[book_gpu.git] / BookGPU / Chapters / chapter18 / ch18.tex

\chapterauthor{Raphaël Couturier and Christophe Guyeux}{Femto-ST Institute, University of Franche-Comte, France}
%\chapterauthor{Christophe Guyeux}{Femto-ST Institute, University of Franche-Comt\'{e}}


\chapter{Pseudorandom number generator on GPU}
\label{chapter18}

\section{Introduction}

Randomness is of importance in many fields such as scientific
simulations or cryptography.  Random numbers can mainly be
generated by either a deterministic and reproducible algorithm called
a pseudorandom number generator (PRNG)\index{PRNG}, or by a physical nondeterministic 
process having all the characteristics of a random noise, called a truly random number
generator (TRNG). In this chapter, we focus on
reproducible generators, useful for instance in MonteCarlo-based
simulators.  These domains need PRNGs that are statistically
irreproachable.  In some fields such as in numerical simulations,
speed is a strong requirement that is usually attained by using
parallel architectures. In that case, a recurrent problem is that a
deflation of the statistical qualities is often reported, when the
parallelization of a good PRNG is realized.  This
is why adhoc PRNGs for each possible architecture must be found to
achieve both speed and randomness.  On the other hand, speed is not
the main requirement in cryptography: the most important point is to
define \emph{secure} generators able to withstand malicious
attacks. Roughly speaking, an attacker should not be able in practice
to make the distinction between numbers obtained with the secure
generator and a true random sequence.  Or, in an equivalent
formulation, he or she should not be able (in practice) to predict the
next bit of the generator, having the knowledge of all the binary
digits that have been already released~\cite{Goldreich}. ``Being able in practice''
refers here to the possibility to achieve this attack in polynomial
time and to the exponential growth of the difficulty of this
challenge when the size of the parameters of the PRNG increases.


Finally, a small part of the community working in this domain focuses on a
third requirement: to define chaotic generators~\cite{kellert1994wake,  Wu20051195,gleick2011chaos}.
The main idea is to  benefit from a chaotic dynamical system to obtain a
generator that is unpredictable, disordered, sensible to its seed, or in other words, chaotic.
These scientists' desire is to map a given chaotic dynamics into a sequence that seems random 
and unassailable due to chaos.
However, the chaotic maps used as a pattern are defined in the real line 
whereas computers deal with finite precision numbers.
This distortion leads to a deflation of both chaotic properties and speed.
Furthermore, authors of such chaotic generators often claim their PRNG
as secure due to their chaos properties, but there is no obvious relation
between chaos and security as it is understood in cryptography.
This is why the use of chaos for PRNG still remains marginal and disputable.
However, we have established in previous researches that these flaws can 
be circumvented by using a tool called choatic iterations. 
Such investigations have led to the definition of a new
family of PRNGs that are chaotic while being fast and statistically perfect,
or cryptographically secure~\cite{bgw09:ip,bgw10:ip,bfgw11:ij,bfg12a:ip}. This family is improved and adapted for GPU 
architectures in this chapter.


Let us finish this introduction by noticing that, in this chapter, 
statistical perfection refers to the ability to pass the whole 
{\it BigCrush} battery of tests, which is widely considered as the most
stringent statistical evaluation of a sequence claimed as random.
This battery can be found in the well-known TestU01 package~\cite{LEcuyerS07}\index{TestU01}.
More precisely, each time we performed a test on a PRNG, we ran it
twice in order to observe if all $p$-values were inside [0.01, 0.99]. In
fact, we observed that few $p$-values (fewer than 10 out of 160) are sometimes
outside this interval but inside [0.001, 0.999], so that is why a
second run has allowed us to confirm that the values outside are not for
the same test. With this approach all our PRNGs pass the {\it
  BigCrush} successfully and all $p$-values are at least once inside
[0.01, 0.99].
Chaos, for its part, refers to the well-established definition of a
chaotic dynamical system defined by Devaney~\cite{Devaney}.

The remainder of this chapter is organized as follows. 
Basic definitions and terminologies about both topological chaos
and chaotic iterations are provided in the next section. Various 
chaotic iterations based on pseudorandom number generators are then
presented in Section~\ref{sec:efficient PRNG}. They encompass
naive and improved efficient generators for CPU and for GPU. 
These generators are finally experimented in Section~\ref{sec:experiments}.


\section{Basic remindees}
\label{section:BASIC RECALLS}

This section is devoted to basic definitions and terminologies in the fields of
topological chaos and chaotic iterations. We assume the reader is familiar
with basic notions on topology (see for instance~\cite{Devaney}).


\subsection{A short presentation of chaos}


Chaos theory studies the behavior of dynamical systems that are perfectly predictable, yet appear to be wildly amorphous and meaningless. 
Chaotic systems\index{chaotic!systems} are highly sensitive to initial conditions, 
which is popularly referred to as the butterfly effect. 
In other words, small differences in initial conditions (such as those due to rounding errors in numerical computation) yield widely diverging outcomes, 
in general rendering long-term prediction impossible  \cite{kellert1994wake}. This happens even though these systems are deterministic, meaning that their future behavior is fully determined by their initial conditions, with no random elements involved \cite{kellert1994wake}. That is, the deterministic nature of these systems does not make them predictable \cite{kellert1994wake,Werndl01032009}. This behavior is known as deterministic chaos, or simply chaos. It has been well-studied in mathematics and
physics, leading among other things to the well-established definition of Devaney which can be found next.





\subsection{On Devaney's definition of chaos}\index{chaos}
\label{sec:dev}
Consider a metric space $(\mathcal{X},d)$ and a continuous function $f:\mathcal{X}\longrightarrow \mathcal{X}$, for one-dimensional dynamical systems of the form:
\begin{equation}
x^0 \in \mathcal{X} \textrm{  and } \forall n \in \mathds{N}^*, x^n=f(x^{n-1}),
\label{Devaney}
\end{equation}
the following definition of chaotic behavior, formulated by Devaney~\cite{Devaney}, is widely accepted.

\begin{definition}
 A dynamical system of Form~(\ref{Devaney}) is said to be chaotic if the following conditions hold.
\begin{itemize}
\item Topological transitivity\index{topological transitivity}:

\begin{equation}
\forall U,V \textrm{ open sets of } \mathcal{X},~\exists k>0, f^k(U) \cap V \neq \varnothing .
\end{equation}

Intuitively, a topologically transitive map has points that eventually move under iteration from one arbitrarily small neighborhood to any other. Consequently, the dynamical system cannot be decomposed into two disjoint open sets that are invariant under the map. Note that if a map possesses a dense orbit, then it is clearly topologically transitive.
\item Density of periodic points in $\mathcal{X}$\index{density of periodic points}:

Let $P=\{p\in \mathcal{X}|\exists n \in \mathds{N}^{\ast}:f^n(p)=p\}$ the set of periodic points of $f$. Then $P$ is dense in $\mathcal{X}$:

\begin{equation}
 \overline{P}=\mathcal{X} .
\end{equation}

The density of periodic orbits means that every point in  space is closely approached  by periodic orbits in an arbitrary way. Topologically mixing systems failing this condition may not display sensitivity to initial conditions presented below and, hence,may not be chaotic.
\item Sensitive dependence on initial conditions\index{sensitive dependence on initial conditions}:

$\exists \varepsilon>0,$ $\forall x \in \mathcal{X},$ $\forall \delta >0,$ $\exists y \in \mathcal{X},$ $\exists n \in \mathbb{N},$ $d(x,y)<\delta$ and $d\left(f^n(x),f^n(y)\right) \geqslant \varepsilon.$

Intuitively, a map possesses sensitive dependence on initial conditions if there exist points arbitrarily close to $x$ that eventually separate from $x$ by at least $\varepsilon$ under the iteration of $f$. Not all points near $x$ need to be eventually separate from $x$ under iteration, but there must be at least one such point in every neighborhood of $x$. If a map possesses sensitive dependence on initial conditions, then for all practical purposes, the dynamics of the map defy numerical computation. Small errors in computation that are introduced by round-off may become magnified upon iteration. The results of numerical computation of an orbit, no matter how accurate, may bear no resemblance whatsoever with the real orbit.
\end{itemize}

\end{definition}

When $f$ is chaotic, then the system $(\mathcal{X}, f)$ is chaotic and quoting Devaney~\cite[p. 50]{Devaney}: ``it is unpredictable because of the sensitive dependence on initial conditions. It cannot be broken down or decomposed into two subsystems which do not interact because of topological transitivity. And, in the midst of this random behavior, we nevertheless have an element of regularity.'' Fundamentally different behaviors are consequently possible and occur in an unpredictable way.








\subsection{Chaotic iterations}\index{chaotic!iterations}
\label{subsection:Chaotic iterations}

Let us now introduce an example of a dynamical systems family that has
the potentiality to become chaotic, depending on the choice of the iteration 
function. This family is the basis of the PRNGs we will
develop during this chapter.

\begin{definition}
\label{Chaotic iterations}
The set $\mathds{B}$ denoting $\{0,1\}$, let $f:\mathds{B}^{\mathsf{N}%
}\longrightarrow \mathds{B}^{\mathsf{N}}$ be an ``iteration'' function and $S$ be a 
sequence of subsets of $\llbracket 1, \mathsf{N}\rrbracket$ called a chaotic strategy. Then, the so-called \emph{chaotic iterations} are defined by~\cite{Robert1986}:

\begin{equation}
\label{eq:generalIC}
\left\{\begin{array}{l}
x^0\in \mathds{B}^{\mathsf{N}}, \\
\forall n\in \mathds{N}^{\ast },\forall i\in \llbracket1;\mathsf{N}\rrbracket%
,x_i^n=
\left\{
\begin{array}{ll}
x_i^{n-1} & \text{if}~ i\notin S^n  \\
f(x^{n-1})_{i}  & \text{if}~ i \in S^n.
\end{array} 
\right. 
\end{array}
\right.
\end{equation}
\end{definition}

In other words, at the $nth$ iteration, only the  cells of $S^{n}$ are
iterated. 
Chaotic iterations generate a set of vectors;
they are defined by an initial state $x^{0}$, an iteration function $f$, and a chaotic strategy $S$~\cite{bg10:ij}.
These ``chaotic iterations'' can behave chaotically as defined by Devaney, 
depending on the choice of $f$~\cite{bg10:ij}. For instance,
chaos is obtained when $f$ is the vectorial negation.
Note that, with this example of function, chaotic iterations
defined above can be rewritten as
\begin{equation}
\label{equation Oplus}
x^0 \in \llbracket 0,2^\mathsf{N}-1\rrbracket,~\mathcal{S}^n \in \mathcal{P}\left(\llbracket 1,2^\mathsf{N}-1\rrbracket\right)^\mathds{N},~~ x^{n+1}=x^n \oplus \mathcal{S}^n,
\end{equation}
where $\mathcal{P}(X)$ stands for the set of subsets of $X$, whereas 
$a\oplus b$ is the bitwise exclusive or operation between the binary
representation of the integers $a$ and $b$. Note that the term
 $\mathcal{S}^n$ is directly and obviously linked to the $S^n$ of
Eq.~\ref{eq:generalIC}. Such an iterative sequence 
satisfies the Devaney's definition of chaos.

\section{Toward efficiency and improvement for CI PRNG}
\label{sec:efficient PRNG}

\subsection{First efficient implementation of a PRNG based on chaotic iterations}
%
%Based on the proof presented in the previous section, it is now possible to 
%improve the speed of the generator formerly presented in~\cite{bgw09:ip,guyeux10}. 
%The first idea is to consider
%that the provided strategy is a pseudorandom Boolean vector obtained by a
%given PRNG.
%An iteration of the system is simply the bitwise exclusive or between
%the last computed state and the current strategy.
%Topological properties of disorder exhibited by chaotic 
%iterations can be inherited by the inputted generator, we hope by doing so to 
%obtain some statistical improvements while preserving speed.
%
%%RAPH : j'ai viré tout ca
%% Let us give an example using 16-bits numbers, to clearly understand how the bitwise xor operations
%% are
%% done.  
%% Suppose  that $x$ and the  strategy $S^i$ are given as
%% binary vectors.
%% Table~\ref{TableExemple} shows the result of $x \oplus S^i$.

%% \begin{table}
%% \begin{scriptsize}
%% $$
%% \begin{array}{|cc|cccccccccccccccc|}
%% \hline
%% x      &=&1&0&1&1&1&0&1&0&1&0&0&1&0&0&1&0\\
%% \hline
%% S^i      &=&0&1&1&0&0&1&1&0&1&1&1&0&0&1&1&1\\
%% \hline
%% x \oplus S^i&=&1&1&0&1&1&1&0&0&0&1&1&1&0&1&0&1\\
%% \hline

%% \hline
%%  \end{array}
%% $$
%% \end{scriptsize}
%% \caption{Example of an arbitrary round of the proposed generator}
%% \label{TableExemple}
%% \end{table}




\lstinputlisting[label=algo:seqCIPRNG,caption={C code of the sequential PRNG based on chaotic iterations}]{Chapters/chapter18/code2.cu}


In Listing~\ref{algo:seqCIPRNG} a sequential  version of the proposed PRNG based
on  chaotic  iterations  is  presented, which extends the generator family 
formerly presented in~\cite{bgw09:ip,guyeux10}.   The xor  operator  is  represented  by
\textasciicircum.  This function uses  three classical 64-bit PRNGs, namely the
\texttt{xorshift},         the          \texttt{xor128},         and         the
\texttt{xorwow}~\cite{Marsaglia2003}.  In the following, we call them ``xor-like
PRNGs''.   As each  xor-like PRNG  uses 64-bits  whereas our  proposed generator
works with 32-bits, we use the command \texttt{(unsigned int)}, which selects the
32 least  significant bits  of a given  integer, and the  code \texttt{(unsigned
  int)(t$>>$32)} in order to obtain the 32 most significant bits of \texttt{t}.

Thus producing a pseudorandom number needs 6 xor operations with 6 32-bit numbers
that  are provided by  3 64-bit  PRNGs.  This  version successfully  passes the
stringent {\it BigCrush} battery of tests~\cite{LEcuyerS07}. 
At this point, we have defined an efficient and statistically unbiased generator. Its speed is
directly related to the use of linear operations, but for the same reason,
this fast generator cannot be proven as secure.



\subsection{Efficient PRNGs based on chaotic iterations on GPU}
\label{sec:efficient PRNG gpu}

In order to  benefit from the computing power  of GPU, a program
needs  to have  independent blocks  of  threads that  can be  computed
simultaneously. In general,  the larger the number of  threads is, the
more local  memory is  used, and the  less branching  instructions are
used  (if,  while,  etc.) and so,  the  better the  performances  on  GPU  are.
Obviously, having these requirements in  mind, it is possible to build
a   program    similar   to    the   one   presented    in  Listing~\ref{algo:seqCIPRNG}, which computes  pseudorandom numbers on GPU.  To
do  so,  we  must   first  recall  that  in  the  CUDA~\cite{Nvid10}
environment,    threads    have     a    local    identifier    called
\texttt{ThreadIdx},  which   is  relative  to   the  block  containing
them. Furthermore, in  CUDA, parts of  the code that are executed by the  GPU are
called {\it kernels}.


\subsection{Naive version for GPU}

 
It is possible to deduce from the CPU version a quite similar version adapted to GPU.
The simple principle consists of making each thread of the GPU compute the CPU version of our PRNG.  
Of course,  the  three xor-like
PRNGs  used in these computations must have different  parameters. 
In a given thread, these parameters are
randomly picked from another PRNGs. 
The  initialization stage is performed by  the CPU.
To do this, the   Indirection, Shift, Accumulate, Add, and Count (ISAAC)  PRNG~\cite{Jenkins96} is used to  set  all  the
parameters embedded into each thread.   

The implementation of  the three
xor-like  PRNGs  is  straightforward  when  their  parameters  have  been
allocated in  the GPU memory.  Each xor-like  works with  an internal
number  $x$  that saves  the  last  generated  pseudorandom number. Additionally,  the
implementation of the  {\it xor128}, the {\it xorshift}, and the  {\it xorwow}, respectively, require
4, 5, and 6 unsigned long as internal variables.


\begin{algorithm}
\begin{small}
\KwIn{InternalVarXorLikeArray: array with internal variables of the 3 xor-like
PRNGs in global memory\;
NumThreads: number of threads\;}
\KwOut{NewNb: array containing random numbers in global memory}
\If{threadIdx is concerned by the computation} {
  retrieve data from InternalVarXorLikeArray[threadIdx] in local variables\;
  \For{i=1 to n} {
    compute a new PRNG as in Listing~\ref{algo:seqCIPRNG}\;
    store the new PRNG in NewNb[NumThreads*threadIdx+i]\;
  }
  store internal variables in InternalVarXorLikeArray[threadIdx]\;
}
\end{small}
\caption{main kernel of the GPU ``naive'' version of the PRNG based on chaotic iterations}
\label{algo:gpu_kernel}
\end{algorithm}



Algorithm~\ref{algo:gpu_kernel}  presents a naive  implementation of the proposed  PRNG on
GPU.  Due to the available  memory in the  GPU and the number  of threads
used simultaneously,  the number  of random numbers  that a thread  can generate
inside   a    kernel   is   limited  (i.e.,    the    variable   {\it n}   in
Algorithm~\ref{algo:gpu_kernel}). For instance, if  $100,000$ threads are used and
if $n=100$\footnote{In fact, we need to add the initial seed (a 32-bit number).},
then   the  memory   required   to  store all of the  internal   variables  of both the  xor-like
PRNGs\footnote{We multiply this number by $2$ in order to count 32-bit numbers.}
and  the pseudorandom  numbers generated by  our  PRNG,  is  equal to  $100,000\times  ((4+5+6)\times
2+(1+100))=1,310,000$ 32-bit numbers, that is, approximately $52$Mb.

This generator is able to pass the whole {\it BigCrush} battery of tests, for all
the versions that have been tested depend on their number of threads 
(called NumThreads in our algorithm, tested up to $5$ million).


\subsection{Improved version for GPU}

As GPU cards using CUDA have shared memory between threads of the same block, it
is possible  to use this  feature in order  to simplify the  previous algorithm,
i.e., to use fewer  than 3 xor-like PRNGs. The solution  consists in computing only
one xor-like PRNG by thread, saving  it into the shared memory, and then  using the results
of some  other threads in the  same block of  threads. In order to  define which
thread uses the result of which other  one, we can use a combination array that
contains  the indexes  of  all threads  and  for which  a combination has  been
performed. 

In  Algorithm~\ref{algo:gpu_kernel2},  two  combination  arrays are  used.   The
variable     offset    is     computed    using     the     value    of
combination\_size.   Then we  can compute  o1  and o2
representing the  indexes of  the other  threads whose results  are used  by the
current one.   In this algorithm, we  consider that a 32-bit  xor-like PRNG has
been chosen. In practice, we  use the xor128 proposed in~\cite{Marsaglia2003} in
which  unsigned longs  (64 bits)  have been  replaced by  unsigned  integers (32
bits).

This version  can also pass the whole {\it BigCrush} battery of tests.

\begin{algorithm}
\begin{small}
\KwIn{InternalVarXorLikeArray: array with internal variables of 1 xor-like PRNGs
in global memory\;
NumThreads: Number of threads\;
array\_comb1, array\_comb2: Arrays containing combinations of size combination\_size\;}

\KwOut{NewNb: array containing random numbers in global memory}
\If{threadId is concerned} {
  retrieve data from InternalVarXorLikeArray[threadIdx] in local variables including shared memory and x\;
  offset = threadIdx\%combination\_size\;
  o1 = threadIdx-offset+array\_comb1[offset]\;
  o2 = threadIdx-offset+array\_comb2[offset]\;
  \For{i=1 to n} {
    t=xor-like()\;
    t=t\textasciicircum shmem[o1]\textasciicircum shmem[o2]\;
    shared\_mem[threadId]=t\;
    x = x\textasciicircum t\;

    store the new PRNG in NewNb[NumThreads*threadIdx+i]\;
  }
  store internal variables in InternalVarXorLikeArray[threadIdx]\;
}
\end{small}
\caption{Main kernel for the chaotic iterations based PRNG GPU efficient
version\label{IR}}
\label{algo:gpu_kernel2} 
\end{algorithm}

\subsection{Chaos evaluation of the improved version}

A run of Algorithm~\ref{algo:gpu_kernel2} consists of an operation ($x=x\oplus t$) having 
the form of Equation~\ref{equation Oplus}, which is equivalent to the iterative
system of Eq.~\ref{eq:generalIC}. That is, an iteration of the general chaotic
iterations is realized between the last stored value $x$ of the thread and a strategy $t$
(obtained by a bitwise exclusive or between a value provided by a xor-like() call
and two values previously obtained by two other threads).
To be certain that such iterations correspond to the chaotic one recalled at the
end of  Section~\ref{sec:dev},
we must guarantee that this dynamical system iterates on the space 
$\mathcal{X} =\mathds{B}^\mathsf{N} \times \mathcal{P}\left(\llbracket 1, 2^\mathsf{N} \rrbracket\right)^\mathds{N}$.
The left term $x$ obviously belongs to $\mathds{B}^ \mathsf{N}$.
To prevent any flaws of chaotic properties, we must check that the right 
term (the last $t$ in Algorithm~\ref{algo:gpu_kernel2}), corresponding to the strategies,  can possibly be equal to any
integer of $\llbracket 1, 2^\mathsf{N} \rrbracket$. 

Such a result is obvious; as for the xor-like(), all the
integers belonging to its interval of definition can occur at each iteration, and thus the 
last $t$ respects the requirement. Furthermore, it is possible to
prove by an immediate mathematical induction that, supposing that the initial $x$
is uniformly distributed, %(it is provided by a cryptographically secure PRNG),
the two other stored values shmem[o1] and shmem[o2] are uniformly distributed too
(this is the induction hypothesis), and thus the next $x$ is finally uniformly distributed.

Thus Algorithm~\ref{algo:gpu_kernel2} is a concrete realization of the general
chaotic iterations presented previously, and for this reason, it satisfies the 
Devaney's formulation of  chaotic behavior.

\section{Experiments}
\label{sec:experiments}

Different experiments  have been  performed in order  to measure  the generation
speed. We have used one computer equipped with a Tesla C1060 NVIDIA  GPU card
and an
Intel  Xeon E5530 cadenced  at 2.40  GHz,  and 
a second computer  equipped with a smaller  CPU and  a GeForce GTX  280. 
All the
cards have 240 cores.

\begin{figure}[htpb]
\begin{center}
  \includegraphics[scale=0.65]{Chapters/chapter18/figures/curve_time_xorlike_gpu.pdf}
\end{center}
\caption{Quantity of pseudorandom numbers generated per second with the xorlike-based PRNG.}
\label{fig:time_xorlike_gpu}
\end{figure}

In  Figure~\ref{fig:time_xorlike_gpu} we  compare the  quantity of  pseudorandom numbers
generated per second with various xor-like based PRNGs. In this figure, the optimized
versions use the  xor64 described in~\cite{Marsaglia2003}, whereas the naive versions
embed  the three  xor-like  PRNGs described  in Listing~\ref{algo:seqCIPRNG}.   In
order to obtain the optimal performances, the storage of pseudorandom numbers
to the GPU memory has been removed. This step is time-consuming and slows down the numbers
generation.  Moreover this   storage  is  completely
useless, in case of applications that consume the pseudorandom
numbers  directly   after they have been generated. We can see  that when the number of  threads is greater
than approximately 30,000 and less than 5 million, the number of pseudorandom numbers generated
per second  is almost constant.  With the  naive version, this value ranges from 2.5 to
3GSamples/s.   With  the  optimized   version,  it  is  approximately  equal to
20GSamples/s. Finally  we can remark  that both GPU  cards are quite  similar, but in
practice,  the Tesla C1060  has more  memory than  the GTX  280, and  this memory
should be of better quality.
As a  comparison,   Listing~\ref{algo:seqCIPRNG}  leads   to the  generation of  about
138MSample/s when using one core of the Xeon E5530.








These  experiments allow  us  to conclude  that  it  is possible  to
generate a very large quantity of pseudorandom  numbers statistically perfect with the  xor-like version.


\section{Summary}


In this chapter,  a PRNG based on chaotic iterations is  presented.   It is proven to be
chaotic according to Devaney.   Efficient implementations  on GPU
using xor-like PRNGs  as input generators have shown that  a very large quantity
of pseudorandom  numbers can be  generated per second (about  20Gsamples/s on a Tesla C1060), and
that these proposed PRNGs succeed in passing the hardest battery in TestU01, namely, the {\it BigCrush}.






\putbib[Chapters/chapter18/biblio18]