DeepMapping: Learned Data Mapping for Lossless Compression and Efficient Lookup

Resources for ICDE 2024 Submission

• DeepMapping: Learned Data Mapping for Lossless Compression and Efficient Lookup
  • Dataset
  • Model Searching
  • Benchmark
    • Task: Data Query
    • Task: Data Manipulation

2. DeepMapping is wrapped up as a Python library, please run the following command to install it.

   cd DeepMapping
   pip install -e ./
   

3. We wrapped up the feature extraction as a C-based function for better performance. Run the following command to compile it as a shared library.

   cc -fPIC -shared -o shared_utils.so shared_utils.c
   

Dataset

Our experiments covered synthetic datasets, low/high correlation datasets with different scales(100MB, 1GB, 10GB), and TPC-H, TPC-DS benchmark datasets with scale factors as 1 and 10. We removed all string/continuous columns and uploaded our pre-generated datasets to HERE.

After download it, please unzip it to the root folder of this GitHub repository. Then, you will see a dataset folder here.

List of datasets:

• TPC-H (S1/S10): customer, lineitem, orders, part, supplier.
• TPC-DS (S1/S10): catalog_page, catalog_returns, catalog_sales,customer_address, customer_demographics, customer, item, store_returns, web_returns.
• Synthetic Dataset (100MB, 1GB, 10GB): single_value_column_low_correlation, single_value_column_high_correlation, multiple_value_column_low_correlation, multiple_value_column_high_correlation.

Model Searching

1. Please run python run_search_model.py to perform a NAS with given dataset. You can configure the NAS by editing the run_search_model.py correspondingly. The searched result will be printout.

2. Modify the SEARCH_MODEL_STRUCTURE in run_train_searched_model.py with the output from step 1. And then run python run_train_searched_model.py to train a model.

Benchmark

We provided some demo models for the following 2 tasks. Please go HERE to download:

After download it, please unzip it to the root folder of this GitHub repository. Then, you will see a models folder here.

Note: to optimize the performance for each method, including baselines and DeepMapping. It is recommended to tune the hyperparameters in your local environment and use it. Run run_benchmark_tune.py to run a grid-search.

Task: Data Query

These experiments measured overall storage overhead and end-end query latency for benchmark datasets, i.e. TPC-H and TPC-DS. Run python run_benchmark_data_query.py to benchmark. To benchmark with different dataset, you should modify the file correspondingly by following the instructions provided in the python file.

Task: Data Manipulation

These experiments measured overall storage overhead and end-end query latency for synthetic dataset with data manipulation, i.e. INSERT/UPDATE/DELETE. Run python run_benchmark_data_manipulation.py to benchmark it. To benchmark with different dataset, you should modify the file correspondingly by following the instructions provided in the python file.