Difference between revisions of "Poisson equation, solving with DFT"
| [unchecked revision] | [unchecked revision] |
| Line 278: | Line 278: | ||
=== Scalability of the algorithm and its implementations === | === Scalability of the algorithm and its implementations === | ||
| − | The variable [[ | + | The variable [[Glossary#Start-up parameters|start-up parameters]] of this implementation and their ranges are as follows: |
* the number of processors is a squared integer in the range [4 : 100]; | * the number of processors is a squared integer in the range [4 : 100]; | ||
* the domain size varies in the range [100x100x100 : 250x100x100] with the step 50x100x100. | * the domain size varies in the range [100x100x100 : 250x100x100] with the step 50x100x100. | ||
| − | The experiments resulted in the following range for the [[ | + | The experiments resulted in the following range for the [[Glossary#Implementation efficiency|implementation efficiency]] of this algorithm: |
* the minimum efficiency is 0,132%; | * the minimum efficiency is 0,132%; | ||
* the maximum efficiency is 1,34%. | * the maximum efficiency is 1,34%. | ||
| − | The following two figures show how the [[ | + | The following two figures show how the [[Glossary#Performance|performance]] and efficiency of this implementation depend on the variable start-up parameters. |
[[file:Poisson perf.png|thumb|center|700px|Figure 8. Parallel implementation of the algorithm. Performance as a function of the number of processors and the domain size.]] | [[file:Poisson perf.png|thumb|center|700px|Figure 8. Parallel implementation of the algorithm. Performance as a function of the number of processors and the domain size.]] | ||
| Line 294: | Line 294: | ||
[[file:Poisson eff.png|thumb|center|700px|Figure 9. Parallel implementation of the algorithm. Efficiency as a function of the number of processors and the domain size. ]] | [[file:Poisson eff.png|thumb|center|700px|Figure 9. Parallel implementation of the algorithm. Efficiency as a function of the number of processors and the domain size. ]] | ||
| − | These results are in good qualitative agreement with the estimates given in the preceding section. For instance, one can see from fig. 8 that the scalability of the algorithm improves with the growth of <math>N</math>; that is, the performance improves faster with the growing number of cores. This is consistent with the fact that <math>\alpha</math> decreases with the growth of <math>N</math>, although the dependence revealed by the expression <math>\alpha=12/(6\log_2 N+1)</math> is considerably more weak. On the whole, the scalability of the algorithm is rather weak. Suppose that the domain has the size 250х100х100. When the number of cores increases by a factor 25 (from 4 to 100), the performance of calculations increases only sixfold. | + | These results are in good qualitative agreement with the estimates given in the preceding section. For instance, one can see from fig. 8 that the scalability of the algorithm improves with the growth of <math>N</math>; that is, the performance improves faster with the growing number of cores. This is consistent with the fact that <math>\alpha</math> decreases with the growth of <math>N</math>, although the dependence revealed by the expression <math>\alpha=12/(6\log_2 N+1)</math> is considerably more weak. On the whole, the scalability of the algorithm is rather weak. Suppose that the domain has the size 250х100х100. When the number of cores increases by a factor 25 (from 4 to 100), the performance of calculations increases only sixfold. |
| − | |||
=== Dynamic characteristics and efficiency of the algorithm implementation === | === Dynamic characteristics and efficiency of the algorithm implementation === | ||
Revision as of 11:35, 3 March 2016
Primary authors of this description:V.M.Stepanenko, E.V.Mortikov, Vad.V.Voevodin (section 2.2)
Contents
- 1 Properties and structure of the algorithm
- 1.1 General description of the algorithm
- 1.2 Mathematical description of the algorithm
- 1.3 Computational kernel of the algorithm
- 1.4 Macro structure of the algorithm
- 1.5 Implementation scheme of the serial algorithm
- 1.6 Serial complexity of the algorithm
- 1.7 Information graph
- 1.8 Parallelization resource of the algorithm
- 1.9 Input and output data of the algorithm
- 1.10 Properties of the algorithm
- 2 Software implementation of the algorithm
- 2.1 Implementation peculiarities of the serial algorithm
- 2.2 Locality of data and computations
- 2.3 Possible methods and considerations for parallel implementation of the algorithm
- 2.4 Scalability of the algorithm and its implementations
- 2.5 Dynamic characteristics and efficiency of the algorithm implementation
- 2.6 Conclusions for different classes of computer architecture
- 2.7 Existing implementations of the algorithm
- 3 References
1 Properties and structure of the algorithm
1.1 General description of the algorithm
The Poisson equation for the multidimensional space has the form [math] \sum_{i=1}^{N}\frac{\partial^2 \phi}{\partial x_i^2}=f,~\mathbf{x}\in D. [/math]
Here, [math]D \in \mathbb{R}^N[/math] is the domain in which the solution [math]\phi(\mathbf{x})[/math] is defined, and [math]\mathbf{x}=(x_1,...,x_N)^T[/math] is the vector of independent variables. The Poisson equation is supplemented by the boundary conditions [math] B(\phi)=F, \mathbf{x} \in \Gamma(D), [/math] where [math]\Gamma(D)[/math] is the boundary of [math]D[/math] and [math]B(\phi)[/math] is the operator defining the boundary conditions. The case [math]B(\phi)=\phi[/math] corresponds to the Dirichlet boundary condition, while [math]B(\phi)=\partial\phi/\partial n[/math], where [math]\mathbf{n}[/math] is the outer normal to the boundary [math]\Gamma(D)[/math], corresponds to the Neumann boundary condition. Sometimes mixed boundary conditions [math]B(\phi)=C\phi+\partial\phi/\partial n[/math], where [math]C[/math] is a constant, are also used. The so-called "periodic boundary conditions" may also occur. In this case, the problem is posed on an unbounded domain, but the solution is assumed to be periodic with respect to a subset of variables from [math]\mathbf{x}[/math].
The Poisson equation emerges in many problems of mathematical physics, for instance, in electrostatics (in this case, [math]\phi[/math] is the potential of the electric force) and hydrodynamics ([math]\phi[/math] is the pressure of a fluid or a gas). The parameter [math]N[/math] is 2 and 3 for the plane and three-dimensional problems, respectively.
The analytical form of the solution to the Poisson equation is not known in the case where the right-hand side is arbitrary and the boundary conditions are inhomogeneous. Consequently, in most applications, this equation is solved numerically. The most common discretization of the Poisson equation has the form
[math] \sum_{i=1}^{N}\frac{\phi_{k_1,...,k_i+1,...,k_N}-2\phi_{k_1,...,k_i,...,k_N}+\phi_{k_1,...,k_i-1,...,k_N}}{\Delta x_i^2}=f_{k_1,...,k_N},~(k_1,...,k_N) \in D_N. [/math]
Here, the second derivatives are replaced by second-order finite difference approximations (which creates the cross stencil for the plane problem), and the solution is sought on a discrete subset [math]D_N[/math] of the [math]N[/math]-dimensional space. The boundary conditions are also approximated by finite differences.
1.2 Mathematical description of the algorithm
Here, we examine a finite difference scheme for the most common problem related to the Poisson equation in the three-dimensional space:
[math] \frac{\phi_{i+1,j,k}-2\phi_{i,j,k}+\phi_{i-1,j,k}}{ \Delta x^2}\,+\, \frac{\phi_{i,j+1,k}-2\phi_{i,j,k}+\phi_{i,j-1,k}}{ \Delta y^2}\,+\, \frac{\phi_{i,j,k+1}-2\phi_{i,j,k}+\phi_{i,j,k-1}}{ \Delta z^2} = f_{i,j,k},~(i,j,k) \in D_h, [/math]
where [math]D_h=\{0:N_x-1\}\times\{0:N_y-1\}\times\{0:N_z-1\} [/math] is a parallelepiped in the grid domain. For simplicity, we impose the so-called 3-D periodic boundary conditions
[math] \begin{align} \phi_{0,j,k} &= \phi_{N_x,j,k},\\ \phi_{i,0,k} &= \phi_{i,N_y,k},\\ \phi_{i,j,0} &= \phi_{i,j,N_z}. \end{align} [/math]
The periodic boundary conditions are automatically satisfied if the solution is represented via the conventional discrete inverse Fourier transform:
[math] \phi_{i,j,k}=\frac{1}{N_x N_y N_z}\sum_{l=0}^{N_x-1}\sum_{m=0}^{N_y-1}\sum_{n=0}^{N_z-1}\Phi_{l,m,n}e^{2\pi \overline{i} \left(\frac{il}{N_x}+\frac{jm}{N_y}+\frac{kn}{N_z}\right)}. [/math]
Here, [math]\overline{i}=\sqrt{-1}[/math]. A similar representation is applied to the right-hand side [math]f_{i,j,k}[/math]. The convenience of using the Fourier transform for solving the discrete Poisson equation consists in the fact that the basis functions of the Fourier expansion are eigenfunctions of the discrete Laplace operator. Namely, by substituting the Fourier expansions of the unknown function [math]\phi_{i,j,k}[/math] and the right-hand side [math]f_{i,j,k}[/math] into the original equation, we obtain
[math] \Phi_{l,m,n}=-\frac{F_{l,m,n}}{4\left[\sin^2\left(\frac{\pi l}{N_x}\right) + \sin^2\left(\frac{\pi m}{N_y}\right) + \sin^2\left(\frac{\pi n}{N_z}\right) \right]}, [/math]
where [math]F_{l,m,n}[/math] is the Fourier transform of the right-hand side. This makes obvious the algorithm for solving the equation: first, the right-hand side is expanded into the Fourier series, then the above formula is used for calculating the Fourier coefficients of the solution; finally, the solution is reconstructed by applying the inverse Fourier transform.
1.3 Computational kernel of the algorithm
The one-dimensional Fourier transform is the computational kernel of this algorithm. Indeed, the discrete inverse Fourier transform can be written as
[math] \phi_{i,j,k}=\frac{1}{N_x} \sum_{l=0}^{N_x-1} \left[ e^{2\pi \overline{i}\left(\frac{il}{N_x}\right)} \frac{1}{N_y} \sum_{m=0}^{N_y-1} \left[ e^{2\pi \overline{i}\left(\frac{jm}{N_y}\right)} \frac{1}{N_z} \sum_{n=0}^{N_z-1} \Phi_{l,m,n}e^{2\pi \overline{i} \left(\frac{kn}{N_z}\right)}\right]\right]. [/math]
One can see that the three-dimensional Fourier transform reduces to the sequence of three one-dimensional transforms. A widely used tool for calculating the one-dimensional transform is an efficient algorithm called the fast Fourier transform (FFT) [1].
1.4 Macro structure of the algorithm
From the above discussion, it is clear that the fast Fourier transform is the basic macro operation in the algorithm for solving the Poisson equation. In what follows, we denote this operation by [math]\text{FFT}_i,~i=x,y,z[/math] in accordance with its direction, while the inverse Fourier transforms are denoted by [math]\text{IFFT}_i,~i=x,y,z[/math].
1.5 Implementation scheme of the serial algorithm
For brevity, we use the following notation for the grid functions: [math]\phi_h=\{\phi_{i,j,k},~(i,j,k) \in D_h\},~F_h=\{F_{l,m,n},~l=0,...,N_x-1,~m=0,...,N_y-1,~n=0,...,N_z-1\}[/math]. A similar notation is used for [math]f_h[/math] and [math]\Phi_h[/math]. Now, the algorithm can be written as follows:
- Calculate [math]\text{FFT}_x(f_h)[/math]
- Calculate [math]\text{FFT}_y(\text{FFT}_x(f_h))[/math]
- Calculate [math]\text{FFT}_z(\text{FFT}_y(\text{FFT}_x(f_h)))=F_h[/math]
- Find [math]\Phi_h[/math] from [math]F_h[/math]
- Calculate [math]\text{IFFT}_x(\Phi_h)[/math]
- Calculate [math]\text{IFFT}_y(\text{IFFT}_x(\Phi_h))[/math]
- Calculate [math]\text{IFFT}_z(\text{IFFT}_y(\text{IFFT}_x(\Phi_h)))=\phi_h[/math]
It is remarkable that the entire algorithm can be implemented with the use of a single three-dimensional array of size [math]N_x \times N_y \times N_z[/math]. Indeed, the results of one-dimensional Fourier transforms in one direction can be written to the input array, and the resulting array can be used as an input for transforms in the next direction. Operation (4) is an entrywise array modification.
1.6 Serial complexity of the algorithm
If the domain [math]D_h[/math] is a cube, then [math]N_x=N_y=N_z=N=2^k[/math]. Then the complexity of a one-dimensional FFT in any direction is [math]N(\text{log}_2 N)[/math] operations (in order). At each stage, excepting the fourth, one performs [math]N^2[/math] one-dimensional FFTs, while the number of operations at stage 4 is [math]N^3[/math] (again in order). Thus, the overall number of arithmetic operations is [math]N^3(6\text{log}_2 N+1)[/math].
1.7 Information graph
We present the information graph only for stages 1-4 because stages 5-7 are implemented analogously to stages 1-3.
The information dependence between the layers of the graph can be stated as follows. A one-dimensional FFT depends on the results of an earlier FFT in a perpendicular direction (which was performed at a previous layer) only if the one-dimensional sections processed by these FFTs intersect each other. Consequently, each [math]FFT_{i,y,k},~i=1,...,N[/math], with respect to the second coordinate depends on the results of each [math]FFT_{x,j,k},~j=1,...,N[/math], w.r.t. the first coordinate if and only if these transformations are executed in the same plane [math]k=K[/math]. Accordingly, each [math]FFT_{i,j,z},~j=1,...,N[/math], w.r.t. the third coordinate depends on each [math]FFT_{i,y,k},~k=1,...,N[/math], w.r.t. the second coordinate if they belong to the same plane [math]i=I[/math].
1.8 Parallelization resource of the algorithm
One can distinguish at least two levels of parallelism in the above algorithm. In the first place, the execution of each one-dimensional FFT can be distributed between computational cores. Second, at each stage of the algorithm (that is, at each layer of the graph), one-dimensional FFTs are independent of each other and can be performed in parallel (coordinate parallelism). In order to compare the parallel complexity (under coordinate parallelization) and the serial one, we take the one-dimensional FFT and the entrywise updating of the array (stage 4 of the algorithm) as the basic operations. Then the serial algorithm consists of [math]6N^2[/math] one-dimensional FFTs, executed sequentially, and [math]N^3[/math] operations over individual array entries. As for the parallel algorithm, it performs 6 steps of one-dimensional FFTs and one step of entrywise modification. By expressing the macro operations introduced above in terms of elementary arithmetic operations, we conclude that the complexity of the serial algorithm is [math]6N^3(\text{log}_2N)+N^3[/math] operations, whereas the complexity of the parallel algorithm is [math]6N(\text{log}_2N)+1[/math] operations. It follows that, for large [math]N[/math], the ratio of the serial to parallel complexity tends to [math]\propto N^2(6\text{log}_2N \text{ln}2+\text{ln}2+5)[/math].
1.9 Input and output data of the algorithm
The right-hand side of the equation delivered as a three-dimensional array of size [math]N \times N \times N[/math] is the input of the algorithm, while the solution stored in a three-dimensional array of the same size is its output.
1.10 Properties of the algorithm
As shown above, the computational complexity of the parallel algorithm decreases faster than [math]N^2[/math] if the number of computing units is unlimited and the input data are voluminous.
Since the size of the input data is [math]N^3[/math], the computational power of the serial algorithm is [math]6\text{log}_2N+1[/math], while the computational power of the parallel algorithm is [math](6N\text{log}_2N+1)/N^3 \propto N^{-2}[/math] (for large [math]N[/math]).
2 Software implementation of the algorithm
2.1 Implementation peculiarities of the serial algorithm
Let us give an example of implementation in Fortran-90 with the use of the library FFTW. In this case, the direct three-dimensional Fourier transform of the right-hand side is performed as the sequence of three one-dimensional FFTs (stages 1-3 in section 1.5). The resulting Fourier coefficients are divided by the eigenvalues of the Laplace operator (stage 4 in section 1.5) (see the comments in the code):
! x - transform
do k = 1, k1
do j = 1, j1
workfor_x(1:i1) = cmplx(f_r(1:i1,j,k),f_i(1:i1,j,k))
call FFTW_EXECUTE_DFT(planfor_x,workfor_x,workfor_x)
f_r(1:i1,j,k) = real(real (workfor_x(1:i1)) )
f_i(1:i1,j,k) = real(aimag(workfor_x(1:i1)) )
enddo
enddo
! y - transform
do k = 1, k1
do i = 1, i1
workfor_y(1:j1) = cmplx(f_r(i,1:j1,k),f_i(i,1:j1,k))
call FFTW_EXECUTE_DFT(planfor_y,workfor_y,workfor_y)
f_r(i,1:j1,k) = real(real (workfor_y(1:j1)) )
f_i(i,1:j1,k) = real(aimag(workfor_y(1:j1)) )
enddo
enddo
! z - transform and divide by eigenvalue
do j = 1, j1
do i = 1, i1
workfor_z(1:k1) = cmplx(f_r(i,j,1:k1),f_i(i,j,1:k1))
call FFTW_EXECUTE_DFT(planfor_z,workfor_z,workfor_z)
f_r(i,j,1:k1) = real(real (workfor_z(1:k1)) )
f_i(i,j,1:k1) = real(aimag(workfor_z(1:k1)) )
! Divide by eigenvalue of Laplace operator
sinx = sin((i-1) * Pi / i1)^2 * 4
siny = sin((j-1) * Pi / j1)^2 * 4
do k = 1, k1
sinz = sin((k-1) * Pi / k1)^2 * 4
f_r(i,j,1:k1) = f_r(i,j,1:k1)/(sinx + siny + sinz)
f_i(i,j,1:k1) = f_i(i,j,1:k1)/(sinx + siny + sinz)
enddo
enddo
enddo
As soon as the Fourier coefficients have been found, the inverse Fourier transform is performed in a similar way:
! inverse x - transform
do k = 1, k1
do j = 1, j1
workback_x(1:i1) = cmplx(f_r(1:i1,j,k),f_i(1:i1,j,k))
call FFTW_EXECUTE_DFT(planback_x,workback_x,workback_x)
f_r(1:i1,j,k) = real(real (workback_x(1:i1)) )/real(i1)
f_i(1:i1,j,k) = real(aimag(workback_x(1:i1)) )/real(i1)
enddo
enddo
! inverse y - transform
do k = 1, k1
do i = 1, i1
workback_y(1:j1) = cmplx(f_r(i,1:j1,k),f_i(i,1:j1,k))
call FFTW_EXECUTE_DFT(planback_y,workback_y,workback_y)
f_r(i,1:j1,k) = real(real (workback_y(1:j1)) )/real(j1)
f_i(i,1:j1,k) = real(aimag(workback_y(1:j1)) )/real(j1)
enddo
enddo
! inverse z - transform
do j = 1, j1
do i = 1, i1
workback_z(1:k1) = cmplx(f_r(i,j,1:k1),f_i(i,j,1:k1))
call FFTW_EXECUTE_DFT(planback_z,workback_z,workback_z)
f_r(i,j,1:k1) = real(real (workback_z(1:k1)) )
f_i(i,j,1:k1) = real(aimag(workback_z(1:k1)) )
enddo
enddo
One can see that the way in which the loops are embedded corresponds to the columnwise arrangement of the array entries in Fortran. This makes it possible to significantly reduce the memory access time. The same auxiliary arrays are repeatedly used, which should ensure the fast interaction with the cache for grids of modest size. However, this implementation is not optimal. For instance, instead of repeatedly calling the subroutine FFTW_EXECUTE_DFT, a subroutine performing the prescribed set of one-dimensional FFTs could be called just once.
The above program listings make also obvious the parallelization resource of the algorithm; namely, all the one-dimensional Fourier transforms are executed independently of each other. The simplest way to take advantage of this fact is to insert the corresponding OpenMP instructions before the inner loops, provided that a shared memory multicore computer is used. If the program is executed on a distributed memory cluster, then one should make sure that the input array, containing the right-hand side, is distributed between the processors and intermediate arrays are transposed when transiting from the FFT in one direction to that in the other.
Note that, for performing one-dimensional FFTs, the above implementation calls a procedure from the FFTW library that processes arrays of complex numbers. On the other hand, there exists versions of FFT for processing real arrays that are twice as fast as the complex algorithm (for more details, see the site of the FFTW project ([1]).
An important circumstance is the fact that most variants of FFT, including the Cooley-Tukey algorithm, have very low errors when performed in floating point arithmetic. The relative error of the Cooley-Tukey algorithm is bounded above by [math]O(\epsilon \log N)[/math]. (For comparison, note that the straightforward implementation via the formulas of the discrete Fourier transform has the relative error [math]O(\epsilon N^{3/2})[/math].) [2]. Thus, one can accelerate the algorithm by using single precision floating point numbers, and no significant increase in error will occur.
2.2 Locality of data and computations
2.2.1 Locality of implementation
2.2.1.1 Structure of memory access and a qualitative estimation of locality
Figure 3 presents the memory access profile for an implementation of the discrete Fourier transform as a method for solving the Poisson equation. This profile includes accesses to several service arrays (fragment 1) and to the main array (fragment 2). One can see that the profile can be split into two virtually identical stages (see the vertical yellow line). One notices that, on the whole, the overall number of entries is modest (slightly more than 1000 entries); on the other hand, the profile consists of about 100 thousands accesses. It is seen that the density of accesses is higher in the upper part of the profile. This probably indicates a high level of locality; however, the accesses to the main array are fairly scattered. Let us examine the locality matter in more detail.
Figure 4 shows fragment 1, which contains all the excesses to the service arrays. From this graph, one cannot say what is the structure of accesses inside the blocks; however, this is not required here. The overall number of referenced addresses is small; there are only about 100 of them used throughout the program. Moreover, they are organized into a block structure, which improves the memory interaction. Thus, one can confidently declare that, in this case, the locality (both spatial and temporal) is very high.
Now, we look at the main array. Figure 5 presents fragment 2 of the general profile shown in fig. 3. One can immediately notice the following fact. The graph shows at most several hundreds of accesses, whereas the axis X contains the mark 35000. This means that the vast majority of accesses occurs beyond this fragment; namely, they occur in the service arrays. Thus, on the whole, accesses to the main array occur not very often. An inspection of the original code confirms this conclusion. Indeed, the data stored in the main array are actually only copied to the service arrays. Here, they undergo Fourier transforms, after which the data return to the main array.
According to this fragment, the data that are used in succession are often located close to each other; however, they are rarely used again. Thus, the spatial locality of this fragment is not bad, but its temporal locality is rather low.
Taking into account that the majority of accesses occur to the service arrays and the accesses to the main array have a sufficiently high locality, one can say that, on the whole, the general profile has a high level of locality (both spatial and temporal).
2.2.1.2 Quantitative estimation of locality
The basic fragment of the implementation used for obtaining quantitative estimates is given here [2] (the Kernel function). The start-up parameters are described here [3].
The first estimate is produced on the basis of daps, which estimates the number of memory accesses (reading and writing) per second. This characteristic, used for the interaction with memory, is an analog of the flops estimate. It largely estimates the performance of this interaction rather than locality. Nevertheless, this characteristic (in particular, its comparison with the next characteristic cvg) is a good source of information.
Figure 6 presents the values of daps for implementations of popular algorithms. They are arranged in increasing order (in general, the larger daps, the higher efficiency). Here, we can observe a somewhat unexpected result; namely, despite the assumed high locality, the performance of memory interaction is fairly low.
There seem to be several reasons for this result. The first reason is that the repeated copying of data to and from the service arrays deteriorates both spatial and temporal locality. Second, in the service arrays, the data undergo Fourier transforms based on the Cooley-Tukey algorithm [4]. From this reference, one can see that the performance of memory interaction for the Cooley-Tukey algorithm is very similar to the performance for this Poisson equation solver. The above conclusions about the locality of accesses to the service arrays, where the Fourier transforms are performed, remain also valid for the Cooley-Tukey algorithm; namely, the locality of memory accesses is very high despite a low performance.
The second characteristic, called cvg, is intended for obtaining a locality estimate that would be more machine independent. It determines how often a program needs to fetch data to the cache memory. Accordingly, smaller values of cvg correspond to less often fetch operations and better locality.
Figure 7 presents the values of cvg for the same set of implementations. They are arranged in decreasing order (in general, the smaller cvg, the higher locality). It can be seen that, similarly to the case of the Cooley-Tukey algorithm, the locality is estimated as very high. The implementation of this algorithm takes up a prominent portion of the general profile for the program under discussion. Consequently, the reasons for this disagreement between locality and performance are identical for both cases (see the detailed discussion of the Cooley-Tukey algorithm).
2.3 Possible methods and considerations for parallel implementation of the algorithm
As shown above, there are at least two levels of parallelism in the algorithm under discussion. The first level concerns the parallel implementation of a one-dimensional FFT, while the second level is the coordinate parallelism, that is, a parallel execution of independent one-dimensional FFTs [3]. These two levels can be combined; for instance, different FFTs can be distributed between MPI processes, while each individual FFT can be distributed between threads with shared memory (POSIX, OpenMP, etc.). Below, we present some details of the base parallelization algorithm using the coordinate parallelism, discuss its shortcomings, and list possible ways of improving its scalability.
In the context of this problem, the coordinate parallelism is realized by decomposing a three-dimensional domain into subdomains. In such subdomains, one-dimensional FFTs can be performed independently of each other. Presently, two implementations of this approach are available. They are one-dimensional domain decomposition (slab decomposition) and two-dimensional decomposition (pencil decomposition) [4]. Under one-dimensional decomposition, one requires a smaller number of MPI-Sends; however, the maximum number of processes/cores is bounded by the maximum (over three dimensions) size of the domain. Consequently, from the viewpoint of massive parallelism, two-dimensional decomposition is preferable. In what follows, we discuss exactly this approach.
Under MPI implementation, one must take into account the overhead expenses caused by data transfers between processors. The algorithm under discussion also requires that the so-called array transpositions be performed. At each of the three stages of Fourier transforms (stages 1-3 and 5-7 in section 1.5), these transpositions ensure finding the array entries along the corresponding direction in the memory of the corresponding processor.
Consider the relation between calculations and data transfers under MPI implementation. Suppose that the coordinate parallelization for [math]M[/math] MPI processes is used. Then the computational complexity of the algorithm is [math] \left(6N^3(\log_2 N) + N^3\right)/M[/math], which is at the best [math] 6N(\log_2 N) + N[/math] (if [math]N^2[/math] processes are employed). The total amount of MPI packages (sent and received) is [math]8N_{px}+4N_{py}[/math], while the total amount of array entries (sent and received) is [math]12N^3/(N_{px}N_{py})[/math] for each process (here, [math]N_{px},~N_{py}[/math] are the number of processes along the axes [math]x[/math] and [math]y[/math], respectively, for the two-dimensional MPI decomposition used; see the above figure). Since [math]M=N_{px}N_{py}[/math], we conclude that the ratio "the number of forwarded array entries to the number of arithmetic operations" is [math]\alpha=12/(6\log_2 N + 1)[/math]. Thus, the ratio "the amount of data traffic to the number of operations" is independent of the number of MPI processes, and this ratio decreases with the growth in the problem size. However, the decay becomes slow for large [math]N[/math] (the saturation phase). On the other hand, consider the case of a fixed problem size. Then an increase in the number of processes does not alter the parameter [math]\alpha[/math], but it causes an increase in the number of packages ([math]\propto \sqrt{M}[/math], in view of [math]N_{px} \propto N_{py} \propto \sqrt{M}[/math] and the above formulas). Due to the network latency, this increases the time for transferring data; consequently, the acceleration of the algorithm slows down.
The above analysis can only serve for approximate estimates because no times for computation and MPI exchanges were given. More detailed calculations for one- and two-dimensional domain decompositions are abundant in the literature [5][6].
In the following sections, we profile the program code of a parallel implementation of this algorithm. In this code, calls of one-dimensional FFTs in three directions shown in the listing in section 2.1 alternate with calls on special procedures effecting array transpositions. The real and imaginary parts of an array are transposed separately (which is not optimal because better results are obtained by sending complex numbers) with the use of point-to-point communications. Asynchronous forwarding based on MPI_ISEND and MPI_IRECV is used.
2.4 Scalability of the algorithm and its implementations
The variable start-up parameters of this implementation and their ranges are as follows:
- the number of processors is a squared integer in the range [4 : 100];
- the domain size varies in the range [100x100x100 : 250x100x100] with the step 50x100x100.
The experiments resulted in the following range for the implementation efficiency of this algorithm:
- the minimum efficiency is 0,132%;
- the maximum efficiency is 1,34%.
The following two figures show how the performance and efficiency of this implementation depend on the variable start-up parameters.
These results are in good qualitative agreement with the estimates given in the preceding section. For instance, one can see from fig. 8 that the scalability of the algorithm improves with the growth of [math]N[/math]; that is, the performance improves faster with the growing number of cores. This is consistent with the fact that [math]\alpha[/math] decreases with the growth of [math]N[/math], although the dependence revealed by the expression [math]\alpha=12/(6\log_2 N+1)[/math] is considerably more weak. On the whole, the scalability of the algorithm is rather weak. Suppose that the domain has the size 250х100х100. When the number of cores increases by a factor 25 (from 4 to 100), the performance of calculations increases only sixfold.
2.5 Dynamic characteristics and efficiency of the algorithm implementation
2.6 Conclusions for different classes of computer architecture
2.7 Existing implementations of the algorithm
Below, we list the presently available implementations of the three-dimensional FFT and their basic characteristics.
| Implementation | Domain decomposition | Parallelization technologies | Exchanges in MPI |
|---|---|---|---|
| FFTW | 1D | MPI+OpenMP | |
| FFTE | 1D & 2D | MPI+OpenMP | MPI_ALLTOALL |
| P3DFFT | 1D & 2D | MPI | |
| cuFFT | |||
| AccFFT | 1D & 2D | MPI, CUDA | MPI_ALLTOALL |
| MKL FFT | 1D | MPI | |
| PFFT | MPI |
3 References
- ↑ Marchuk G.I. Methods of Computational Mathematics, Moscow, 1977, 456 p. (in Russian)
- ↑ Gentleman, W. M.; Sande, G. (1966). "Fast Fourier transforms—for fun and profit". Proc. AFIPS 29: 563–578. doi:10.1145/1464291.1464352
- ↑ Anshu Dubey and Daniele Tessera. Redistribution strategies for portable parallel FFT: a case study. Concurrency and Computation: Practice and Experience, 13(3):209–220, 2001.
- ↑ Orlando Ayala, Lian-Ping Wang, Parallel implementation and scalability analysis of 3D Fast Fourier Transform using 2D domain decomposition, Parallel Computing, Volume 39, Issue 1, January 2013, Pages 58-77, ISSN 0167-8191, http://dx.doi.org/10.1016/j.parco.2012.12.002.
- ↑ Orlando Ayala, Lian-Ping Wang, Parallel implementation and scalability analysis of 3D Fast Fourier Transform using 2D domain decomposition, Parallel Computing, Volume 39, Issue 1, January 2013, Pages 58-77, ISSN 0167-8191, http://dx.doi.org/10.1016/j.parco.2012.12.002.
- ↑ P. Dmitruk, L.-P. Wang, W.H. Matthaeus, R. Zhang, D. Seckel, Scalable parallel FFT for spectral simulations on a Beowulf cluster, Parallel Comput. 27 (2001) 1921–1936.
