Skip to content

TianyiChen0316/LOGER

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LOGER: A Learned Optimizer towards Generating Efficient and Robust Query Execution Plans

Requirements

Before running LOGER, please install PyTorch and Deep Graph Library (DGL) following the instructions on the pages, and run the command pip install -r requirements.txt to install psycopg2 and cx-Oracle. In addition, we use psqlparse to parse queries. Since we found that installing psqlparse with pip may result in errors, we recommend building it directly with setup.py. We provide a compiled version in our repository.

Running

To train LOGER with default settings, please use the following command.

python train.py

We provide a list of available arguments to change settings.

  • -D DATABASE: To specify the PostgreSQL database name. By default, we set the name to imdb.
  • -U USER: To specify the PostgreSQL user name. The default value is postgres.
  • -P PASSWORD: To specify the password of the PostgreSQL user. By default, the password is not used.
  • --port PORT: To specify the port of PostgreSQL.
  • -d TRAIN TEST: To train with the specified training and testing workloads. TRAIN is the folder of training workload, and TEST is for testing workload.
  • -e EPOCHS: To train with the specified number of epochs. The default value is 200.
  • -F FILE_ID: To change the output file ID of experiment results and checkpoints. The default value is 1.
  • -b BEAM_SIZE: To change the beam size of $\epsilon$-beam search. The default value is 4. To use $\epsilon$-greedy, please use a value less than 0.
  • -l NUM_LAYERS: To specify the number of Graph Transformer layers used in the query representation module. The default value is 1.
  • -w WEIGHT: To specify the weight factor of reward weighting. The default value is 0.1.
  • -N: To directly choose physical operators instead of using restricted operator.
  • --bushy: To allow bushy plan generation.
  • --no-exploration: To use the original beam search instead of $\epsilon$-beam search.
  • --no-expert-initialization: To forbid initializing the experience dataset with expert knowledge.
  • --seed VALUE: To set the random seed of training.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages