782.lbm_r
SPEC CPU®2026 Benchmark Description

Benchmark Name

782.lbm_r

Benchmark Program General Category

Computational Fluid Dynamics, Lattice Boltzmann Method

Benchmark Author

Thomas Pohl <thomas.pohl [at] gmail [dot] com>

Benchmark Description

This program implements the so-called "Lattice Boltzmann Method" (LBM) to simulate incompressible fluids in 3D as described in [1]. It is the computationally most important part of a larger code which is used in the field of material science to simulate the behavior of fluids with free surfaces, in particular the formation and movement of gas bubbles in metal foams. For benchmarking purposes and easy optimization for different architectures, the code makes extensive use of macros which hide the details of the data access. A visualization of the results of the submitted code can be seen below (flow through a porous medium, grid size 200x200x130, 1200 time steps).

Flow of a fluid through an array of spheres

Input Description

The lbm program requires several command line arguments:

lbm <time steps> <result file> <0: nil, 1: cmp, 2: str> <0: ldc, 1: channel flow> [<obstacle file>]

Description of the arguments:

<time steps>

number of time steps that should be performed before storing the results

<result file>

name of the result file

<0: nil, 1: cmp, 2: str>

determines what should be done with the specified result file: action '0' does nothing; with action '1' the computed results are compared with the results stored in the specified file; action '2' stores the computed results (if the file already exists, it will be overwritten)

<0: ldc, 1: channel flow>

chooses among two basic simulation setups, lid-driven cavity (shear flow driven by a "sliding wall" boundary condition) and channel flow (flow driven by inflow/outflow boundary conditions)

[<obstacle file>]

optional argument that specifies the obstacle file which is loaded before the simulation is run

The basic steps of the simulation code are as follows:

1. If an obstacle file was specified it is read and the obstacle cells are set accordingly.
2. The specified number of time steps are calculated in the selected simulation setup (lid-driven cavity or channel flow).
3. Depending on the action chosen the result is either stored, compared to an existing result file, or thrown away.

Benchmarking

For benchmarking purposes, where the SPEC® tools are used to validate the solution, only the computed results are stored.

In the Lattice Boltzmann Method, a steady state solution is achieved by running a sufficient number of model time steps. For the refrate workload, 1200 time steps are computed for an obstacle file. The simulation domain is x=200, y=200 and z=130.

The geometry used in the test and training workload is the same as used in the reference benchmark workload with fewer timesteps.

Obstacle File Format

The file format which specifies the location of obstacle and fluid cells is a simple ASCII format. The dot character '.' stands for a fluid cell, while all other characters (here '#') denote an obstacle cell. Each line represents the cells along the x axis. After each line a newline has to be included. After a complete x/y plane another newline has to be included.

Below you see an example of an obstacle file for the simulation domain x = 6, y = 5, and z = 3. The red comments just show the corresponding coordinates for each line and must not be included in the obstacle file itself.

......                (y = 0, z = 0)
...#..                (y = 1, z = 0)
..##..                (y = 2, z = 0)
.###..                (y = 3, z = 0)
......                (y = 4, z = 0)

......                (y = 0, z = 1)
......                (y = 1, z = 1)
...#..                (y = 2, z = 1)
..##..                (y = 3, z = 1)
......                (y = 4, z = 1)

......                (y = 0, z = 2)
......                (y = 1, z = 2)
......                (y = 2, z = 2)
...#..                (y = 3, z = 2)
......                (y = 4, z = 2)

Output Description

By default, the benchmark uses action '0' and no output file is generated. Instead, the statistical summary information is printed to stdout and used for validation.

If the store action '2' is specified in the command line arguments, a result file containing the 3D velocity vector for each cell is stored.

The default file format is a sequence of binary single precision values (little endian) with the following ordering:

v_x(0,0,0), v_y(0,0,0), v_z(0,0,0),     v_x(1,0,0), v_y(1,0,0), v_z(1,0,0),     v_x(2,0,0), v_y(2,0,0), v_z(2,0,0),         ...,     v_x(X,0,0)  v_y(X,0,0)  v_z(X,0,0),
v_x(0,1,0), v_y(0,1,0), v_z(0,1,0),     v_x(1,1,0), v_y(1,1,0), v_z(1,1,0),     v_x(2,1,0), v_y(2,1,0), v_z(2,1,0),         ...,     v_x(X,1,0)  v_y(X,1,0)  v_z(X,1,0),
...
v_x(0,Y,0), v_y(0,Y,0), v_z(0,Y,0),     v_x(1,Y,0), v_y(1,Y,0), v_z(1,Y,0),     v_x(2,Y,0), v_y(2,Y,0), v_z(2,Y,0),         ...,     v_x(X,Y,0)  v_y(X,Y,0)  v_z(X,Y,0),
...
v_x(0,Y,Z), v_y(0,Y,Z), v_z(0,Y,Z),     v_x(1,Y,Z), v_y(1,Y,Z), v_z(1,Y,Z),     v_x(2,Y,Z), v_y(2,Y,Z), v_z(2,Y,Z),         ...,     v_x(X,Y,Z)  v_y(X,Y,Z)  v_z(X,Y,Z)

Although the default format cannot be altered via command line arguments, it is possible to change the output precision to double precision in config.h.

If the computed result should be compared to an existing result (action '1'), the program returns the maximum absolute difference of the velocity comparing each cell individually. If the difference is smaller than a certain threshold, the two results are considered to be equal.

Programming Language

ANSI C

Differences from SPEC CPU 2017 519.lbm_r

782.lbm_r is highly similar to 519.lbm_r, with the exception that the CPU 2026 version has been resized to use 1.6 GiB of memory, which is about 4 times the size of the CPU 2017 version. SPEC decided to retain the benchmark in the new CPU 2026 suite because it significantly exercises main memory.

Known portability issues

When using larger inputs than those included with the the refrate workload, systems that do not support allocation of large static arrays may need to use the portability flag "SPEC_LBM_NO_HUGE_ARRAY".

Sources and Licensing

LBM is licensed directly to SPEC by the author, Thomas Pohl.

References

1. Y.-H. Qian, D. d'Humieres, and P. Lallemand: Lattice BGK models for Navier-Stokes equation. Europhys. Lett. 17(6): 479-484, 1992
2. Thomas Pohl, Markus Kowarschik, Jens Wilke, Klaus Iglberger, and Ulrich Rüde: Optimization and Profiling of the Cache Performance of Parallel Lattice Boltzmann Codes. Parallel Processing Letter 13(4) 549-560, 2003, postscript copy available here

Copyright © 2026 Standard Performance Evaluation Corporation (SPEC®)