network models interpret the context
as the node id of the node which generated an event. This is why it is
important to call ScheduleWithContext in ns3::Channel subclasses
because we are generating an event from node i to node j and we want
to make sure that the event which will run on node j has the right
context.

Available Simulator Engines
===========================

|ns3| supplies two different types of basic simulator engine to manage
event execution.  These are derived from the abstract base class `SimulatorImpl`:

*  `DefaultSimulatorImpl`  This is a classic sequential discrete event
   simulator engine which uses a single thread of execution.  This engine
   executes events as fast as possible.
*  `DistributedSimulatorImpl` This is a classic YAWNS distributed ("parallel")
   simulator engine. By labeling and instantiating your model components
   appropriately this engine will execute the model in parallel across many
   compute processes, yet in a time-synchronized way, as if the model had
   executed sequentially. The two advantages are to execute models faster
   and to execute models too large to fit in one compute node.  This engine also
   attempts to execute as fast as possible.
*  `NullMessageSimulatorImpl`  This implements a variant of the Chandy-
   Misra-Bryant (CMB) null message algorithm for parallel simulation.
   Like `DistributedSimulatorImpl` this requires appropriate labeling and
   instantiation of model components. This engine attempts to execute
   events as fast as possible.

You can choose which simulator engine to use by setting a global variable,
for example::

  GlobalValue::Bind("SimulatorImplementationType",
                    StringValue("ns3::DistributedSimulatorImpl"));

or by using a command line argument

.. sourcecode:: console

  $ ./ns3 run "...  --SimulatorImplementationType=ns3::DistributedSimulatorImpl"

In addition to the basic simulator engines there is a general facility used
to build "adapters" which provide small behavior modifications to one of
the core `SimulatorImpl` engines.  The adapter base class is
`SimulatorAdapter`, itself derived from `SimulatorImpl`.  `SimulatorAdapter`
uses the `PIMPL (pointer to implementation) <https://en.cppreference.com/w/cpp/language/pimpl>`_
idiom to forward all calls to the configured base simulator engine.
This makes it easy to provide small customizations
just by overriding the specific Simulator calls needed, and allowing
`SimulatorAdapter` to handle the rest.

There are few places where adapters are used currently:

*  `RealtimeSimulatorImpl`  This adapter attempts to execute in real time
   by pacing the wall clock evolution.  This pacing is "best effort",
   meaning actual event execution may not occur exactly in sync, but
   close to it. This engine is normally only used with the
   `DefaultSimulatorImpl`, but it can be used to keep a distributed
   simulation synchronized with real time.  See the :doc:`realtime` chapter.
*  `VisualSimulatorImpl`  This adapter starts a live visualization of the
   running simulation, showing the network graph and each packet traversing
   the links.
*  `LocalTimeSimulatorImpl`  This adapter enables attaching noisy local clocks
   to `Nodes`, then scheduling events with respect to the local noisy clock,
   instead of relative to the true simulator time.

In addition to the PIMPL idiom of `SimulatorAdapter` there is a special
per-event customization hook::

  SimulatorImpl::PreEventHook( const EventId & id)

One can use this to perform any housekeeping actions before the next event
actually executes.

The distinction between a core engine and an adapter is the following: there
can only ever be one core engine running, while there can be several adapters
chained up each providing a variation on the base engine execution.
For example one can use noisy local clocks with the real time adapter.

A single adapter can be added on top of the `DefaultSimulatorImpl` by the same
two methods above: binding the `"SimulatorImplementationType"` global value or
using the command line argument.  To chain multiple adapters a different
approach must be used; see the `SimulatorAdapter::AddAdapter()`
API documentation.

The simulator engine type can be set once, but must be set before the
first call to the `Simulator()` API.  In practice, since some models have
to schedule their start up events when they are constructed, this means
generally you should set the engine type before instantiating any other
model components.

The engine type can be changed after `Simulator::Destroy()` but before
any additional calls to the Simulator API, for instance when executing
multiple runs in a single |ns3| invocation.


Time
****

|ns3| internally represents simulation times and durations as
64-bit signed integers (with the sign bit used for negative durations).
The time values are interpreted with respect to a "resolution" unit in the
customary SI units: fs, ps, ns, us, ms, s, min, h, d, y.
The unit defines the minimum Time value.
It can be changed once before any calls to `Simulator::Run()`.
It is not stored with the 64-bit time value itself.

Times can be constructed from all standard numeric types
(using the configured default unit)
or with explicit units (as in `Time MicroSeconds (uint64_t value)`).
Times can be compared, tested for sign or equality to zero, rounded to
a given unit, converted to standard numeric types in specific units.
All basic arithmetic operations are supported
(addition, subtraction, multiplication or division
by a scalar (numeric value)). Times can be written to/read from IO streams.
In the case of writing it is easy to choose the output unit, different
from the resolution unit.


Scheduler
*********

The main job of the `Scheduler` classes is to maintain the priority queue of
future events.  The scheduler can be set with a global variable,
similar to choosing the `SimulatorImpl`::

  GlobalValue::Bind("SchedulerType",
                    StringValue("ns3::DistributedSimulatorImpl"));

The scheduler can be changed at any time via `Simulator::SetScheduler()`.
The default scheduler is `MapScheduler` which uses a `std::map<>` to
store events in time order.

Because event distributions vary by model there is no one
best strategy for the priority queue, so |ns3| has several options with
differing tradeoffs.  The example `utils/bench-scheduler.c` can be used
to test the performance for a user-supplied event distribution.
For modest execution times (less than an hour, say) the choice of priority
queue is usually not significant; configuring the build type to optimized
is much more important in reducing execution times.

The available scheduler types, and a summary of their time and space
complexity on `Insert()` and `RemoveNext()`, are listed in the
following table.  See the individual Scheduler API pages for details on the
complexity of the other API calls.

+------------------------+-------------------------------------+-------------+--------------+----------+--------------+
|  Scheduler Type                                              | Complexity                                           |
+------------------------+-------------------------------------+-------------+--------------+----------+--------------+
|                        |                                     | Time                       | Space                   |
|  `SchedulerImpl` Type  |               Method                +-------------+--------------+----------+--------------+
|                        |                                     | Insert()    | RemoveNext() | Overhead |  Per Event   |
+========================+=====================================+=============+==============+==========+==============+
| CalendarScheduler      | `<std::list> []`                    | Constant    | Constant     | 24 bytes | 16 bytes     |
+------------------------+-------------------------------------+-------------+--------------+----------+--------------+
| HeapScheduler          | Heap on `std::vector`               | Logarithmic | Logarithmic  | 24 bytes | 0            |
+------------------------+-------------------------------------+-------------+--------------+----------+--------------+
| ListScheduler          | `std::list`                         | Linear      | Constant     | 24 bytes | 16 bytes     |
+------------------------+-------------------------------------+-------------+--------------+----------+--------------+
| MapScheduler           | `st::map`                           | Logarithmic | Constant     | 40 bytes | 32 bytes     |
+------------------------+-------------------------------------+-------------+--------------+----------+--------------+
| PriorityQueueScheduler | `std::priority_queue<,std::vector>` | Logarithmic | Logarithms   | 24 bytes | 0            |
+------------------------+-------------------------------------+-------------+--------------+----------+--------------+