This repository contains the raw experiment logs and analysis scripts for reproducing the results presented in Section 5.2 of the OSDI 2021 paper "Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning". To get started, we recommend using a virtualenv or conda environment:
$ conda create -n pollux python=3.8
$ conda activate pollux
$ python3 -m pip install seaborn dateutil
Decompress all results files:
$ unzip '*.zip'
The results can then be reproduced using the commands below.
- Pollux (p = -1):
python calc_jcts.py pollux-p-1-a - Optimus+Oracle+TunedJobs:
python calc_jcts.py optimus-tunedjobs - Tiresias+TunedJobs:
python calc_jcts.py tiresias-tunedjobs - Optimus+Oracle:
python calc_jcts.py optimus-realistic - Tiresias:
python calc_jcts.py tiresias-realistic - Pollux (p = +1):
python calc_jcts.py pollux-p1 - Pollux (p = -10):
python calc_jcts.py pollux-p-10
- Figure 5a:
python plot_cluster.py --value allocation pollux-p-1-a optimus-tunedjobs tiresias-tunedjobs - Figure 5b:
python plot_cluster.py --value efficiency pollux-p-1-a optimus-tunedjobs tiresias-tunedjobs
- Left Row 1:
python plot_imagenet.py --value jobs pollux-p-1-b - Left Row 2:
python plot_imagenet.py --value gpus pollux-p-1-b - Left Row 3:
python plot_imagenet.py --value bsz pollux-p-1-b - Left Row 4:
python plot_imagenet.py --value eff pollux-p-1-b - Right Row 1:
python plot_yolov3.py --value jobs pollux-p-1-a - Right Row 2:
python plot_yolov3.py --value gpus pollux-p-1-a - Right Row 3:
python plot_yolov3.py --value bsz pollux-p-1-a - Right Row 4:
python plot_yolov3.py --value eff pollux-p-1-a