GraphScope Refreshes the World Record for the LDBC SNB Benchmark in both Performance and Data Scale

on Tech

benchmark Recently, LDBC released the latest results for the LDBC SNB Interactive benchmark test, where GraphScope Flex leads the pack once again with a score exceeding 127,000 QPS (Queries Per Second), representing a more than 2.6 times improvement over the second place, which was the previous record holder!

Benchmark screenshot

In addition, GraphScope Flex has updated the maximum data scale for the benchmark test, being the only system among all the participants that used the SF1000 dataset for the test. The SF1000 dataset, generated with the LDBC’s official data generator, has a Scale Factor (SF) of 1000. This dataset contains approximately 2.9 billion vertices and 208 billion edges, with the original disk data file size around 1TB, making it an extremely large-scale graph. Previously, the maximum any system participating in the LDBC SNB Interactive had been tested was SF300, which is about a third of the scale of SF1000.

LDBC is an internationally recognized authority in the field of graph computing, developing and leading a series of graph benchmarks, including the Social Network Benchmark (SNB), the Graphalytics Benchmark, and the Financial Benchmark (FinBench), to measure the capabilities and performance of various types of graph systems. The SNB Interactive benchmark test that GraphScope Flex participated in simulates a social network graph of Facebook at different scales, executing operations including basic Create, Read, Update, Delete (CRUD) as well as complex tasks like shortest paths and multi-hop queries, covering commonly used operations and queries in transactional online query scenarios and is recognized as an authoritative benchmark in the field of graph data. During the test process, a driver continuously sends a variety of query requests to the system being tested at high speed, and the system’s performance is measured by its throughput (QPS) in handling queries. The testing procedure is professional and rigorous, with systems submitting their codes to be tested and their scores verified by LDBC’s professional staff.

In recent years, with the collective efforts of R&D personnel in the graph database industry, the performance of graph computing has continuously improved. GraphScope Flex participated in the LDBC SNB Interactive benchmark test for the first time in July 2023, achieving more than a 2.45 times performance improvement over the then world record holder, TuGraph (developed by Ant Group). Then in December 2023, AtlasGraph took the lead with a performance 45% higher than that of GraphScope Flex. One year later, GraphScope Flex has once again won the top spot, achieving approximately 2.6 times the performance of AtlasGraph and testing on the larger data scale SF1000 for the first time. This performance is due to a series of technical upgrades in GraphScope Flex over the year:

  • A more compact graph structure and more efficient graph storage, allowing larger graphs to be loaded into limited memory with higher access efficiency;
  • Further optimized graph queries, with the help of auxiliary data, online cost analysis is utilized to choose the optimal execution sequence, reducing the overhead of queries.

We will provide a detailed analysis of the technical optimizations used in the version of GraphScope Flex for this benchmark test. Stay tuned!