RL for performance optimization

The following sections represent the steps in order to successfully train the RL agent.

Generate the dataset

To generate the dataset, you use the script tiramisu_maker.py from the submodule RandomTiramisu at utils. Here are the options to pass the script and their default values:

  -h, --help            show this help message and exit
  --output-path OUTPUT_PATH [Default = "Dataset_multi"]
                        The location where to save the dataset.
  --first-seed FIRST_SEED  [Default = 105100]
                        The first seed to start generating the programs from. Every seed produces a program,and two different seeds produce two different programs.
  --nb-programs NB_PROGRAMS  [Default = 20000]
                        The number of programs to generate.

Configuring the repository

To configure the repository first copy the template files into the same location without the .template extension, like follows:

cp config.yaml.template config.yaml
cp scripts/env.sh.template scripts/env.sh
cp scripts/run_rllib_slurm.sh.template scripts/run_rllib_slurm.sh

Then, change the fields to their approriate value based on the environment the code is running on.

RL system configuration

All the configurations related to the RL system can be found in the file config.yml. The most important fields in this file, that the code will not run without setting are:

• ray.base_path: The absolute path to where the code is located on disk.
• environment.dataset_path: The path to the training dataset.
• tiramisu.tiramisu_path: The absolute path to the complete tiramisu installation, with the autoscheduler also installed.
• tiramisu.model_checkpoint: The path to the surrogate model used for prediction. A model chackpoint is provided under weights/surrogate_model.

The other flags can also be set for a desired purpose. For instance, you can change the policy model properties. You can also control whether or not you want to keep the dataset files after using them for an episode.

SLURM configuration

The file scripts/env.sh is used in order to configure the job that runs on SLURM. The most important fields in this file, that the code will not run without setting are:

• CONDA_DIR: The path to the conda installation.
• CONDA_ENV: The path to the specific conda environment. Other flags can be set in order to control the number of workers to use in training.

To get the conda environment location, if the desired conda is already activated, you follow these steps:

conda activate base
which python
# example: /home/user/miniconda3/bin/python

Then, you only keep the parts before bin/python and you set it to the variable CONDA_DIR in the scripts/env.sh file. You can follow the same steps, by activating the desired working environment and get the path to it using the which command without the bin/python part. This value is the stored to the variable CONDA_ENV.

The file scripts/run_rllib_slurm.sh can be configured to your needs in the file heading. The settings that needs to be configure are:

• The partition name on line 2.
• Number of needed nodes on line 3.
• The number of cpus per node on line 6, and
• The time limit for excution on line 7

When using SLURM, the outputs folder needs to be created:

mkdir outputs

Compilation commands

Depending on the c++ installation and the machine, the compilation command may differ. To change the compilation command, change the variables compile_tiramisu_cmd and compile_wrapper_cmd in the file utils/rl_autoscheduler_config.py. Examples of compilation commands:

# Command for compiling the Tiramisu autoschedule code when -lz is supported
c++ -I${TIRAMISU_ROOT}/3rdParty/Halide/include -I${TIRAMISU_ROOT}/include -I${TIRAMISU_ROOT}/3rdParty/isl/include  -Wl,--no-as-needed -ldl -g -fno-rtti -lz -lpthread -std=c++11 -O0 -o ${FILE_PATH}.o -c ${FILE_PATH};\
c++ -Wl,--no-as-needed -ldl -g -fno-rtti -lz -lpthread -std=c++11 -O0 ${FILE_PATH}.o -o ./${FILE_PATH}.out   -L${TIRAMISU_ROOT}/build  -L${TIRAMISU_ROOT}/3rdParty/Halide/lib  -L${TIRAMISU_ROOT}/3rdParty/isl/build/lib  -Wl,-rpath,${TIRAMISU_ROOT}/build:${TIRAMISU_ROOT}/3rdParty/Halide/lib:${TIRAMISU_ROOT}/3rdParty/isl/build/lib -ltiramisu -ltiramisu_auto_scheduler -lHalide -lisl 

# Command for compiling the Tiramisu autoschedule code when -lz is not supported
${CXX} -I${TIRAMISU_ROOT}/3rdParty/Halide/include -I${TIRAMISU_ROOT}/include -I${TIRAMISU_ROOT}/3rdParty/isl/include  -Wl,--no-as-needed -ldl -g -fno-rtti   -lpthread -std=c++11 -O0 -o ${FILE_PATH}.o -c ${FILE_PATH};\
${CXX} -Wl,--no-as-needed -ldl -g -fno-rtti   -lpthread -std=c++11 -O0 ${FILE_PATH}.o -o ./${FILE_PATH}.out   -L${TIRAMISU_ROOT}/build  -L${TIRAMISU_ROOT}/3rdParty/Halide/lib  -L${TIRAMISU_ROOT}/3rdParty/isl/build/lib  -Wl,-rpath,${TIRAMISU_ROOT}/build:${TIRAMISU_ROOT}/3rdParty/Halide/lib:${TIRAMISU_ROOT}/3rdParty/isl/build/lib -ltiramisu -ltiramisu_auto_scheduler -lHalide -lisl


# Command for compiling the Tiramisu wrapper code when -lz is supported
g++ -shared -o ${FUNC_NAME}.o.so ${FUNC_NAME}.o;\
g++ -std=c++11 -fno-rtti -I${TIRAMISU_ROOT}/include -I${TIRAMISU_ROOT}/3rdParty/Halide/include -I${TIRAMISU_ROOT}/3rdParty/isl/include/ -I${TIRAMISU_ROOT}/benchmarks -L${TIRAMISU_ROOT}/build -L${TIRAMISU_ROOT}/3rdParty/Halide/lib/ -L${TIRAMISU_ROOT}/3rdParty/isl/build/lib -o ${FUNC_NAME}_wrapper -ltiramisu -lHalide -ldl -lpthread -lz -lm -Wl,-rpath,${TIRAMISU_ROOT}/build ./${FUNC_NAME}_wrapper.cpp ./${FUNC_NAME}.o.so -ltiramisu -lHalide -ldl -lpthread -lz -lm

# Command for compiling the Tiramisu wrapper code when -lz is not supported
${GXX} -shared -o ${FUNC_NAME}.o.so ${FUNC_NAME}.o;\
${CXX} -I${TIRAMISU_ROOT}/3rdParty/Halide/include -I${TIRAMISU_ROOT}/include -I${TIRAMISU_ROOT}/3rdParty/isl/include -Wl,--no-as-needed -ldl -g -fno-rtti -lpthread -std=c++11 -O3 -o ${FUNC_NAME}_wrapper ${FUNC_NAME}_wrapper.cpp ./${FUNC_NAME}.o.so -L${TIRAMISU_ROOT}/build  -L${TIRAMISU_ROOT}/3rdParty/Halide/lib  -L${TIRAMISU_ROOT}/3rdParty/isl/build/lib  -Wl,-rpath,${TIRAMISU_ROOT}/build:${TIRAMISU_ROOT}/3rdParty/Halide/lib:${TIRAMISU_ROOT}/3rdParty/isl/build/lib -ltiramisu -ltiramisu_auto_scheduler -lHalide -lisl

Running the script

To run the script on one worker, you use the following command:

python train_ppo.py --num-workers 1

To run on slurm, you execute the following command:

sbatch scripts/run_rllib_slurm.sh

Visualization

We use tensorboard for visualization, to visualize the experiements, use the following command:

tensorboard dev upload --logdir ray_results/

Logs

Logs are available under outputs. The filename job.[job_id].out (where [job_id] is the id attributed to the job by SLURM) contains all the execution logs, while the file job.[job_id].err contains all the errors.