correct

[book_gpu.git] / BookGPU / Chapters / chapter17 / ch17.tex

\chapterauthor{Guillaume Laville, Christophe Lang, Bénédicte Herrmann and Laurent Philippe}{Femto-ST Institute, University of Franche-Comte, France}
%\chapterauthor{Christophe Lang}{Femto-ST Institute, University of Franche-Comte, France}
\chapterauthor{Kamel Mazouzi}{Franche-Comte Computing Center, University of Franche-Comte, France}
\chapterauthor{Nicolas Marilleau}{UMMISCO, Institut de Recherche pour le Developpement (IRD), France}
%\chapterauthor{Bénédicte Herrmann}{Femto-ST Institute, University of Franche-Comte, France}
%\chapterauthor{Laurent Philippe}{Femto-ST Institute, University of Franche-Comte, France}

\newlength\mylen
\newcommand\myinput[1]{%
  \settowidth\mylen{\KwIn{}}%
  \setlength\hangindent{\mylen}%
  \hspace*{\mylen}#1\\}

\chapter{Implementing Multi-Agent Systems on GPU}
\label{chapter17}


\section{Introduction}
\label{ch17:intro}

In this chapter we introduce the use of Graphical Processing Units for
multi-agents based systems as an example of a not-so-regular
application that could benefit from the GPU computing
power. Multi-Agent Systems or MAS are a simulation paradigm used to
study the behavior of dynamic systems. Dynamic systems as physical
systems are often modeled by mathematical representations and their
dynamic behavior simulated by differential equations. The simulation
of the system thus often relay on the resolution of a linear system
that can be efficiently computed on a graphical processing unit as
shown in the preceding chapters. But when the behavior of the system
elements is not uniformly driven by the same law, when these elements
have their own behavior, the modeling process is too complex to rely
on formal expressions. In this context MAS is a recognized approach to
model and simulate systems where individuals have an autonomous
behavior that cannot be simulated by the evolution of a set of
variables driven by mathematical laws. MAS are often used to simulate
natural or collective phenomenons whose individuals are too numerous or
various to provide a unified algorithm describing the system
evolution. The agent-based approach is to divide these complex systems
into individual self-contained entities with their smaller set of
attributes and functions. But, as for mathematical simulations, when
the size of the Multi-Agent System (MAS) increases the need of computing
power and memory also increases. For this reason, multi-agent systems
should benefit from the use of distributed computing
architectures. Clusters and Grids are often identified as the main
solution to increase simulation performance but Graphical Processing
Units (GPU) are also a promising technology with an attractive
performance/cost ratio.

Conceptually a MAS\index{Multi-Agent System} is a distributed system as it favors the definition
and description of large sets of individuals, the agents, that can be
run in parallel. As a large set of agents could have the same behavior
a SIMD model should fit the simulation execution. Most of the
agent-based simulators are however designed with a sequential scheme
in mind and these simulators seldom use more than one core for their
execution. Due to simulation scheduling constraints, data sharing and
exchange between agents and the huge amount of interactions between
agents and their environment, it is indeed rather difficult to
distribute an agent based simulator, for instance, to take advantage of new
multi-threaded computer architectures. Thus, guidelines and tools
dedicated to MAS paradigm and HPC is now a need for other complex
system communities. Note that, from the described structure (large
number of agents sharing data), we can conclude that MAS would more
easily benefit from many-cores architectures than from other kinds of
parallelism.

Another key point that advocates for the use of many-core in MAS is
the growing need for multi-scale simulations. Multi-scale simulations
explore problems with interactions between several scales. The
different scales use different granularity of the structure and
potentially different models. Most of the time the lower scale
simulations provide results to higher scale simulations. In that case
the execution of the simulations can easily be distributed between the
local cores and a many-core architecture, i.e. a GPU device.

We explore in this chapter the use of many-core architectures to
execute agent-based simulations. We illustrate our reflexion with two
uses cases: the Colembola simulator designed to simulate the diffusion
of Colembola between plots of land and the MIOR simulator that
reproduces effects of earthworms on bacteria dynamics in a bulked
soil. In Section \ref{ch17:ABM} we present the work related to MAS and
parallelization with a special focus on many-core use. In sections
\ref{ch17:sec:1stmodel} and \ref{ch17:sec:2ndmodel} we present in
details two multi-agent models, their GPU implementations, the
conducted experiments and their performance results. The first model,
given in Section \ref{ch17:sec:1stmodel}, illustrates the use of a GPU
architecture to speed up the execution of some computation intensive
functions while the main model is still executed on the central
processing unit. The second model, given in Section
\ref{ch17:sec:2ndmodel}, illustrates the use of a GPU architecture to
implement the whole model on the GPU processor which implies deeper
changes in the initial algorithm. In Section \ref{ch17:analysis} we
propose a more general reflexion on these implementations and we
provide some guidelines. Then we conclude in Section
\ref{ch17:conclusion} on the possible generalization of our work.


\section{Running Agent-Based Simulations}
\label{ch17:ABM}

In this section, we present the context of multi-agent systems, MAS, 
their parallelization and we report several existing works on using
GPU to simulate multi-agent systems.

\subsection{Multi-agent systems and parallelism}

Agent-Based systems are often used to simulate natural or collective
phenomenons whose actors are too numerous or various to provide a
simple unified algorithm describing the studied system
dynamic~\cite{Schweitzer2003}. The implementation of an agent based
simulation usually starts by designing the underlying Agent Based
Model (ABM). Most ABM are based around a few types of entities such as Agents,
Environment and around an Interaction Organization~\cite{Vowel02}. In the
complex system domain, the environment often describes a real space,
its structure (e.g. soil textures and porosities) and its dynamics
(e.g. organic matter decomposition). It is a virtual world in which
agents represent studied entities (e.g. biotic organisms)
evolution. The actual agent is animated by a behavior that can range
between reactivity (only react to external stimuli) and cognition
(lives its own process alongside other individuals). Interaction
and Organization define functions, types and patterns of
communications of their member agents in the
system~\cite{Odell:2003:RRD:1807559.1807562}. Note that, depending on
the MAS, agents can communicate either directly through special
primitives or indirectly through the information stored in the
environment.

Agent based simulations have been used for more than one decade to
reproduce, understand even predict complex system dynamic. They have proved their interest in various
scientific communities. Nowadays generic agent based frameworks are
promoted such as Repast~\cite{repast_home} or
NetLogo~\cite{netlogo_home} to implement simulators. Many ABM such as
the crown model representing a city-wide
scale~\cite{Strippgen_Nagel_2009} tend however to require a large
number of agents to provide a realistic behavior and reliable global
statistics. Moreover, an achieved model analysis needs to run many
simulations identified into an experiment plan to obtain enough
confidence in a simulation. In this case the available computing power
often limits the simulation size and the result range thus requiring
the use of parallelism to explore bigger configurations.

For that, three major approaches can be identified:
\begin{enumerate}
\item parallelizing experiments plan on a cluster or a grid (one or
  few simulations are submitted to each
  core)~\cite{Blanchart11,Chuffart2010}.
\item parallelizing the simulator on a cluster (the environment of the
  MAS is split and run on several distributed
  nodes)~\cite{Cosenza2011,MAR06}
\item optimizing simulator by taking advantage of computer resources
  (multi-threading, GPU, and so on) \cite{Aaby10}
\end{enumerate}

In the first case, experiments are run independently each others and only simulation parameters change between two runs so
that a simple version of an existing simulator can be used. This
approach does not, however, allow to run larger models. 
% It does not imply
%model code changes except for Graphical Unit Interface extracting as
%the experiments does not run interactively but in batch. 
In the second
and the third case, model and code modifications are necessary. Only few
frameworks however introduce distribution in agent simulation
(Madkit~\cite{Gutknecht2000}, MASON~\cite{Sean05}, repastHPC~\cite{Collier11}) and parallel
implementations are often based on the explicit use of threads using
shared memory~\cite{Guy09clearpath} or cluster libraries such as
MPI~\cite{Kiran2010}.

Parallelizing a multi-agent simulation is however complex due to space
and time constraints. Multi-agent simulations are usually based on a
synchronous execution:
%by the agents that share the same environment. 
at each time step, numerous events (space data
modification, agent motion) and interactions between agents happen.
Distributing the simulation on several computers or grid nodes thus
implies to guaranty a distributed synchronous execution and
coherency. This often leads to poor performance or complex synchronization
problems. Multi-cores execution or delegating part of this execution
to others processors as GPUs~\cite{Bleiweiss_2008} are thus usually easier
to implement since all the threads share the data and the local clock.

% Different agent patterns can be adopted in an ABMs such as
% cognitive and reactive ones~\cite{Ferber99}. Cognitive agents act on
% the environment and interact with other agents according to a complex
% behavior. This behavior takes a local perception of the virtual world
% and the agent past (a memory characterized by an internal state and
% belief, imperfect knowledge about the world) into account. Reactive
% agents have a much systematic pattern of action based on stimuli
% response schemes (no or few knowledge and state conservation in agent). The
% evolution of the ABM environment, in particular, is often represented with
% this last kind of agents. As their behavior is usually simple, we
% propose in this chapter to delegate part of the environment and of the
% reactive agents execution to the graphical processing unit of the
% computer. This way we can balance the load between both CPU and GPU
% execution resources.

% In the particular case of multi-scale simulations such as the Sworm
% simulation~\cite{BLA09} the environment may be used at different
% levels. Since the representation of the whole simulated environment
% (i.e. the soil area) would be costly, the environment is organized as
% a multi-level tree of small soil cubes which can be lazily
% instantiated during the simulation. This allows to gradually refine
% distribution details in units of soil as agents progress and need
% those information, by using a fractal process based on the
% bigger-grained already instantiated levels. This characteristic,
% especially for a fractal model, could be the key of the
% distribution. For instance, each branch of a fractal environment could
% be identified as an independent area and parallelized. In addition
% Fractal is a famous approach to describe multi-scale environment (such
% as soil) and its organization~\cite{perrier}. In that case the lower
% scale simulations can also be delegated to the GPU card to limit the
% load of the main (upper scale) simulation.

\subsection{MAS Implementation on GPU}
\label{ch17:subsec:gpu}

The last few years saw the apparition of new generations of graphic
cards based on more general purpose execution units which are
promising for less regular systems as MAS. 
% These units can be programmed using GPGPU languages to solve various
% problems such as linear algebra resolutions, usually based on matrix
% manipulations.
The matrix-based data representation and SIMD computations are
however not straightforward in Multi-Agent Systems, where data
structures and algorithms are tightly coupled to the described
simulation. However, works from existing literature show that MAS can
benefit from these performance gains on various simulation types, such
as traffic simulation~\cite{Strippgen_Nagel_2009}, cellular
automatons~\cite{Dsouza2007}, mobile-agent based
path-finding~\cite{Silveira:2010:PRG:1948395.1948446} or genetic
algorithms~\cite{Maitre2009}. Note that an application-specific
adaptation process was required in the case of these MAS: some of the
previous examples are driven by mathematical laws (path-finding) or
use a natural mapping between a discrete environment (cellular
automaton) and GPU cores. Unfortunately, this mapping often requires
algorithmic adaptations in other models but the experience shows that
the more reactive a MAS is the more adapted its implementation is to
GPU.

The first step in the adaptation of an ABM to GPU platforms is the
choice of the language. On the one hand, the Java programming language
is often used for the implementation of MAS due to its availability
on numerous platforms or frameworks and its focus on high-level,
object-oriented programming. On the other hand, GPU platforms can
only run specific languages as OpenCL or CUDA. OpenCL (supported on
AMD, Intel and NVIDIA hardware) better suits the portability concerns
across a wide range of hardware needed the agent simulators, as
opposed to CUDA which is a NVIDIA-specific library.

OpenCL is a C library which provides access to the underlying CPUs or
GPUs threads using an asynchronous interface. Various OpenCL functions allow
the compilation and the execution of programs on these execution
resources, the copy of data buffers between devices, or the collection
of profiling information.

This library is based around three main concepts:

\begin{itemize}
\item the \emph{kernel} (similar to a CUDA kernel) which represents a runnable program
  containing instructions to be executed on the GPU.
\item the \emph{work-item} (equivalent to a CUDA thread), which is analogous to the concept
  of thread on GPU, in that it represents one running instance of a
  GPU kernel.
\item the \emph{work-group} (or execution block) which is a set of work-items
  sharing some memory to speed up data accesses and
  computations. Synchronization operations such as barrier can only be
  used across the same work-group.
\end{itemize}

Running an OpenCL computation consists in launching numerous work-items
that execute the same kernel. The work-items are submitted to a
submission queue to optimize the available cores usage. A calculus is
achieved once all these kernel instances have terminated.

The number of work-items used in each work-group is an important
implementation choice which determines how many tasks will share the
same cache memory. Data used by the work-items can be stored as
N-dimensions matrices in local or global GPU memory.  Since the size
of this memory is often limited to a few hundred of kilobytes,
choosing this number often implies a compromise between the model
synchronization or data requirements and the available resources.

In the case of agent-based simulations, each agent can be naturally
mapped to a work-item. Work-groups can then be used to represents groups
of agents or simulations sharing common data (such as the environment)
or algorithms (such as the background evolution process).

In the following examples the JOCL~\cite{jocl_home} binding is used
to access the OpenCL platform from the Java programming language.

In the next sections we present two practical cases that will be
studied in details, from the model to its implementation and 
performance. 

\section{A first practical example}
\label{ch17:sec:1stmodel}

The first model, the Collembola model, simulates the propagation of
collembolas in fields and forests. It is based on a diffusion
algorithm which illustrates the case of agents with a simple behavior
and few synchronization problems.

\subsection{The Collembola model\index{Collembola model}}
\label{ch17:subsec:collembolamodel}

The Collembola model is an example of multi-agent system using GIS (Geographical Information System)
 and survey data (population count)
 to model the evolution of the biodiversity
across land plots. A first version of this model has been developed with Netlogo framework by Bioemco and UMMISCO researchers. In this model, the biodiversity is modeled by populations of
athropod individuals, the Collembola, which can reproduce and diffuse
to favorable new habitats. The simulator allows to study the diffusion
of collembola, between plots of land depending on their use
(artifical soil, crop, forest\ldots). In this
model the environment is composed of the studied land and the
colembola are agents. Every land plot is divided into several cells,
each cell representing a surface unit (16*16 meters). A number of individuals per collembola species is associated to each cell. The model evolution
is then based on a common diffusion model that diffuse individuals between
cells. Each step of the simulation is based on four stages, as shown
on Figure~\ref{ch17:fig:collem_algorithm}:

% \begin{enumerate}
% \item arrival of new individuals
% \item reproduction in each cell
% \item diffusion between cells
% \item updating of colembola lifetime
% \end{enumerate}

\begin{figure}[h]
\centering
\includegraphics[width=0.6\textwidth]{Chapters/chapter17/figs/algo_collem.pdf}
\caption{Evolution algorithm of Collembola model}
\label{ch17:fig:collem_algorithm}
\end{figure}

The algorithm is quite simple but includes two costly operations, the
reproduction and the diffusion, that must be parallelized to improve
the model performances.

The {\bf reproduction} stage consists in updating the total population
of each plot by taking the individuals arrived at the preceding
computation step. This stage implies to process the whole set of cells
of the environment to sum their population. The computed value is
recorded in the plot associated to each cell. This process can be
assimilated to a reduction operation on all the population cells
associated to one plot to obtain its population.

The {\bf diffusion} stage simulates the natural behavior of the
collembola that tends toward occupying the whole space over time. This
stage consists in computing a new value for each cell depending on
the population of the neighbor cells. This process can be assimilated
to a linear diffusion at each iteration of the population of the cells
across their neighbors.

These two processes are quite common in numerical computations so
that the collembola model can be adapted to a GPU execution without
much difficulties.

\subsection{Collembola Implementation}

In the Collembola simulator biodiversity is modeled by populations
of Collembola individuals, which can reproduce and diffuse to
favorable new habitats. This is implemented in the algorithm as a fixed
reproduction factor, applied to the size of each population, followed
by a linear diffusion of each cell population to its eight
neighbors. These reproduction and diffusion processes are followed by
two more steps on the GPU implementation. The first one consist of
culling of populations in a inhospitable environment, by checking each
cell value, landuse, and setting its population to zero if
necessary.  The final simulation step is the reduction of the cells
populations for each plot, to obtain an updated plot population
for statistic purposes. This separate computation step, done while
updating each cell population in the reference sequential algorithm,
is motivated by synchronization problems and allow the reduction of
the total numbers of access needed to updated those populations.

%\lstinputlisting[language=C,caption=Collembola OpenCL
%kernels,label=fig:collem_kernels]{Chapters/chapter17/code/collem_kernels.cl}
\pagebreak
\lstinputlisting[caption=Collembola OpenCL Diffusion kernel,label=ch17:listing:collembola-diffuse]{Chapters/chapter17/code/collem_kernel_diffuse.cl}

The reproduction, diffusion and culling steps are implemented on GPU
(Figure~\ref{ch17:fig:collem_algorithm}) as a straight mapping of each cell
to an OpenCL work-item (GPU thread). Listing
\ref{ch17:listing:collembola-diffuse} gives the kernel for the
diffusion implementation.  To prevent data coherency problems, the
diffusion step is split into two phases, separated by an execution
{\it barrier}. In the first phase, each cell diffusion overflow is
calculated, and divided by the number of neightbors. Note that, at the
model frontier, populations can also overflow outside the environment
grid but we do not manage those external populations, since there are
no reason to assume our model to be isolated of its surroundings. The
overflow by neighbors value is stored for each cell, before
encountering the barrier. After the barrier is met, each cell read the
overflows stored by all its neighbors at the previous step and applies
them to its own population. In this manner, only one barrier is
required to ensure the consistency of populations value, since no
cells ever modify a value other than its own.

\lstinputlisting[caption=Collembola OpenCL reduction kernel,label=ch17:listing:collembola-reduc]{Chapters/chapter17/code/collem_kernel_reduc.cl}

Listing \ref{ch17:listing:collembola-reduc} gives the kernel for the
reduction implementation.  The only step requiring numerous
synchronized accesses is the reduction one: in a first approach, we
chose to use {\it atomic\_add} operation to implement this process, but more
efficient implementations using partial reduction and local GPU memory
could be implemented.

\subsection{Collembola performance}

In this part we present the performance of the Collembola model on
various CPU and GPU execution
platforms. Figure~\ref{ch17:fig:mior_perfs_collem} shows that the
number of cores and the processor architecture as a strong influence on the obtained results

% : the
% dual-core processor equipping our Thinkpad platform has two to six
% longer executions times, compared to a six-core Phenom X6. 

% % \begin{figure}[h]
% %begin{minipage}[t]{0.49\linewidth}
% \centering \includegraphics[width=0.7\linewidth]{./Chapters/chapter17/figs/collem_perfs.pdf}
% \caption{Performance CPU et GPU du modèle Collemboles}
% \label{ch17:fig:mior_perfs_collem}
% %end{minipage}
% \end{figure}

% In figure~\ref{ch17:fig:mior_perfs_collem2} the Thinkpad curve is removed
% to make other trends clearer. Two more observation can be made, using
% this more detailled scale:

\begin{itemize}
\item Older GPU cards can be slower than modern processors. This can
  be explained by the new cache and memory access optimizations
  implemented in newer generations of GPU devices. These optimizations
  reduce the penalties associated with irregular and frequent global
  memory accesses. They are not available on our Tesla nodes.
\item GPU curves exhibit a odd-even pattern in their performance
  results. Since this phenomenon is visible on two distinct
  manufacturer hardware, drivers and OpenCL implementation, it is
  likely the result of the decomposition process based on warp of
  fixed, power-of-two sizes.
\item The number of cores is not the only determining factor: an Intel
  Core i7 2600K processor, even with only four cores, can provide
  better performance than a Phenom one.
\end{itemize}

\begin{figure}[h]
%begin{minipage}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.7\linewidth]{./Chapters/chapter17/figs/collem_perfs_nothinkpad.pdf}
\caption{Performance of the Collembola model on CPU and GPU}
\label{ch17:fig:mior_perfs_collem}
%end{minipage}
\end{figure}

Both figures shows that using the GPU to parallelize part of the
simulator results in tangible performance gains over a CPU execution
on modern hardware. These gains are more mixed on older GPU platforms
due to the limitations when dealing with irregular memory or execution
patterns often encountered in MAS systems. This can be closely linked
to the availability of caching facilities on the GPU hardware and its
dramatically effects depending on the locality and frequency of data
accesses. In this case, even if the Tesla architectures provides more
execution cores and is the far costlier solution, more recent, cheaper,
solutions such as high-end GPU provide better performance when the
execution is not constrained by memory size.

\section{Second example}
\label{ch17:sec:2ndmodel}

The second model, the MIOR model, simulates the behavior of microbian
colonies. Its execution model is more complex so that it requires
changes in the initial algorithm and the use of synchronization to
benefit from the GPU architecture.

\subsection{The MIOR model}
\label{ch17:subsec:miormodel}

The MIOR~\cite{C.Cambier2007} model was developed to simulate local
interactions in a soil between microbial colonies and organic
matters. It reproduces each small cubic units ($0.002 m^3$) of soil as a
MAS.

Multiple implementations of the MIOR model have already been
realized, in Smalltalk and Netlogo, in 2 or 3 dimensions. The last
implementation, used in our work and referenced as MIOR in the
rest of the chapter, is freely accessible online as
WebSimMior~\footnote{http://www.IRD.fr/websimmior/}.

MIOR is based around two types of agents: (i) the Meta-Mior (MM),
which represents microbial colonies consuming carbon and (ii) the
Organic Matter (OM) which represents carbon deposits occurring in soil.

The Meta-Mior agents are characterized by two distinct behaviors:
\begin{itemize}
\item \emph{breath}: the action converts mineral carbon from the soil
  to carbon dioxide $CO_{2}$ that is released in the soil.
\item \emph{growth}: by this action each microbial colony fixes the
  carbon present in the environment to reproduce itself (augments its
  size). This action is only possible if the colony breathing needs
  where covered, i.e. enough mineral carbon is available.
\end{itemize}

These behaviors are described in Algorithm~\ref{ch17:seqalgo}.

\begin{algorithm}[h]
\caption{Evolution step of each Meta-Mior (microbial colony) agent} 
\label{ch17:seqalgo}
\KwIn{A static array $mmList$ of MM agents}
\myinput{A static array $omList$ of OM agents}
\myinput{A MIOR environment $world$}
$breathNeed \gets world.respirationRate \times mm.carbon$\;
$growthNeed \gets world.growthRate \times mm.carbon$\;
$availableCarbon \gets totalAccessibleCarbon(mm)$\;
\uIf{$availableCarbon > breathNeed$}{
  \tcc{ Breath }
  $mm.active \gets true$\;
  $availableCarbon \gets availableCarbon - consumCarbon(mm, breathNeed)$\;
  $world.CO2 \gets world.CO2 + breathNeed$\;
  \If{$availableCarbon > 0$}{
    \tcc{ Growth }
    $growthConsum \gets max(totalAccessCarbon(mm), growthNeed)$\;
    $consumCarbon(mm, growthConsum)$\;
    $mm.carbon \gets mm.carbon + growthConsum$\;
  }
}
\Else{
  $mm.active \gets false$
}
\end{algorithm}

Since this simulation takes place at a microscopic scale, a large
number of these simulations must be executed for each macroscopic
simulation step to model realistic-sized unit of soil. This lead to
large computing needs despite the small computation cost of each
individual simulation.


\subsection{MIOR Implementation}

As pointed out previously, the MIOR implementation implied more
changes for the initial code to be run on GPU.  As a first attempt, we
realized a simple GPU implementation of the MIOR simulator, with only
minimal changes to the CPU algorithm. Execution times showed the
inefficiency of this approach, and highlighted the necessity of
adapting the simulator to take advantage of the GPU execution
capabilities~\cite{lmlm+12:ip}. In this part, we show the main changes
which were realized to adapt the MIOR simulator on GPU architectures.

\subsubsection{Execution mapping on GPU}

\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Chapters/chapter17/figs/repartition.pdf}
\caption{Execution distribution retained on GPU}
\label{ch17:fig:gpu_distribution}
\end{figure}

Each MIOR simulation is represented by a work-group, and each agent by
a work-item. A kernel is in charge of the life cycle process for each
agent of the model. This kernel is executed by all the work-items of the simulation
each on its own gpu core.

The usage of one work-group for each simulation allows to easily
execute multiple simulations in parallel, as shown on
figure~\ref{ch17:fig:gpu_distribution}.  By taking advantage of the
execution overlap possibilities provided by OpenCL it then becomes
possible to exploit all the cores at the same time, even if an unique
simulation is to small to use all the available GPU cores. However,
the maximum size of a work-group is limited (to $512$), which only allows
us to execute one simulation per work-group when using $310$ threads
(number of OM in the reference model) to execute the simulation.

The usage of the GPU to execute multiple simulations is initiated by
the CPU. The later keeps total control of the simulator execution
flow. Thus, optimized scheduling policies (such as submitting kernels
in batch, or limiting the number of kernels, or asynchronously
retrieving the simulation results) can be defined to minimize the cost
related to data transfers between CPU and GPUs.

\subsubsection{Data structures translation}
\label{ch17:subsec:datastructures}

The adaptation of the MIOR model to GPU requires the mapping of the
data model to OpenCL data structures. The environment and the agents
are represented by arrays of structures, where each structure describes the
state of one entity. The behavior of these entities are implemented
as OpenCL functions to be called from the kernels during execution.

Since the main program is written in Java, JOCL is responsible for the
allocation and the mapping of the object data structures to OpenCL
ones before execution.

Four main data structures
(Figure~\ref{ch17:listing:mior_data_structures}) are defined: (i) an
array of MM agents, representing the state of the microbial
colonies. (ii) an array of OM agents, representing the state of the
carbon deposits. (iii) a topology matrix, which stores accessibility
information between the two types of agents of the model (iv) a world
structure, which contains all the global input data (metabolism rate,
numbers of agents) and output data (quantity of $CO_{2}$ produced) of
the simulation. These data structures are initialized by the CPU and
then copied on the GPU.

\lstinputlisting[caption=Main data structures used in a MIOR simulation,label=ch17:listing:mior_data_structures]{Chapters/chapter17/code/data_structures.cl}

The world topology is stored as a two-dimension matrix which contains
OM indexes on the abscissa and the MM index on the ordinate. Each
agent walks through its line/column of the matrix at each iteration to
determinate which agents can be accessed during the simulation.  Since
many agents are not connected this matrix is sparse, which introduces
a big number of useless memory accesses. To reduce the impact of these
memory accesses we use a compacted, optimized representation of this
matrix based on~\cite{Gomez-Luna:2009:PVS:1616772.1616869}, as
illustrated on the figure~\ref{ch17:fig:csr_representation}. This compact
representation considers each line of the matrix as an index list, and
only stores accessible agents continuously, to reduce the size of the
list and the number of non productive accesses.

\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{Chapters/chapter17/figs/grid.pdf}
%\psfig{file=figs/grid, height=1in}
\caption{Compact representation of the topology of a MIOR simulation}
\label{ch17:fig:csr_representation}
\end{figure}

Since dynamic memory allocation is not possible yet in OpenCL and only
provided in the latest revisions of the CUDA standard, these matrices
are statically allocated. The allocation is based on the worst-case
scenario where all OM and MM are linked as the real occupation of the
matrix cells cannot be deduced without some kind of preprocessing
computations.

\subsubsection{Critical resources access management}
\label{ch17:subsec:concurrency}

One of the main issue in a MIOR model is to ensure that all the
microbial colonies will have an equitable access to carbon resources,
when multiple colonies share the same deposits. Access
synchronizations are mandatory in these cases to prevent conflicting
updates on the same data that may lead to calculus error (e.g. loss of
matter).

On a massively parallel architecture such as GPUs, this kind of
synchronization conflicts can however lead to an inefficient
implementation by enforcing a quasi-sequential execution. It is
necessary, in the case of MIOR as well as for other ABM, to ensure
that each work-item is not too constrained in its execution.

\pagebreak
\lstinputlisting[caption=Main MIOR kernel,label=ch17:listing:mior_kernels]{Chapters/chapter17/code/mior_kernels.cl}

From the sequential algorithm 1 where all the agents share the same
data, we have developed a parallel algorithm composed of three
sequential stages separated by synchronization barriers. This new
algorithm is based on the distribution of the available OM carbon
deposits into parts at the beginning of each execution step. The three
stages, illustrated on Listing~\ref{ch17:listing:mior_kernels}, are the following:

\begin{enumerate}
\item \emph{distribution}: the available carbon in each carbon deposit
  (OM) is equitably dispatched between all accessible MM in the form
  of parts.
\item \emph{metabolism}: each MM consumes carbon in its allocated parts
  for its breathing and growing processes.
\item \emph{gathering}: unconsumed carbon in parts is gathered back
  into the carbon deposits.
\end{enumerate}

This solution suppresses the data synchronization needed by the first
algorithm and thus the need for synchronization barriers, and requires only
one kernel launch from Java as described on Listing~\ref{ch17:fig:mior_launcher}.

\lstinputlisting[caption=MIOR simulation launcher,label=ch17:fig:mior_launcher]{Chapters/chapter17/code/mior_launcher.java}

\subsubsection{Termination detection}

The termination of a MIOR simulation is reached when the model
stabilizes and no more $CO_{2}$ is produced. This termination
detection can be done either on the CPU or the GPU but it requires a
global view on the system execution.

In the first case, when the CPU controls the GPU simulation process,
the detection is done in two steps: (i) the CPU starts the execution
of a simulation step on the GPU, (ii) the CPU retrieves the GPU data
and determines if another iteration must be launched or if the
simulation has terminated. This approach allows a fine-grain control
over the GPU execution, but it requires many costly data transfers as
each iteration results must be sent from the GPU to the CPU. In the
case of the MIOR model these costs are mainly due to the inherent
PCI-express port latencies rather than to bandwidth limitation since
data sizes remains rather small, in the order of few dozens of
Megabytes.

In the second case the termination detection is directly implemented
on the GPU by checking the amount of available carbon between two
iterations. The CPU does not have any feedback while the simulation is
running, but retrieves the results once the kernel execution is
finished. This approach minimizes the number of transfers between the
CPU and the GPU.

\subsection{Performance of MIOR implementations}
\label{ch17:subsec:miorexperiments}

In this part we present several MIOR GPU implementations using the
distribution/gathering process described in the previous section and
compare their performance on two distinct hardware platform; i.e two
different GPU devices. Five incremental MIOR implementations were
realized with an increasing level of adaptation for the algorithm: in
all cases, we choose the average time over 50 executions as a
performance indicator.

\begin{itemize}
\item the \textbf{GPU 1.0} implementation is a direct implementation
  of the existing algorithm and its data structures where data
  dependencies were removed and that use the non-compact topology
  representation described in Section~\ref{ch17:subsec:datastructures}
\item the \textbf{GPU 2.0} implementation uses the previously
  described compact representation of the topology and remains
  otherwise identical to the GPU 1.0 implementation.
\item the \textbf{GPU 3.0} implementation introduces the manual
  copy into local (private) memory of often-accessed global data, such
  as carbon parts or topology information.
\item the \textbf{GPU 4.0} implementation is a variant of the
  GPU 1.0 implementation which allows the execution of
  multiple simulations for each kernel execution.
\item the \textbf{GPU 5.0} implementation is a multi-simulation of the
  GPU 2.0 implementation, using the execution of multiple simulations
  for each kernel execution as for GPU 4.0
\end{itemize}

The two last implementations -- \textbf{GPU 4.0} and \textbf{GPU 5.0}
-- illustrate the gain provided by a better usage of the hardware
resources, thanks to the driver execution overlapping capabilities. A
sequential version of the MIOR algorithm, labeled as \textbf{CPU}, is
provided for comparison purpose. This sequential version is developed
in Java, the same language used for GPU implementations and the Sworm
model.

For these performance evaluations, two platforms are used. The first
one is representative of the kind of hardware which is available on
HPC clusters. It is a cluster node dedicated to GPU computations with
two Intel X5550 processors running at $2.67GHz$ and two Tesla C1060
GPU devices running at $1.3$GHz and composed of $240$ cores ($30$
multi-processors). Only one of these GPU is used in these experiments,
at the moment. The second platform illustrates what can be expected
from a personal desktop computer built a few years ago. It uses a
Intel Q9300 CPU, running at $2.5$GHz, and a Geforce 8800GT GPU card
running at $1.5$GHz and composed of $112$ cores ($14$
multi-processors). The purpose of these two platforms is to assess the
benefit that could be obtained either when a scientist has access to
specialized hardware as a cluster or tries to take benefit from its
own personal computer.

\begin{figure}[h]
%begin{minipage}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.7\linewidth]{Chapters/chapter17/figs/mior_perfs_tesla.pdf}
%\caption{Performance CPU and GPU sur carte Tesla C1060}
\caption{CPU and GPU performance on a Tesla C1060 node}
\label{ch17:fig:mior_perfs_tesla}
%end{minipage}
\end{figure}

Figures~\ref{ch17:fig:mior_perfs_tesla}~and~\ref{ch17:fig:mior_perfs_8800gt}
show the execution time for $50$ simulations on the two hardware
platforms. A size factor is applied to the problem: at scale 1,
the model contains $38$ MM and $310$ OM, while at the scale 6 these
numbers are multiplied by six. The size of the environment is modified
as well to maintain the same average agent density in the model. This
scaling factor display the impact of the chosen size of simulation on
performance.

%hspace{0.02\linewidth}
%begin{minipage}[t]{0.49\linewidth}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\linewidth]{Chapters/chapter17/figs/mior_perfs_8800gt.pdf}
\caption{CPU and GPU performance on a personal computer with a Geforce 8800GT}
\label{ch17:fig:mior_perfs_8800gt}
%end{minipage}
\end{figure}

The charts show that for small problems the execution times of all
the implementations are very close. This is because the GPU execution
does not have enough threads (representing agents) for an optimal
usage of GPU resources. This trend changes after scale $5$ where GPU
2.0 and GPU 3.0 take the advantage on the GPU 1.0 and CPU
implementations. This advantage continues to grow with the scaling
factor, and reaches a speedup of $10$ at the scale $10$ between the
fastest single-simulation GPU implementation and the first, naive one
GPU 1.0.

Multiple trends can be observed in these results. First, optimizations
for the GPU hardware show a big, positive impact on performance,
illustrating the strong requirements on the algorithm properties to
reach execution efficiency. These charts also show that despite the
vast difference in numbers of cores between the two GPUs platforms the
same trends can be observed in both case. We can therefore expect
similar results on other GPU cards, without the need for more
adaptations.

\begin{figure}[h]
\centering
\includegraphics[width=0.7\linewidth]{Chapters/chapter17/figs/monokernel.pdf}
\caption{Execution time of one multi-simulation kernel on the Tesla
  platform}
\label{ch17:fig:monokernel_graph}
\end{figure}

\begin{figure}[h]
\centering
\includegraphics[width=0.7\linewidth]{Chapters/chapter17/figs/multikernel.pdf}
\caption{Total execution time for 1000 simulations on the Tesla
  platform, while varying the number of simulations for each kernel}
\label{ch17:fig:multikernel_graph}
\end{figure}

There are two ways to measure simulations performance: (i) by
executing only one kernel, and varying its size (the number of
simulations executed in parallel), as shown on
Figure~\ref{ch17:fig:monokernel_graph}, to test the costs linked to the
parallelization process or (ii) by executing a fixed number of
simulations and varying the size
of each kernel, as shown on Figure~\ref{ch17:fig:multikernel_graph}.

Figure~\ref{ch17:fig:monokernel_graph} illustrates the execution time for
only one kernel. It shows that for small numbers of simulations run in
parallel, the compact implementation of the model topology is faster
than the two-dimension matrix representation. This trends reverse with
more than $50$ simulations in parallel, which can be explained either by
the non linear progression of the synchronization costs or by the
additional memory required for the access-efficient representation.

Figure~\ref{ch17:fig:multikernel_graph} illustrates the execution time of a
fixed number of simulations. It shows that for a small number of
simulations run in parallel, the costs resulting of program setup,
data copies and launch on GPU are determinant and very detrimental to
performance. Once the number of simulations executed for each kernel
grows, these costs are counterbalanced by computation costs. This
trend is more marked in the case of the sparse implementation (GPU
4.0) than the compact one but appears on both curves. With more than
$30$ simulations for each kernel, execution times stalls, since
hardware limits are reached. This indicates that the cost of preparing
and launching kernels become negligible compared to the computing
time once a good GPU occupation rate is achieved.

\section{Analysis and recommendations}
\label{ch17:analysis}

In this section we synthesize the observations done on the two models
and identify some recommendations on implementing complex systems on
GPU platforms.

\subsection{Analysis}

In both the Collembola and the MIOR model, a critical problematic is
the determination of the part of the simulation which are to be run on
GPU and which are to remain on the CPU. The determination of these
parts is a classic step of any algorithm parallelization and must
take into account considerations such as the cost of the different
parts of the algorithm and the expected gains.

In the case of the Collembola model two steps of the algorithm were
ported to GPU. Both steps use straightforward, easily-parallelizable,
operations where a direct gain can be expected by using more
executions cores without important modifications of the algorithm.

In the MIOR model case however no such inherently parallelizable parts
are evident in the original sequential algorithm. This is mainly
explained by the rate of interactions between agents in this model in
the form of many operations (breathing, growth) on shared carbon
resources. In this case the algorithm had to be more profoundly
modified while keeping in head the need to remain true to the original
model, to synchronize the main execution step of all agents in the
model, to ensure equity, and to minimize the numbers of
synchronizations. The minimization is done by factoring the repartition of
carbon in the model in two separated step at the beginning and the end
of each iterations rather than at multiple point of the execution.

\subsection{MAS execution workflow}

Many MAS simulations decompose their execution process into discrete
evolution steps where each step represents a quantum of time (minimal
unit of time described). At the end of each step many global
information, graphical displays or results are updated. This
execution model may not correctly fit on GPU platforms as they assume
more or less a batch-like workflow model. The execution model must be
split into the following ever repeating steps:

\begin{itemize}
\item Allocation of GPU data buffers
\item Copy of data from CPU to GPU
\item GPU kernels execution
\item Copy of results from GPU to CPU
\end{itemize}

This workflow works well if the considered data transfer time is
negligible compared to GPU execution or can be done in parallel,
thanks to the asynchronous nature of OpenCL. If we are to update the
MAS model after each iteration then performance risk being
degraded. This is illustrated in the MIOR model by the fact that the
speedup observed on GPU is much more significant for bigger
simulations, which implies longer GPU execution times. Our solution to
this problem is to desynchronize the execution of the SMA model and
its GPU parts by requesting the execution of multiple steps of the
GPU simulations for each launch.

Another related prerequisite of GPU execution is the ability to have
many threads to be executed, to allow an efficient exploitation of the
superior number of cores provided by the architecture. In the case of
MAS models, this means that executing one agent at a time on GPU is
meaningless in regard to GPU usage, copying cost, and actual gain in
execution time, if the process is not already parallel and costly at
this scale. In the MIOR and the Collembola models, this is solved by
executing the computations for all agents of the model at the same
time. If the model has however only chronic needs for intensive
computations then some kind of batching mechanism is required to store
waiting treatments in a queue, until the total sum of waiting
computations justify the transfers cost to the GPU platform.

\subsection{Implementation challenges}

Besides the execution strategy challenges described above, some
implementation challenges also occur when implementing an OpenCL
version of a MAS model, mainly related to the underlying limitations
of the execution platform.

The first one is the impossibility (except in latest CUDA versions)
to dynamically allocate memory during execution. This is a problem in
the case of models where the number of agent can vary during the
simulation, such as prey-predator models. In this case, the only
solution is to overestimate the size of arrays or data structures to
accommodate these additional individuals, or using the CPU to resize
data structures when these situations occur. Both approaches require
to trend either memory or performance and are not always practical.

Another limitation is the impossibility to store pointers in data
structures, we restraint any OpenCL to use only one dimension static
arrays or specific data structures such as textures. This preclude the
usage of structures like linked-list, graphs or sparse matrix not
represented by some combination of static arrays, and can be another
source of memory or performance losses.

In the case of MIOR, this problem is especially exacerbated in the case
of neighboring storage: both representations consume much more memory
that actually required, since the worst case (all agents have
access to all others agents) has to be taken into account defensively
when dimensioning the data structure.

The existence of different generations of GPU hardwares is also a
challenge. Older implementations, such as the four year old Tesla
C1060 cards, have very strong expectation in term of memory accesses
and requires very regular access patterns to perform efficiently. MAS
having many random accesses, such as MIOR, or many small global memory
accesses, such as Collembola, are penalized on these older
cards. Fortunately, these requirements are less present is modern
cards, which offer caches and other facilities traditionally present
on CPU to offset these kind of penalties.

The final concern is related to the previous ones and often result in
more memory consumption. The amount of memory available on GPU cards
is much more limited and adding new memory capabilities is more costly
compared to expending a CPU RAM. On computing clusters, hardwares
nodes with 128GB of memory or more have become affordable, whereas
newer Tesla architecture remains limited to 16GB of memory. This can
be a concern in the case of big MAS models, or small ones which can
only use memory-inefficient OpenCL structures.

\subsection{MCSMA}
\label{ch17:Mcsma}

As shown in the previous sections, many data representations choices
are common to entire classes of MAS. The paradigm of grid, for example,
is often encountered in models where each cell constitutes either the
elementary unit of simulation (SugarScape~\cite{Dsouza2007}) or a
discretization of the environment space
(Pathfinding~\cite{Guy09clearpath}). These grids can be considered as
two or three-dimensional matrices, whose processing can be directly
distributed.

Another common data representation encountered in MAS system is the
usage of 2d or 3d coordinates to store the position of each agent of
the model. In this case, even if the environment is no longer
discrete, the location information still imply computations
(euclidean distances) which can be parallelized.

MCSMA~\cite{lmlm+13:ip} is a framework developed to provide to the MAS designer
those basic data structures, and the associated operations, to
facilitate the portage of existing MAS on GPU. Two levels of
utilization are provided to the developer, depending on its usage
profile.

\begin{itemize}
\item A high-level library, constituted of modules regrouping classes
  of operations. A module provides multiple methods of
  distance computations in 1d, 2d or 3d grids, another one the
  diffusion algorithm...
\item A low level API which allows the developed direct access to the
  GPU and the inner working of MCSMA, to develop its own module in
  the case where the required operations are not yet provided by the platform.
\end{itemize}

Both usage levels were illustrated in the above two practical cases. In
MIOR, the whole algorithm (baring initialization) is ported on GPU as
a specific plugin which allows executing $n$ MIOR simulations and
retrieve their results. This first approach requires extensive
adaptations to the original algorithm.  In Collembola, on the
contrary, the main steps of the algorithm remain executed on the CPU,
and only specific operations are delegated to generic, already
existing, diffusion and reduction kernels. The fundamental algorithm
is not modified and GPU execution is only applied to specific parts of
the execution which may benefit from it. These two programming
approaches allow incremental adaptations of existing Java MAS to
accelerate their execution on GPU, while retaining the option to
develop their own reusable or more efficient module to supplement the
already existing ones.

\section{Conclusion}
\label{ch17:conclusion}

This chapter addresses the issue of complex system simulation by using
agent based paradigm and GPU hardware. From the experiments on two
existing Agent Based Models of soil science we intend to provide
useful information on the architecture, the algorithm design and, the
data management to run Agent based simulations on GPU, and more
generally to run compute intensive applications that are not based on
a not-so-regular model. The first result of this work is that adapting
the algorithm to a GPU architecture is possible and suitable to speed
up agent based simulations as illustrated by the MIOR model. Coupling
CPU with GPU seems even to be an interesting way to better take
advantage of the power given by computers and clusters as illustrated
by the Collembola model. From our point of view the adaptation process
is lighter than a traditional parallelization on distributed nodes and
not much difficult than a standard multi-threaded parallelization,
since all the data remains on the same host and can be shared in
central memory. The OCL adaptation also enables a portable simulator
that can be run on different graphical units. Even using a mainstream
card as the GPU card of a standard computer can lead to significant
performance improvements. This is an interesting result as it opens
the field of inexpensive HPC to a large community.

In this perspective, we are working on MCSMA a development platform
that would facilitate the use of GPU or many-core architectures for
multi-agent simulations. Our first work has been the definition of
common, efficient, reusable data structures, such as grids or
lists. Another goal is to provide easier means to control the
distribution of specific processes between CPU or GPU, to allow the
easy exploitation of the strengths of each platform in the same
multi-agent simulation. We think that the same approach, i.e
developing specific environments that facilitate the developer
access to the GPU power, can be applied in many domains with
compute intensive needs to open the GPU use to a larger community. 

\putbib[Chapters/chapter17/biblio]