BDGS: A Scalable Big Data Generator Suite in Big Data Benchmarking

Zijian Ming, Chunjie Luo, Wanling Gao, Rui Han, Qiang Yang, Lei Wang, Jianfeng Zhan

TL;DR

BDGS (Big Data Generator Suite) is a comprehensive tool designed to create massive, synthetic datasets that mimic the "4V" characteristics of real-world big data: Volume, Variety, Velocity, and Veracity. By using sophisticated statistical models like Latent Dirichlet Allocation (LDA) and Kronecker graphs, it allows researchers to generate petabyte-scale data that preserves the complex internal structures of real-world information for more accurate system benchmarking.

What Problem Does This Paper Address?

Benchmarking big data systems (like Hadoop, NoSQL databases, or Spark) requires massive amounts of data. However, researchers face three major hurdles:

  1. Privacy and Confidentiality: Companies are often unwilling to share real-world datasets.
  2. Logistics: Transferring terabytes or petabytes of real data over the internet is prohibitively expensive and slow.
  3. The "Veracity" Gap: Simple random data generation doesn't reflect the complexity of real data. If a benchmark uses data that is "too simple," the performance results won't translate to real-world applications.

Existing tools often focus on a single data type (like structured tables) or a single platform. The authors introduce BDGS to provide a unified, scalable way to generate structured, semi-structured, and unstructured data that actually "looks" like the real thing.

Methods

The BDGS methodology follows a four-step pipeline: Selection, Processing (Modeling), Generation, and Format Conversion. The core innovation lies in how it models different data types to ensure "veracity":

  • Text Generation: Instead of random strings, BDGS uses Latent Dirichlet Allocation (LDA). It trains on real text (like Wikipedia) to learn topic and word distributions, then generates new documents that maintain those realistic statistical properties.
  • Graph Generation: To simulate social networks or web graphs, it employs the Kronecker graph model. This captures essential real-world graph patterns, such as the "Power Law" distribution of connections and the "Shrinking Diameter" effect (where graphs become "smaller" as they grow).
  • Table/Semi-structured Generation: For traditional databases, it uses the Parallel Data Generation Framework (PDGF). For NoSQL-style semi-structured data (like resumes), it uses a Bernoulli distribution to model the presence or absence of optional fields, reflecting the "schema-less" nature of modern data.

Datasets

BDGS uses six representative real-world datasets as "seeds" to build its generation models:

  1. Wikipedia Entries: Unstructured text (4.3 million articles).
  2. Amazon Movie Reviews: Semi-structured text with metadata (7.9 million reviews).
  3. Google Web Graph: Unstructured graph data (875k nodes, 5.1M edges).
  4. Facebook Social Graph: Unstructured social network data (4k nodes, 88k edges).
  5. E-commerce Transactions: Structured relational table data.
  6. Personal Resumes: Semi-structured, schema-less data (278k resumes).

Key Insights and Ideas

  • Veracity is the "New" Requirement: The authors argue that volume isn't enough. If the synthetic data doesn't have the same "shape" (distribution, sparsity, connectivity) as real data, the benchmark is effectively useless.
  • Model-Based Scaling: Rather than just duplicating data, BDGS scales by learning the underlying parameters of a model. This allows it to generate a 100GB file that has the same statistical "DNA" as a 1GB sample.
  • Linear Scalability: A critical design goal was ensuring that generation time grows linearly with data volume. If it takes exponentially longer to generate more data, the tool wouldn't be useful for petabyte-scale testing.

Key Results

  • Consistent Performance: The generators maintain a roughly constant rate regardless of volume. For example, text generation from the Wikipedia model averaged 63.23 MB/s, while Amazon movie reviews reached 71.3 MB/s.
  • Efficiency: The authors demonstrated that generating 1 TB of Wikipedia-like text takes approximately 4.7 hours on a single node.
  • Graph Scaling: The graph generator successfully scaled up to $2^{20}$ edges while maintaining the structural characteristics of the input graphs.
  • Linear Time: Experimental plots show a near-perfect linear relationship between data volume (GB) and generation time (minutes), proving the suite's scalability.

Limitations and Future Work

  • Memory Bottlenecks: The authors noted that the Graph Generator is particularly memory-intensive because it needs to compute large maps in-memory. This was identified as a primary performance bottleneck.
  • Single-Node Testing: While the architecture is designed for parallel deployment, the results in this paper were primarily gathered on a single node.
  • Future Directions: The team plans to implement a fully parallelized version of BDGS and expand the suite to include multimedia data (images, video, audio). They also intend to develop more rigorous metrics to evaluate exactly how well the synthetic data matches the statistical characteristics of the original "seed" data.

Summary updated

Abstract

Data generation is a key issue in big data benchmarking that aims to generate application-specific data sets to meet the 4V requirements of big data. Specifically, big data generators need to generate scalable data (Volume) of different types (Variety) under controllable generation rates (Velocity) while keeping the important characteristics of raw data (Veracity). This gives rise to various new challenges about how we design generators efficiently and successfully. To date, most existing techniques can only generate limited types of data and support specific big data systems such as Hadoop. Hence we develop a tool, called Big Data Generator Suite (BDGS), to efficiently generate scalable big data while employing data models derived from real data to preserve data veracity. The effectiveness of BDGS is demonstrated by developing six data generators covering three representative data types (structured, semi-structured and unstructured) and three data sources (text, graph, and table data).
BibTeX
@inproceedings{ming2013bdgs,
  author = {Zijian Ming and Chunjie Luo and Wanling Gao and Rui Han and Qiang Yang and Lei Wang and Jianfeng Zhan},
  title = {{BDGS: A Scalable Big Data Generator Suite in Big Data Benchmarking}},
  booktitle = {WBDB},
  year = {2013},
  doi = {10.1007/978-3-319-10596-3\_11},
}

Loading related content...

Figure