branching instructions are used (if, while, ...), the better performance is
obtained on GPU. So with algorithm \ref{algo:seqCIprng} presented in the
previous section, it is possible to build a similar program which computes PRNG
-on GPU. In the CUDA [ref] environment, threads have a local identificator,
-called \texttt{ThreadIdx} relative to the block containing them.
+on GPU. In the CUDA~\cite{Nvid10} environment, threads have a local
+identificator, called \texttt{ThreadIdx} relative to the block containing them.
\subsection{Naive version for GPU}
each thread of the GPU. Of course, it is essential that the three xor-like
PRNGs used for our computation have different parameters. So we chose them
randomly with another PRNG. As the initialisation is performed by the CPU, we
-have chosen to use the ISAAC PRNG [ref] to initalize all the parameters for the
-GPU version of our PRNG. The implementation of the three xor-like PRNGs is
-straightforward as soon as their parameters have been allocated in the GPU
-memory. Each xor-like PRNGs used works with an internal number $x$ which keeps
-the last generated random numbers. Other internal variables are also used by the
-xor-like PRNGs. More precisely, the implementation of the xor128, the xorshift
-and the xorwow respectively require 4, 5 and 6 unsigned long as internal
-variables.
+have chosen to use the ISAAC PRNG~\ref{Jenkins96} to initalize all the
+parameters for the GPU version of our PRNG. The implementation of the three
+xor-like PRNGs is straightforward as soon as their parameters have been
+allocated in the GPU memory. Each xor-like PRNGs used works with an internal
+number $x$ which keeps the last generated random numbers. Other internal
+variables are also used by the xor-like PRNGs. More precisely, the
+implementation of the xor128, the xorshift and the xorwow respectively require
+4, 5 and 6 unsigned long as internal variables.
\begin{algorithm}
by the current thread. In the algorithm, we consider that a 64-bits xor-like
PRNG is used, that is why both 32-bits parts are used.
-This version also succeed to the BigCrush batteries of tests.
+This version also succeeds to the {\it BigCrush} batteries of tests.
\begin{algorithm}
\KwOut{NewNb: array containing random numbers in global memory}
\If{threadId is concerned} {
- retrieve data from InternalVarXorLikeArray[threadId] in local variables\;
+ retrieve data from InternalVarXorLikeArray[threadId] in local variables including shared memory\;
offset = threadIdx\%permutation\_size\;
o1 = threadIdx-offset+tab1[offset]\;
o2 = threadIdx-offset+tab2[offset]\;
\For{i=1 to n} {
t=xor-like()\;
- shared\_mem[threadId]=(unsigned int)t\;
- x = x $\oplus$ (unsigned int) t\;
- x = x $\oplus$ (unsigned int) (t>>32)\;
- x = x $\oplus$ shared[o1]\;
- x = x $\oplus$ shared[o2]\;
+ t=t$\oplus$shmem[o1]$\oplus$shmem[o2]\;
+ shared\_mem[threadId]=t\;
+ x = x $\oplus$ t\;
store the new PRNG in NewNb[NumThreads*threadId+i]\;
}
\subsection{Theoretical Evaluation of the Improved Version}
-A run of Algorithm~\ref{algo:gpu_kernel2} consists in four operations having
+A run of Algorithm~\ref{algo:gpu_kernel2} consists in three operations having
the form of Equation~\ref{equation Oplus}, which is equivalent to the iterative
-system of Eq.~\ref{eq:generalIC}. That is, four iterations of the general chaotic
+system of Eq.~\ref{eq:generalIC}. That is, three iterations of the general chaotic
iterations are realized between two stored values of the PRNG.
To be certain that we are in the framework of Theorem~\ref{t:chaos des general},
we must guarantee that this dynamical system iterates on the space
\section{Experiments}
\label{sec:experiments}
-Different experiments have been performed in order to measure the generation
-speed. In Figure~\ref{fig:time_gpu} we compare the number of random numbers generated per second.
+Different experiments have been performed in order to measure the generation
+speed. We have used a computer equiped with Tesla C1060 NVidia GPU card and an
+Intel Xeon E5530 cadenced at 2.40 GHz for our experiments and we have used
+another one equipped with a less performant CPU and a GeForce GTX 280. Both
+cards have 240 cores.
+
+In Figure~\ref{fig:time_gpu} we compare the number of random numbers generated
+per second. The xor-like prng is a xor64 described in~\cite{Marsaglia2003}. In
+order to obtain the optimal performance we remove the storage of random numbers
+in the GPU memory. This step is time consumming and slows down the random number
+generation. Moreover, if you are interested by applications that consome random
+numbers directly when they are generated, their storage is completely
+useless. In this figure we can see that when the number of threads is greater
+than approximately 30,000 upto 5 millions the number of random numbers generated
+per second is almost constant. With the naive version, it is between 2.5 and
+3GSample/s. With the optimized version, it is approximately equals to
+20GSample/s. Finally we can remark that both GPU cards are quite similar. In
+practice, the Tesla C1060 has more memory than the GTX 280 and this memory
+should be of better quality.
\begin{figure}[htbp]
\begin{center}
\end{figure}
-First of all we have compared the time to generate X random numbers with both
-the CPU version and the GPU version.
+In comparison, Listing~\ref{algo:seqCIprng} allows us to generate about
+138MSample/s with only one core of the Xeon E5530.
+
-Faire une courbe du nombre de random en fonction du nombre de threads,
-éventuellement en fonction du nombres de threads par bloc.
