/Lucid

Lucid: A Non-Intrusive, Scalable and Interpretable Scheduler for Deep Learning Training Jobs

Primary LanguagePythonOtherNOASSERTION

Artifact for Lucid

DOI

This repository contains the artifact for our ASPLOS '23 paper "Lucid: A Non-Intrusive, Scalable and Interpretable Scheduler for Deep Learning Training Jobs". It includes following parts:

  • simulation: It contains code and data for reproducing key results in our paper.

  • workloads: The Pytorch implementation of 14 different workloads used in experiments.

  • profile: It contains the code to collect traces of each training job type.

Getting Started

Results Reproduction (for ASPLOS '23 Artifact Evaluation)

simulation (adopted from Helios) contains instructions for reproducing the Venus cluster experiments shown in Section 4. These scripts have been tested on Ubuntu 20.04 with Python 3.9.

0. Structure

The contents inside simulation folder are summarized as follows:

  • data/ contains Venus cluster job trace and cluster configuration used for evaluation.
  • analyzer/ contains the Packing Analyze Model and profiled workloads information used in our experiment.
  • estimator/ contains the Workload Estimate Model and job duration estimation for both Lucid and QSSF.
  • plot/ contains notebook for visualizing experiment results.
  • policy/ contains implementations of the Lucid scheduling policy, and baseline policies including FIFO, SJF, QSSF, Tiresias.
  • predictor/ contains the Throughput Predict Model and cluster throughput estimation in Venus September.
  • profiler/ contains the Least-GPU-First and Auto-Scaling Profiler implementation for Lucid.
  • cluster.py, job.py and updater.py contain implementations of the GPU cluster and workload logic.
  • simulator.py is the main entry of the simulator.

1. Environment Preparation

We suggest using a conda environment to install the dependencies:

conda create -n lucid python=3.9
conda activate lucid
cd simulation
pip install -r requirements.txt

Besides, we recommend execute Jupyter notebook (.ipynb) files with VSCode or JupyterLab (conda install jupyterlab).

2. Lucid Model Training and Interpretation

We train Throughput Predict Model as a reproduction example. Please follow below steps:

  • Enter predictor folder and open predictor.ipynb file

  • Run all cells inside the notebook. It contains the interpretable model (Primo EBM) used in Lucid and other ML baselines (LightGBM, XGBoost, Random Forest, DNN).

  • Table 7: Interpretable Model Performance: Check Result Comparison cell, the MAE scores of all baselines are listed.

  • Figure 13 (a): Throughput Predict Performance: Check Prediction Visualization cell (or Venus_throughput.pdf output file), both the real and predicted throughput are plotted. Generated figures should have similar patterns as the paper. The difference is because we release the Venus Job throughput prediction code but we plot Saturn Job throughput prediction in our paper.

  • Figure 7 (a)(b): Global Model Interpretation and Learned Shape Function: Check Model Interpretation cell (or interpret_Venus_throughput.pdf & interpret_Venus_shapefunc.pdf output files). Generated figures should have similar patterns as the paper. The difference is because we release the Venus Job throughput prediction code but we plot Saturn GPU throughput prediction in our paper.

More model training codes are also provided (estimator/estimator_lucid.ipynb and analyzer/analyzer.py).

3. Reproduce Baseline Results

Use the following command to run all baselines simultaneously

cd simulation
python simulator.py --sweep

The output of this script looks like this:

2022 Oct 08 14:32:57 | MainProcess | Total Job Number in Cluster Training: 23859
2022 Oct 08 14:32:59 | ForkPoolWorker-1 | vcEwI | Time: 13220000 | Total Job: 7603 | End job: 13 | Running job: 2 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-2 | vcWoR | Time: 13220000 | Total Job: 2826 | End job: 0 | Running job: 0 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-1 | vcEwI | Time: 13230000 | Total Job: 7603 | End job: 120 | Running job: 4 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-2 | vcWoR | Time: 13230000 | Total Job: 2826 | End job: 0 | Running job: 1 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-1 | vcEwI | Time: 13240000 | Total Job: 7603 | End job: 120 | Running job: 4 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-3 | vcHvQ | Time: 13220000 | Total Job: 2654 | End job: 1 | Running job: 1 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-2 | vcWoR | Time: 13240000 | Total Job: 2826 | End job: 0 | Running job: 1 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-1 | vcEwI | Time: 13250000 | Total Job: 7603 | End job: 121 | Running job: 4 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-4 | vcvGl | Time: 13220000 | Total Job: 1452 | End job: 0 | Running job: 0 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-2 | vcWoR | Time: 13250000 | Total Job: 2826 | End job: 0 | Running job: 2 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-3 | vcHvQ | Time: 13230000 | Total Job: 2654 | End job: 2 | Running job: 0 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-1 | vcEwI | Time: 13260000 | Total Job: 7603 | End job: 162 | Running job: 9 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-5 | vc8Gr | Time: 13220000 | Total Job: 710 | End job: 0 | Running job: 0 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-4 | vcvGl | Time: 13230000 | Total Job: 1452 | End job: 1 | Running job: 2 | Pending job: 0
2022 Oct 08 14:32:59 | ForkPoolWorker-5 | vc8Gr | Time: 13230000 | Total Job: 710 | End job: 0 | Running job: 1 | Pending job: 0

4. Reproduce Lucid Results

Similarly, use the following command to run all baselines simultaneously

python simulator.py -s lucid

The output of this script looks like this:

2022 Oct 08 14:45:07 | MainProcess | Total Job Number in Cluster Training: 23859
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13220000 | Total Job: 23859 | End job: 17 | Running job: 1 | Pending job: 0 | Avail Nodes: 2
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13230000 | Total Job: 23859 | End job: 134 | Running job: 0 | Pending job: 0 | Avail Nodes: 2
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13240000 | Total Job: 23859 | End job: 134 | Running job: 0 | Pending job: 0 | Avail Nodes: 2
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13250000 | Total Job: 23859 | End job: 136 | Running job: 0 | Pending job: 0 | Avail Nodes: 2
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13260000 | Total Job: 23859 | End job: 249 | Running job: 3 | Pending job: 4 | Avail Nodes: 1
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13270000 | Total Job: 23859 | End job: 385 | Running job: 3 | Pending job: 2 | Avail Nodes: 1
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13280000 | Total Job: 23859 | End job: 589 | Running job: 2 | Pending job: 0 | Avail Nodes: 1
2022 Oct 08 14:45:08 | MainProcess | profvc | Time: 13290000 | Total Job: 23859 | End job: 780 | Running job: 2 | Pending job: 0 | Avail Nodes: 2

After the program is executed, you can check the result in the log folder. The job log and time sequence of each VC are provided separately.

5. Visualize the Key Results

We provide simulation analysis and plot scripts to generate the figures shown in our paper. Please follow below steps:

  • Enter plot folder and open result_plot.ipynb file

  • Run all cells inside the notebook.

  • Table 4: Scheduling Performance: Check Table 4: Result Summary cell (or result_summary.csv output file), the Average JCT, Average Queuing Delay and Queuing Delay 99.9 Quantile of all policies are listed.

  • Table 5: Scheduling Performance (workload analysis): Check Table 5: Result Summary of Different Scales of Workloads cell, the Average JCT, Average Queuing Delay of large and small jobs are listed.

  • Figure 8: CDF of JCT: Check Plot Result 8: JCT cell (or result_cdf_jct.pdf output file), JCT CDF of all policies are plotted.

  • Figure 9: Queue Time in each VC: Check Plot Result 9: Queue Time in each VC cell (or result_bar_queue.pdf output file), queuing delay of all policies are plotted.

Workloads Profiling

This part profile contains code for profiling metrics of multiple workloads.

Directory

Note that ./result/ will be created when main_co.py or main_single.py is launched.

Basic Usage

Run main_co.py will generate the colocated jobs' metrics under ./result/colocate. Run main_single.py will generate single jobs' metrics under ./result/. Some specific settings can be set in each workload's profiling file, e.g.profile_cifar.py. The output will be like this:

imagenet + imagenet
co-locate:
==> Training mobilenet_v3_small model with 32 batchsize, 0 mp..
==> Training mobilenet_v3_small model with 32 batchsize, 0 mp..
co-locate:
==> Training mobilenet_v3_small model with 32 batchsize, 0 mp..
==> Training mobilenet_v3_small model with 32 batchsize, 1 mp..
co-locate:
==> Training mobilenet_v3_small model with 32 batchsize, 1 mp..
==> Training mobilenet_v3_small model with 32 batchsize, 1 mp..
imagenet + cifar10
co-locate:
Files already downloaded and verified
==> Training ResNet18 model with 32 batchsize, 0 mp..
==> Training mobilenet_v3_small model with 32 batchsize, 0 mp..
...

Datasets

The data path storing all datasets is specified in ./workloads/settings.py as data_dir. You can also specify the total runtime of some workloads by changing total_runtime.

  • CIFAR-10: The cifar10 dataset will be downloaded automatically(if not exist) when ./workloads/cifar/profile_cifar.py is run.

  • ImageNet: The dataset is generated automatically in ./workloads/imagenet/profile_imagenet.py.

  • LSUN: The dataset is generated automatically in ./workloads/dcgan/profile_dcgan.py. You can change the custom image size of generated data via --imageSize. The default value is 64.

  • ShapeNet: Use the following command to download dataset under directory data_dir/shapenetcore/:

    wget https://shapenet.cs.stanford.edu/ericyi/shapenetcore_partanno_segmentation_benchmark_v0.zip --no-check-certificate
unzip shapenetcore_partanno_segmentation_benchmark_v0.zipcollect_metric/workloadspointnet.pytorch.

  • SQuAD: The data can be downloaded with the following link and should be saved under data_dir/SQUAD_DIR/ directory.

    train-v1.1.json

  • Wikitext2: The dataset can be downloaded from

    wikitext-2

    File test.txt, train.txt and valid.txt should be saved in data_dir/wikitext-2/ directory.

  • Multi30k: First download the Moses tokenizer(http://www.statmt.org/moses/) for data preparation:

    wget https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts/tokenizer/tokenizer.perl
wget https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts/share/nonbreaking_prefixes/nonbreaking_prefix.de
wget https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts/share/nonbreaking_prefixes/nonbreaking_prefix.en
sed -i "s/$RealBin\/..\/share\/nonbreaking_prefixes//" tokenizer.perl
wget https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts/generic/multi-bleu.perl

    These files should be downloaded in ./workloads/translation/.

    Then download data in data_dir/multi30k/:

    mkdir -p data/multi30k
wget http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/training.tar.gz &&  tar -xf training.tar.gz -C data/multi30k && rm training.tar.gz
wget http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/validation.tar.gz && tar -xf validation.tar.gz -C data/multi30k && rm validation.tar.gz
wget http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/mmt16_task1_test.tar.gz && tar -xf mmt16_task1_test.tar.gz -C data/multi30k && rm mmt16_task1_test.tar.gz

    Preprocess the data:

    for l in en de; do for f in ~/data/multi30k/*.$l; do if [[ "$f" != *"test"* ]]; then sed -i "$ d" $f; fi;  done; done
for l in en de; do for f in ~/data/multi30k/*.$l; do perl tokenizer.perl -a -no-escape -l $l -q  < $f > $f.atok; done; done
python preprocess.py -train_src ~/data/multi30k/train.en.atok -train_tgt ~/data/multi30k/train.de.atok -valid_src ~/data/multi30k/val.en.atok -valid_tgt ~/data/multi30k/val.de.atok -save_data ~/data/multi30k.atok.low.pt

    Referenced from: https://github.com/Eathoublu/attention-is-all-you-need-pytorch.

  • MovieLens: Use the following command to download the dataset in data_dir/ml-1m/:

    wget https://github.com/hexiangnan/neural_collaborative_filtering/raw/master/Data/ml-1m.test.negative
wget https://github.com/hexiangnan/neural_collaborative_filtering/raw/master/Data/ml-1m.test.rating
wget https://github.com/hexiangnan/neural_collaborative_filtering/raw/master/Data/ml-1m.train.rating