C3 Distributed

This serves as a codebase for the distributed version of C3 based on Orca.

1. Create a cloudlab job with 17 nodes (16 nodes for actors, 1 node for learner) using the orca profile (which will have the linux-learner image preinstalled once the nodes start)'
2. Run ./cloudlab/config.sh to setup the nodes (run this from YOUR machine).
3. The next thing you will have to do is ./cloudlab/setup_params.sh to generate rl-module/params_distributed.json. (run this on node0 on cloudlab)
4. Once this is done, you should use v9_multi_train.sh to start the training job - do this within a tmux session. (run this on cloudlab node0)
5. ssh into the actor nodes and look at ~/actor_logs/ to see the stdout (and stderr) of each actors. Look inside ~/ConstrainedOrca/rl-module/training_log for the train log.
6. Once training is done, use scripts/collate_train_files.sh to collect everything into one place. Backup the folder generated by this for the future.

Eval

1. Move the checkpoint you desire into ~/ConstrainedOrca/rl-module/train_dir/seed0/. When you do ls inside this seed0, it should show you one directory that looks something like learner0-v9_actorNum256_multi_lambda0.0_ksymbolic5_k1_raw-sym_threshold25_seed0/.
2. Run ./scripts/eval_orca.sh <model_name> <trace_dir> <results_dir> <start_run> <end_run> <constraints_id>.
3. <trace_dir> is /proj/verifiedmlsys-PG0/ConstrainedOrca/sage_traces/traces for SAGE traces.
4. <results_dir> is /proj/verifiedmlsys-PG0/sigcomm_results/new_result_dir/constraint_id_*.

Plotting and analysis

0. Use cd scripts && ./process_down_file.sh to trim stuff.
1. Use ./scripts/plots/plot_thr.py for motivation figures
2. Use ./scripts/plots/plot_thr_delay.py for thr vs delay plots.

Other useful info

1. baseline_v4 is the baseline used for NSDI submission.