Merzbild.jl benchmarks
Various benchmarks and comparisons to other open-source codes are provided here for reference test cases.
Couette flow, serial, small grid
Comparison with SPARTA are provided for a single-species (argon) Couette flow test case with 50000 particles and 50 cells (averaging over 36k timesteps after t>14000). The computation is serial. Timing in Merzbild.jl providedd by TimerOutputs.jl, timing in SPARTA provided by the inbuilt timers. No surface quantities are being computed. The input can be found in simulations/1D/couette_benchmarking.jl.
Merzbild.jl version 0.7.0, run with --check-bounds=no -O3.
SPARTA version 20Jan2025, compiled with -O3.
Intel Core i9-13900K, 128 GB RAM
Ubuntu 22.04.5, Julia version 1.11.2, SPARTA compiled with gcc version 11.4.0.
Merzbild.jl
──────────────────────────────────────────────────────────────────────────
Time Allocations
─────────────────────── ────────────────────────
Tot / % measured: 28.4s / 97.3% 126MiB / 9.8%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────
sort 50.0k 10.5s 37.9% 209μs 0.00B 0.0% 0.00B
convect 50.0k 6.29s 22.8% 126μs 0.00B 0.0% 0.00B
collide 2.50M 5.90s 21.3% 2.36μs 0.00B 0.0% 0.00B
props compute 36.0k 4.88s 17.7% 136μs 0.00B 0.0% 0.00B
I/O 1 85.4ms 0.3% 85.4ms 9.32MiB 75.3% 9.32MiB
avg physprops 36.0k 5.54ms 0.0% 154ns 0.00B 0.0% 0.00B
sampling 1 2.52ms 0.0% 2.52ms 3.05MiB 24.7% 3.05MiB
──────────────────────────────────────────────────────────────────────────SPARTA
Loop time of 30.4417 on 1 procs for 50000 steps with 50000 particles
MPI task timing breakdown:
Section | min time | avg time | max time |%varavg| %total
---------------------------------------------------------------
Move | 8.0618 | 8.0618 | 8.0618 | 0.0 | 26.48
Coll | 10.862 | 10.862 | 10.862 | 0.0 | 35.68
Sort | 2.793 | 2.793 | 2.793 | 0.0 | 9.18
Comm | 0.0034328 | 0.0034328 | 0.0034328 | 0.0 | 0.01
Modify | 8.719 | 8.719 | 8.719 | 0.0 | 28.64
Output | 0.00057459 | 0.00057459 | 0.00057459 | 0.0 | 0.00
Other | | 0.002392 | | | 0.01M1 Pro (Macbook Pro), 32 GB RAM
MacOS 15.4.1, Julia version 1.11.2, SPARTA compiled with Apple clang version 17.0.0.
Merzbild.jl
──────────────────────────────────────────────────────────────────────────
Time Allocations
─────────────────────── ────────────────────────
Tot / % measured: 33.5s / 97.7% 125MiB / 9.9%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────
sort 50.0k 11.6s 35.5% 232μs 0.00B 0.0% 0.00B
collide 2.50M 8.01s 24.4% 3.20μs 0.00B 0.0% 0.00B
convect 50.0k 7.44s 22.7% 149μs 0.00B 0.0% 0.00B
props compute 36.0k 5.60s 17.1% 156μs 0.00B 0.0% 0.00B
I/O 1 91.8ms 0.3% 91.8ms 9.30MiB 75.3% 9.30MiB
avg physprops 36.0k 5.33ms 0.0% 148ns 0.00B 0.0% 0.00B
sampling 1 2.78ms 0.0% 2.78ms 3.05MiB 24.7% 3.05MiB
──────────────────────────────────────────────────────────────────────────SPARTA
Loop time of 47.4825 on 1 procs for 50000 steps with 50000 particles
MPI task timing breakdown:
Section | min time | avg time | max time |%varavg| %total
---------------------------------------------------------------
Move | 12.001 | 12.001 | 12.001 | 0.0 | 25.28
Coll | 13.299 | 13.299 | 13.299 | 0.0 | 28.01
Sort | 2.8246 | 2.8246 | 2.8246 | 0.0 | 5.95
Comm | 0.0020463 | 0.0020463 | 0.0020463 | 0.0 | 0.00
Modify | 19.352 | 19.352 | 19.352 | 0.0 | 40.76
Output | 0.0019748 | 0.0019748 | 0.0019748 | 0.0 | 0.00
Other | | 0.0009062 | | | 0.00Couette flow, serial, large grid
The physical parameters for this test case are the same as for the previous one, but a larger (2000 cells) grid is used, with 250 particles per cell at t=0. So the number of grid cells is 40x higher than for the small grid test case, and the number of particles is 10x higher.
In addition, surface properties are also computed and averaged. The numerical setup corresponds to the simulations/1D/couette_with_surface_quantities.jl file with the following parameters parameters for the run command:
run(1234, 300.0, 500.0, 5e-4, 5e22, 2000, 250, 2.59e-9, 1000, 50000, 14000; do_benchmark=true)
Setting do_benchmark to true turns off computation of the degree of particle index fragmentation.
Intel Core i9-13900K, 128 GB RAM
Ubuntu 22.04.5, Julia version 1.11.2, SPARTA compiled with gcc version 11.4.0.
Merzbild.jl
──────────────────────────────────────────────────────────────────────────────────────
Time Allocations
─────────────────────── ────────────────────────
Tot / % measured: 795s / 99.5% 192MiB / 17.1%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────────────────
sort 50.0k 269s 34.0% 5.37ms 0.00B 0.0% 0.00B
convect + surface compute 36.0k 209s 26.4% 5.80ms 2.20MiB 6.7% 64.0B
props compute 36.0k 131s 16.6% 3.63ms 0.00B 0.0% 0.00B
collide 100M 114s 14.4% 1.14μs 0.00B 0.0% 0.00B
convect 14.0k 68.1s 8.6% 4.86ms 0.00B 0.0% 0.00B
avg physprops 36.0k 196ms 0.0% 5.43μs 0.00B 0.0% 0.00B
sampling 1 35.1ms 0.0% 35.1ms 30.5MiB 93.3% 30.5MiB
avg surfprops 36.0k 12.6ms 0.0% 351ns 0.00B 0.0% 0.00B
I/O 15 2.26ms 0.0% 151μs 3.58KiB 0.0% 244B
──────────────────────────────────────────────────────────────────────────────────────SPARTA
Loop time of 1155.34 on 1 procs for 50000 steps with 500000 particles
MPI task timing breakdown:
Section | min time | avg time | max time |%varavg| %total
---------------------------------------------------------------
Move | 598.55 | 598.55 | 598.55 | 0.0 | 51.81
Coll | 262.52 | 262.52 | 262.52 | 0.0 | 22.72
Sort | 87.103 | 87.103 | 87.103 | 0.0 | 7.54
Comm | 0.018935 | 0.018935 | 0.018935 | 0.0 | 0.00
Modify | 207.14 | 207.14 | 207.14 | 0.0 | 17.93
Output | 0.002845 | 0.002845 | 0.002845 | 0.0 | 0.00
Other | | 0.01218 | | | 0.00M1 Pro (Macbook Pro), 32 GB RAM
MacOS 15.4.1, Julia version 1.11.2, SPARTA compiled with Apple clang version 17.0.0.
Merzbild.jl
──────────────────────────────────────────────────────────────────────────────────────
Time Allocations
─────────────────────── ────────────────────────
Tot / % measured: 1164s / 99.5% 190MiB / 17.2%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────────────────
convect + surface compute 36.0k 359s 31.0% 10.0ms 2.20MiB 6.7% 64.0B
sort 50.0k 337s 29.0% 6.73ms 0.00B 0.0% 0.00B
props compute 36.0k 181s 15.6% 5.02ms 0.00B 0.0% 0.00B
collide 100M 171s 14.8% 1.71μs 0.00B 0.0% 0.00B
convect 14.0k 110s 9.5% 7.87ms 0.00B 0.0% 0.00B
avg physprops 36.0k 198ms 0.0% 5.51μs 0.00B 0.0% 0.00B
sampling 1 26.1ms 0.0% 26.1ms 30.5MiB 93.3% 30.5MiB
avg surfprops 36.0k 14.8ms 0.0% 410ns 0.00B 0.0% 0.00B
I/O 15 1.77ms 0.0% 118μs 3.58KiB 0.0% 244B
──────────────────────────────────────────────────────────────────────────────────────SPARTA
Loop time of 1418.29 on 1 procs for 50000 steps with 500000 particles
MPI task timing breakdown:
Section | min time | avg time | max time |%varavg| %total
---------------------------------------------------------------
Move | 885.53 | 885.53 | 885.53 | 0.0 | 62.44
Coll | 274.42 | 274.42 | 274.42 | 0.0 | 19.35
Sort | 45.409 | 45.409 | 45.409 | 0.0 | 3.20
Comm | 0.0055656 | 0.0055656 | 0.0055656 | 0.0 | 0.00
Modify | 212.91 | 212.91 | 212.91 | 0.0 | 15.01
Output | 0.0032787 | 0.0032787 | 0.0032787 | 0.0 | 0.00
Other | | 0.006827 | | | 0.00AMD EPYC 9374F, 378 GB RAM
Ubuntu 24.04.3, Julia version 1.11.6.
Merzbild.jl
──────────────────────────────────────────────────────────────────────────────────────
Time Allocations
─────────────────────── ────────────────────────
Tot / % measured: 1467s / 99.5% 192MiB / 17.0%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────────────────
convect + surface compute 36.0k 584s 40.0% 16.2ms 2.20MiB 6.7% 64.0B
sort 50.0k 369s 25.3% 7.39ms 0.00B 0.0% 0.00B
collide 100M 173s 11.8% 1.73μs 0.00B 0.0% 0.00B
convect 14.0k 170s 11.6% 12.1ms 0.00B 0.0% 0.00B
props compute 36.0k 164s 11.2% 4.54ms 0.00B 0.0% 0.00B
avg physprops 36.0k 306ms 0.0% 8.51μs 0.00B 0.0% 0.00B
sampling 1 31.4ms 0.0% 31.4ms 30.5MiB 93.3% 30.5MiB
avg surfprops 36.0k 21.8ms 0.0% 605ns 0.00B 0.0% 0.00B
I/O 15 7.44ms 0.0% 496μs 3.58KiB 0.0% 244B
──────────────────────────────────────────────────────────────────────────────────────Couette flow, multi-threaded, large grid
The numerical and physical parameters are the same as for the serial large grid case (2000 cells, 250 particles per cell at t=0). The simulation file is simulations/1D/couette_multithreaded.jl.
Intel Core i9-13900K, 128 GB RAM
Ubuntu 22.04.5, Julia version 1.11.2. Shown is the speed-up compared to a serial execution on the same computer (see above). DLB denotes dynamic load balancing (currently not used).
| 2 cores | 4 cores | 8 cores | |
|---|---|---|---|
n_chunks=n_threads, no DLB | 1.85 | 3.24 | 5.06 |
M1 Pro (Macbook Pro), 32 GB RAM
MacOS 15.4.1, Julia version 1.11.2. Shown is the speed-up compared to a serial execution on the same computer (see above). DLB denotes dynamic load balancing (currently not used).
| 2 cores | 4 cores | 8 cores | |
|---|---|---|---|
n_chunks=n_threads, no DLB | 2.27 | 3.75 | 5.76 |
AMD EPYC 9374F, 378 GB RAM
Ubuntu 24.04.3, Julia version 1.11.6. Shown is the speed-up compared to a serial execution on the same computer (see above). DLB denotes dynamic load balancing (currently not used).
| 2 cores | 4 cores | 8 cores | 16 cores | **32 cores ** | |
|---|---|---|---|---|---|
n_chunks=n_threads, no DLB | 1.88 | 3.68 | 5.24 | 5.70 | 6.57 |