- version: 1.0.1 (based on FFB 8.1)
- date: 2015/06/02
- contact: [email protected]
This mini app, FFB-MINI, is based on a general purpose thermal flow simulation program FrontFlow/blue (FFB), being developed by Institute of Industrial Science, the University of Tokyo. FFB solves the unsteady incompressible Navier-Stokes equations by finite element method. FFB is MPI parallelized with domain decomposition. Although not using thread based parallelism explicitly, the current version of FFB includes optimizations for the K computer with the automatic thread parallelization functionality.
To refer to the original FFB application:
Contact point of the original FFB application: Yoshinobu Yamade [email protected]
To build and run the program, Fortran90 and C compilers, an MPI library and GNU Make are required. To utilize the automatic parallelization with the Fujitsu compiler for the K computer and Fujitsu FX10 system, METIS library is required. For calculations using automatic mesh-refinement (REFINER), REVOCAP_Refiner library is required.
-
Obtain the package and extract its contents.
-
Go to the src directory and edit the file "make_setting" according to your compilers and libraries. There are several examples:
- make_setting.intel : For using Intel compilers
- make_setting.gcc : For using GCC compilers
- make_setting.fx10 : For K
- make_setting.fx10 : For Fujitsu FX10
-
Run make command in the src directory. After successful compilation, an executable file named "les3x.mpi" and its wrapper program "ffb_mini" are generated in the src directory.
Test shell scripts are provided in the test directory. To run the test interactively, simply run "make test" in the src or test directory. ("make test_k" for the K computer, "make test_fx10" for the FX10 system at the University of Tokyo)
These test scripts runs the program ffb_mini with 8 MPI processes, and compares the computed results with reference data, then judges whether the calculation is correctly completed or not.
This mini app is based on a fluid simulation code les3x.mpi included in the FFB package. The basic loop structure of les3x.mpi is as below:
time-step-loop {
# solve velocity fields (U,V,W) by Crank-Nicolson scheme
U-loop {
# BiCGSTAB iteration
}
V-loop {
# BiCGSTAB iteration
}
W-loop {
# BiCGSTAB iteration
}
# solve pressure Poisson eq.
RCM-outer-loop {
RCM-inner-loop {
# BiCGSTAB iteration
}
}
# velocity filed correction
}
To solve the velocity field equations formulated by Crank-Nicolson scheme, BiCGSTAB is used for each of directions.
The pressure Poisson equation is solved by Residual Cutting Method (RCM). BiCGSTAB with the iteration number fixed to 10 is used as the inner-loop-solver of RCM.
When a wrapper program ffb_mini starts, first it makes input data for les3x.mpi, then runs les3x.mpi internally. ffb_mini calculates 3D cavity flow problems, in which the computational domain is a rectangular region surrounded by fixed walls, except for "+Y wall" moving X direction.
Notes that les3x.mp and ffb_mini always perform floating point operations in single-precision.
To run les3x.mpi, please refer to the original FFB manual "doc/Manual_FrontFlow_blue.8.1_ja.pdf".
For ffb_mini, all the necessary parameters should be given as the command line options:
> ffb_mini Npx Npy Npz N [key1=val1 key2=val2 ...]
where Npx, Npy, Npz and N are mandatory. Mesh size and how to decompose
the domain to MPI processes are specified by these four parameters in a weak
scaling fashion. ffb_mini solves the 3D cavity flow in a rectangular region
composed of NpxxNpyxNpz cubes. Each cube has N^3 hexahedral elements and
is assigned to an MPI process.
key1=val1, key2=val2, ..., are optional.
The list of optional parameters are as below:
ntime=int computing time steps (default 100)
isolp=int type of the Poisson eq. solver, 1 (BiCGSTAB) or 2 (RCM)
(default 2)
dt=float time step delta (default 0.005)
nmaxp=int max. number of the Poisson iterations (total number of RCM inner
loop iterations) (default 50)
nmaxt=int max. number of the velocity fields iterations (default 5)
epsp=float Poisson iteration convergence tolerance
(if epsp=0.0, iterates nmaxp times irrespective of the
convergence) (default 0.0)
epst=float velocity fields iteration convergence tolerance
(if epst=0.0, iterates nmaxt times irrespective of the
convergence) (default 0.0)
color_partsize=int
size of coloring regions (if color_partsize=0, no coloring)
(default 0)
reorder_ndiv=int
size of the block used for node number reordering
(if reorder_ndiv=0, no reordering) (default 0)
unroll=str loop unrolling flog for matrix-vector multiplication in
subroutine CALAX3
on (full unrolling) or off (no unrolling) (default off)
For the K computer and Fujitsu FX10 system, we recommend following options:
color_partsize=2000 reorder_ndiv=10 unroll=on
Wrapper program ffb_mini outputs following files:
- MESH.PXXXX ... mesh data
- BOUN.PXXXX ... boundary condition data
- PARMLES3X.PXXXX ... runtime parameters
- FLOW.PXXXX ... calculation result (flow data)
- les3x.log.PXXXX ... runtime logs
where "XXXX" in suffix PXXXX describes an integer between 1 and Np, the number of MPI processes. For the rank 0 process, runtime logs are sent to stdout instead of les3x.log.P0001.
Notes that MESH.PXXXX, BOUN.PXXXX and PARMLES3X.PXXXX are used as input data to les3x.mpi, which runs internally in ffb_mini.
The current model setting targeted at exa-scale simulations is as follows:
- 10^12 elements system size
- less than 100-hours real time with 4x10^4 simulation time steps (about 10sec/step)
-
Delete all programs other than les3x.mpi
-
Reduce code size by limiting the functionality:
- restrict to unsteady flow calculations using hexahedral element mesh in pressure-defined-on-element mode
- delete following functionality: Large Eddy Simulations (LES), cavitation model, Volume of Fluid (VOF), overset and multi-frame, mid-node, Arbitrary Lagrangian and Eulerian Method (ALE), fan model and porous model
- delete sampling output and history output
-
Remove METIS and REVOCAP_Refiner libraries
-
Add a wrapper program ffb_mini
-
Add test programs
-
Other code clean-up
There are three major hot spot subroutines in the mini app version of les3x.mpi.
- GRAD3X: calculates velocity gradient
- FLD3X2: calculates divergence of velocity gradient
- CALAX3: calculates matrix-vector multiplications
GRAD3X and FLD3X2 are called from subroutine CALLAP, which calculates the Laplacian to solve the pressure Poisson equation. CALAX3 is called in the velocity fields BiCGSTAB loop (RCM-inner-loop).
The call counts of the hot spot subroutines for one time step iteration are:
- GRAD3X: 2.4 * NMAXP + 1
- FLD3X2: 2.4 * NMAXP + 1
- CALAX3: 6 * NMAXT + 3
where NMAXP and NMAXT are the number of the pressure Poisson loop iterations (total iteration number of RCM-inner loop) and the number of the velocity field BiCGSTAB iterations, respectively.
There are two major parts of the inter-process communications in the time step loop:
- Superimposing the data on the nodes located simultaneously on the neighboring sub-domains, in subroutine DDCMX
- All-reduce communications for the global inner products, in subroutine DDCOM2
In subroutine RCMELM, which solves the pressure Poisson equation by RCM, the call counts of subroutine DDCMX and DDCOM2 for one time step iteration are
- DDCOMX: 2.4 * NMAXP + 1
- DDCOM2: 5.5 * NMAXP + 2
In subroutine VEL3D1, which solve velocity fields by BiCGSTAB,
- DDCOMX: 6 * NMAXT + 3
- DDCOM2: 15 * NMAXT + 6
- Minami et al., "Performance Tuning and Evaluation of Sparse matrix-vector multiplication on the K computer", IPSJ High Performance Computing Symposium, 2012. pp.23-31, 2012. (in Japanese)
- Kumahata et al., "Kernel performance improvement for the FEM-based fluid analysis code on the K computer", Procedia Computer Science Vol.18, pp.2496-2499, 2013.
- Kumahata et al.,"Performance Tuning for Gradient Kernel of the FrontFlow/blue on the K computer", IPSJ Transactions on Advanced Computing Systems Vol.6, No.3 pp.31-42, 2013. (in Japanese)
-
1.0.0 (May 21, 2015)
- Inital release
-
1.0.1 (June 2, 2015)
- Fix a bug in makemesh