Experiment 1
First, compile the simulation on either Windows or Linux. To generate graphs, you will need python 3.10+ and be able to open .ipynb files. One method is to use Jupyter, or interface with the python files directly if not wanting to use the notebook.
The simulation has three modes: generating task sets, simulating task sets, and running multiple simulations simultaenously. Generating one task set and then simulating it is helpful when wanting to visualize only one simulation pass.
You may wish to fine-tune the parameters to optimally run on your machine. To do so, see gedf_sim.cpp and note that it creates two threads for every value of Theta. Importantly, see gedf_multisim.cpp and note the PARALLEL_SETS definition (by default, PARALLEL_SETS=5). When running multiple simulations, a PARALLEL_SETS number of gedf_sims are created, thus, the total number of simulation threads running concurrently will be at most 2 * |ThetaSet| * PARALLEL_SETS. When using the parameters given below, at most 40 threads will be running concurrently.
When running the binary without any arguments, the necessary parameters are printed out. To generate results similar to what were used by the paper, we recommend the following arguments:
.\ECRTS24ArtifactEval.exe -a -Mmin 4 -Mmax 16 -Mstep 4 -tminset 3 10 50 -tmaxset 33 100 200 -n 150 -taskSets 1000 -utilMin 0.2 -utilMax 0.9 -utilStep 0.1 -Hmin 8 -Hmax 64 -Hstep 8 -Tmin 1.5 -Tmax 3 -Tstep 0.5 -p 0.5 -o p0.5.csv
These arguments will generate 1000 task sets for every combination of U, M, and Tmin/Tmax. These task sets are then run for every combination of H and Theta. If you do not have enough time to simulate a large set of data, then consider reducing -taskSets 1000 to a lower value, or download this csv of results for use in graph generation.
Note: the generated csv file will contain millions of lines, which not all csv file viewers may be able to render.
Once you have generated results, this python notebook can parse the data and generate graphs.
Repeat the above experiment with different values for p (such as 0.1 and 0.25).
The left graph is when H=16, M=8, Theta=2.5, Tmin=10, Tmax=100, p=0.5. The right graph uses the same parameters but for a fixed value of U=0.6 across the shown values of H.
You can compare your generated results to the graph above.
Note how the pi-blocking for the OMLP increases dramatically as system utilization increases (where more jobs are issuing requests). The pi-blocking observed by the SMLP also increases, but because the SMLP only uses as many SMs as needed to optimally run the GPU kernel, the effect is significantly mitigated.
Note: if you used a small number of taskSets or small values for n, then the graph might not be as smooth.
