Optimizing Parquet String Data Compression with RAPIDS

Parquet writers offer various encoding and compression options that are turned off by default. Enabling these options can provide better lossless compression for your data, but understanding which options to use is crucial for optimal performance, according to the NVIDIA Technical Blog.

Understanding Parquet Encoding and Compression

Parquet's encoding step reorganizes data to reduce its size while preserving access to each data point. The compression step further reduces the total size in bytes but requires decompression before accessing the data again. The Parquet format includes two delta encodings designed to optimize string data storage: DELTA_LENGTH_BYTE_ARRAY (DLBA) and DELTA_BYTE_ARRAY (DBA).

RAPIDS libcudf and cudf.pandas

RAPIDS is a suite of open-source accelerated data science libraries. In this context, libcudf is the CUDA C++ library for columnar data processing. It supports GPU-accelerated readers, writers, relational algebra functions, and column transformations. The Python cudf.pandas library accelerates existing pandas code by up to 150x.

Benchmarking with Kaggle String Data

A dataset of 149 string columns, comprising 4.6 GB total file size and 12 billion total character count, was used to compare encoding and compression methods. The study found less than 1% difference in encoded size between libcudf and arrow-cpp and a 3-8% increase in file size when using the ZSTD implementation in nvCOMP 3.0.6 compared to libzstd 1.4.8+dfsg-3build1.

String Encodings in Parquet

String data in Parquet is represented using the byte array physical type. Most writers default to RLE_DICTIONARY encoding for string data, which uses a dictionary page to map string values to integers. If the dictionary page grows too large, the writer falls back to PLAIN encoding.

Total File Size by Encoding and Compression

For the 149 string columns in the dataset, the default setting of dictionary encoding and SNAPPY compression yields a total 4.6 GB file size. ZSTD compression outperforms SNAPPY, and both outperform uncompressed options. The best single setting for the dataset is default-ZSTD, with further reductions possible using delta encoding for specific conditions.

When to Choose Delta Encoding

Delta encoding is beneficial for data with high cardinality or long string lengths, generally achieving smaller file sizes. For string columns with less than 50 characters, DBA encoding can provide significant file size reductions, especially for sorted or semi-sorted data.

Reader and Writer Performance

The GPU-accelerated cudf.pandas library showed impressive performance compared to pandas, with 17-25x faster Parquet read speeds. Using cudf.pandas with an RMM pool further improved throughput to 552 MB/s read and 263 MB/s write speeds.

Conclusion

RAPIDS libcudf offers flexible, GPU-accelerated tools for reading and writing columnar data in formats such as Parquet, ORC, JSON, and CSV. For those looking to leverage GPU acceleration for Parquet processing, RAPIDS cudf.pandas and libcudf provide significant performance benefits.