Nikolaos Bezirgiannis Google+ Twitter Github RSS

Building a PDES in Erlang(Part 1)

by Nikolaos Bezirgiannis on January 26, 2013

Tagged as: simulation, PDES, erlang.


I’ll go ahead and implement a parallel and distributed discrete-event simulation (PDES) in the Erlang language. But first, let me introduce you and give you some notes on the theory of simulation.

Simulation is the process of ‘imitating’ the behaviour of a particular system. To do that, we first construct an abstract model of the system. Why we do that, you say? Well, in certain circumstances measuring the performance of an actual system is extremely difficult to do; reasons can be the costs for building such a system or the time required to execute it. These systems can be too complex to compute an analytical solution for. Instead, what we do, is simulate the system by estimating the performance of its constructed model.

A model of a system contains the algorithms and/or equations that best describe the system’s behaviour. We don’t need to construct an elaborate specification of the system, because we risk overcomplicating the problem and thus making it intractable. We only require a ‘simplified’ version of the system. On the other hand, we must be careful not to oversimplify it, as we might lose in the detail of the model’s outcome.

There are many types of simulations: pen&paper vs computer simulation, stochastic vs deterministic, steady-state vs dynamic, continuous vs discrete, local vs distributed. Not to stress you with extra terms, here, I will only consider computer distributed discrete-event simulations. As you probably guessed it, we are going to run these simulations using computers. That is, we will write a model using a computer language and execute it on a machine. The simulations are going to be distributed; that means, we will comprise a bunch of multiple processing units (being distributed over a network or using SMP technology). Each processing unit, called the processor or logical process, will itself be an instance of a simulation engine. The simulation engine’s role is to receive events in discrete time, process them in timestamp order, and schedule new events in the future to itself or other logical processes (for example by sending an event message to a remote simulation engine). The simulation engine stops when there are no more events to process, or a certain condition is met.

Why did I choose Erlang to implement such a simulation? Well, firstly Erlang is a functional language and I, in general, like functional languages :). Secondly, the difficult part of implementing a simulation engine is not how to process the incoming events (they are just linked to arbitrary code which gets executed by the engine), but more how to easily communicate between logical processes (sending and receiving events). This can be easily done in Erlang, since message-passing is a first-class citizen of the language. Events are simply modelled as Erlang messages, and logical processes, likewise, are implemented as Erlang processes. Another reason is that, in Erlang, running on many processors (SMP) or on multiple distributed machines is transparent, that is we do not have to write extra code to handle these distinct cases.

When talking about PDES we have to be clear about what synchronization approach we make use of. It is the case that, in parallel and distributed simulations certain conflicts will arise that must be resolved. I’m not getting into much detail on this, I have to better direct you to the excellent book on PDES 1. There are three different synchronization approaches: the conservative approach, that avoids conflicts at all costs, the optimistic that allows conflicts to happen but later has to go back and correct them, and the mixed approach that employs the conservative on some and optimistic on other logical processes. In my implementation, I will use a conservative non-zero-lookahead mechanism, influenced by the Chandy/Misra/Bryant null message protocol algorithm 2.

Conservative mechanisms are easier to implement by the simulation developer, but require extra (lookahead) information from the end-user; optimistic mechanisms on the other hand don’t require such information by the end-user, but are much more difficult to implement. We can say that, in most cases, an optimistic approach is faster in execution time than a conservative simulation. Here, however, for the sake of simplicity, I demonstrate a conservative mechanism written in Erlang.

ErlangTW is another similar simulation middleware written in Erlang, although it instead follows an optimistic approach. You can find more information on their recently published paper. Their implementation is hosted on GitHub, and I have to say that they provide a clean and easy-to-grasp codebase.

μsik3 is a classic simulation microkernel written in C++. The advantage of μsik is that it can dynamically alter the deployed synchronization mechanism of the simulation to conservative, optimistic or mixed. It looks like that the microkernel and its kernel processes emulate how an Erlang VM actually works and communicates with other machines.

How it is going to look like

In action, the simulation program will be comprised of two entities: the simulation application and the simulation engine. The simulation application has the model specification in it, not mathematically defined, but rather through a computer language.
The simulation engine is also written in a programming language (since it has to be executed); it takes the the simulation application as input and runs it accordingly. The two entities can, but don’t have to, be hosted on the same programming language. In this case, I’m choosing Erlang for both the simulation application and the simulation engine.

What follows is an example definition of a simulation application. It is an Erlang module, that follows the OTP principles and thus is defined as a sim_proc behaviour (more about this later, in a different post). What the application is responsible for is the definition of state (the state record), simulation initialization (the init function), a series of callbacks (handle_event function) and simulation termination cleanup (the terminate function). The event callbacks are simply associating a possible incoming event to specific code that should be executed. In this example, the simulation application is modelling an airport, which schedules arrival, landed and departure events. I am not going into detail here; more on this in the followup, where I’m going to talk about the structure and the building blocks of a simulation application.


%% constants
-define(R, 10).
-define(G, 5).

%% state variables
-record(state, {in_the_air,

init(_Args) ->
    %% initialize state_variables
    State = #state{in_the_air = 0, 
                   on_the_ground = 0, 
                   runway_free = true},

    %% schedule initial events
    sim_proc:schedule(arrival, 30),
    sim_proc:schedule(arrival, 10),

    %% create links

    %% set correct lookahead
    sim_proc:link_to(lax, 3),

    {ok, State, 70}.

handle_event(arrival, State) ->
    In_the_air_ = State#state.in_the_air + 1,
    Runway_free_ = case State#state.runway_free of
                       true -> sim_proc:schedule(landed, ?R),
                       false -> false
    {ok, State#state{in_the_air = In_the_air_, runway_free = Runway_free_}};

handle_event(landed, State) ->
    In_the_air_ = State#state.in_the_air - 1,
    On_the_ground_ = State#state.on_the_ground + 1,
    sim_proc:schedule(departure, ?G),
    Runway_free_ = case In_the_air_ > 0 of
                       true -> sim_proc:schedule(landed, ?R),
                       false -> true
    {ok, State#state{in_the_air = In_the_air_, on_the_ground = On_the_ground_, runway_free = Runway_free_}};

handle_event(departure, State) ->
    On_the_ground_ = State#state.on_the_ground - 1,
    sim_proc:schedule(lax, arrival, 5),

    {ok, State#state{on_the_ground = On_the_ground_}};

terminate(normal, _State) ->
    sim_proc:println("Finished simulation");
terminate(timeout, _State) ->
    sim_proc:println("Timeout reached").

  1. Fujimoto, R. M. “Parallel Simulation: Parallel and Distributed Simulation Systems.” In Proceedings of the 33nd Conference on Winter Simulation, 147–157, 2001.

  2. Misra, J. “Distributed Discrete-event Simulation.” ACM Computing Surveys (CSUR) 18, no. 1 (1986): 39–65.

  3. Perumalla, K.S. “Mu;sik - a Micro-kernel for Parallel/distributed Simulation Systems.” In Workshop on Principles of Advanced and Distributed Simulation, 2005. PADS 2005, 59 – 68, 2005. doi:10.1109/PADS.2005.1.

comments powered by Disqus