765.roms_r
ROMS is a Regional Ocean Modeling System
The Regional Ocean Modeling System (ROMS) is written and maintained by "The ROMS/TOMS Group", which is described as:
765.roms_r was submitted to the SPEC CPU v8 Benchmark Search Program by Hans Joraandstad hans [dot] joraandstad [at] gmail [dot] com.
ROMS is a free-surface, terrain-following, primitive equations ocean model. It is coded in parallel for multi-threaded and multi-process architectures. ROMS has been used for applications from planetary scales down to local estuarine environments. Forecasts include predictions of water temperature, ocean currents, salinity, and sea surface height.
765.roms_r uses the generic BENCHMARK data set.
TITLE = Benchmark Test, Idealized Southern Ocean, Small Grid
The files used are:
varinfo.yaml Model description, used by all workloads.
Sets up over 600 variables.
roms_benchmarkN.in.x where N=[0123] representing test, train,
refrate, and refspeed.
The roms_benchmarkN.in.x files are read from stdin. They are preprocessed for the OpenMP threaded version to find an optimal and legal set of values for tiling of the I and J dimensions, resulting in files roms_benchmarkN.in
To create larger or smaller datasets the following parameters in the roms_benchmarkN.in.x files can be used. The example below is from the very short "test" workload:
Lm == 512 ! Number of I-direction INTERIOR RHO-points
Mm == 64 ! Number of J-direction INTERIOR RHO-points
N == 30 ! Number of vertical levels
NTIMES == 2 ! Total number time-steps in current run.
DT == 150.0d0 ! Time-Step size in seconds
The 3 first parameters affect both size (memory) and run time for the benchmark. The above settings for Lm, Mm, and N are used for both the test and train workloads, which consume about 372 MiB of memory on a Linux system tested by SPEC.
NTIMES affects only run time. The test workload does only 2 steps because the test workload is intended merely to verify that a valid binary has been built and that it can open its files. All the other workloads (train, refrate, refspeed) have adjusted for NTIMES according to SPEC's desired run times during pre-release testing.
For the refrate workload (which is used by the SPECrate version), these values are used, and memory increases to about 1.2 GiB on the same Linux system:
Lm == 1024 ! Number of I-direction INTERIOR RHO-points
Mm == 128 ! Number of J-direction INTERIOR RHO-points
N == 30 ! Number of vertical levels
The refspeed workload (for the SPECspeed version of the benchmark) uses the values shown below. When tested with a single thread it used about 22 GiB rss (resident size) and about the same vsz (virtual size). When many OpenMP threads were used, rss increased to about 24 GiB and vsz to about 30 GiB. Your memory needs will vary depending on your compiler, OpenMP library, number of threads, and your setting for environment variable OMP_STACKSIZE.
Lm == 8192 ! Number of I-direction INTERIOR RHO-points
Mm == 256 ! Number of J-direction INTERIOR RHO-points
N == 40 ! Number of vertical levels
Optional debugging information: SPEC builds include a -DNDEBUG option. If this is omitted the output file will include information on tiling, thread and memory usage information, for example:
Resolution, Grid 01: 512x64x30, Parallel Threads: 4, Tiling: 4x4
Tile partition information for Grid 01: 512x64x30 tiling: 4x4
tile Istr Iend Jstr Jend Npts
0 1 128 1 16 61440
1 129 256 1 16 61440
...
15 385 512 49 64 61440
Tile minimum and maximum fractional coordinates for Grid 01:
tile Xmin Xmax Ymin Ymax grid
0 -1.50 515.50 0.50 66.50 RHO-points
1 -2.50 515.50 0.50 66.50 RHO-points
...
15 -2.50 514.50 -0.50 65.50 RHO-points
0 -2.00 515.50 0.50 66.50 U-points
...
15 -2.50 514.00 -0.50 65.50 U-points
0 -1.50 515.50 0.00 66.50 V-points
...
15 -2.50 514.50 -0.50 65.00 V-points
Dynamic and Automatic memory (MB) usage for Grid 01: 512x64x30 tiling: 4x4
tile Dynamic Automatic USAGE
0 301.96 16.83 318.79
...
15 0.00 16.83 16.83
TOTAL 301.96 269.22 571.19
After each time step, various calculated energies and volume are printed. After the run, these are validated against a SPEC-supplied set of expected answers. Note that all of the values printed in the table after each step are calculated. The fact that 3 of the 4 values printed are the same on each step is an expected behavior.
The output is written to stdout, which is redirected to: roms_benchmarkN.log where N is again 0, 1, 2, or 3.
Fortran
The SPECrate version is single-threaded.
The SPECspeed version uses OpenMP. (Note: although the ROMS itself supports MPI, the SPEC CPU version does not use MPI.)
ROMS is highly portable. It is a modular code written in modern Fortran. It uses C-preprocessing (for SPEC CPU: specpp) to activate the various physical and numerical options. Several coding standards have been established to facilitate model readability, maintenance, and portability.
As mentioned in
system-requirements.html#memory, you may need to adjust your stack size(s).
For the SPECrate version 765.roms_r, Linux and Unix users may want to set
ulimit -s unlimited
and users of the Intel Compiler on Windows might need compiler options such as
/F1800000000
to adjust the main process stack size. For the SPECspeed version 865.roms_s,
you may need to adjust the main process stack (as shown just above); and you may also need
to adjust the stack size for child processes using OMP_STACKSIZE and
preenv.
The SPEC CPU 2017
FAQ item
for 627.cam4 may be helpful to understand how to set the stack sizes.
765.roms_r is based on https://github.com/myroms/roms tag: roms-4.1.
ROMS uses an MIT license, as found in License_ROMS.txt (link goes to the local copy in the benchmark Docs/ directory). In August 2023 the file was renamed, but accessing tag roms-4.1 allows it to be found via the original name License_ROMS.txt (GitHub copy).
The SPEC CPU 2026 benchmark 765.roms_r differs from the SPEC CPU 2017 benchmark versions 554.roms_r and 654.roms_s:
The CPU 2017 benchmark was based on ROMS 3.2 (2009); the CPU 2026 version is based on ROMS 4.1 (2023). Many algorithms have been refined or rewritten. The code has been updated to more modern and portable Fortran. It is more modular, making it easier to create new applications by changes to a minimal number of source files.
Workload sizes and run times have been modified to match the requirements for SPEC CPU 2026
Copyright © 2026 Standard Performance Evaluation Corporation (SPEC®)