Reproducibility Data for "High-performance in-situ ML Inference with dalotia: A Lightweight Tensor Loader API for Science Codes"

Reproducing the experiments

Get the experiment code. We used version 8356430aa5bc3ca84f4583a32421f425f9288cb2 of dalotia_evaluation with version beb9e03f762ccf475bfc7f23698a60df1e008513 of dalotia.

```
git clone https://github.com/RIKEN-RCCS/dalotia_evaluation
cd dalotia_evaluation
```

Put the `build.sh` and `experiments.sh` scripts into this same folder. Follow the prerequisite instructions in the build script (the Dockerfile may have additional hints on how to get the dependencies working). Execute the build script, and adapt where necessary.

Execute the experiments script:

```
# single-threaded:
for b in DeepRLEddy SubgridLES ; do for m in runtime memory energy ; do echo $b $m ; ./experiments.sh $m $b 1 &> out_${m}_${b}_t1.txt ; done ; done
# increasing the number of OpenMP threads:
for b in DeepRLEddy SubgridLES ; do for m in runtime memory energy ; do echo $b $m ; ./experiments.sh $m $b n &> out_${m}_${b}_tN.txt ; done ; done
```

This will generate separate results files for each graph in the paper.

Extract the numbers by the python script in the tools folder, for example:

```
python3 tools/extract_avg_values.py out_runtime_DeepRLEddy_tN.txt runtime
python3 tools/extract_avg_values.py out_energy_SubgridLES_t1.txt flops
```

This script will generate a csv file with the same file name as the input .txt file (except the .csv ending).
To generate the graphs, rename them to the expected file name and build the plot file, for example:

```
latexmk -pdf runtimeplot_cnn_parallel
```