LDBC SNB Interactive workload implementations

Reference implementations of the LDBC Social Network Benchmark's Interactive workload (paper, specification on GitHub pages, specification on arXiv).

To get started with the LDBC SNB benchmarks, check out our introductory presentation: The LDBC Social Network Benchmark (PDF).

Notes

⚠️ There are some quirks to using this repository:

The goal of the implementations in this repository is to serve as reference implementations which other implementations can cross-validated against. Therefore, our primary objective was readability and not absolute performance when formulating the queries.
SNB data sets of different scale factors require different configurations for the benchmark runs. Therefore, make sure you use the correct properties (update_interleave value and query frequencies) based on the files provided in the sf-properties/ directory.
The default workload contains updates which are persisted in the database. Therefore, the database needs to be reloaded or restored from backup before each run. Use the provided scripts/backup-database.sh and scripts/restore-database.sh scripts to achieve this.
We expect most systems-under-test to use multi-threaded execution for their benchmark runs. To allow running the updates on multiple threads, the update stream files need to be partitioned accordingly by the generator. We have pre-generated these for frequent partition numbers (1, 2, ..., 1024 and 24, 48, 96, ..., 768) and scale factors up to 1000 (their deployment is in progress).

Implementations

We provide two reference implementations:

Additional implementations:

For detailed instructions, consult the READMEs of the projects.

To build a subset of the projects, use Maven profiles, e.g. to build the reference implementations, run:

mvn clean package -DskipTests -Pcypher,postgres

User's guide

Building the project

To build the project, run:

scripts/build.sh

Inputs

The benchmark framework relies on the following inputs produced by the SNB Datagen:

Initial data set: the SNB graph in CSV format (social_network/{static,dynamic})
Update streams: the input for the update operations (social_network/updateStream_*.csv)
Substitution parameters: the input parameters for the complex queries. It is produced by the Datagen (substitution_parameters/)

Driver modes

For each implementation, it is possible to perform to perform the run in one of the SNB driver's three modes. All three should be started withe the initial data set loaded to the database.

Create validation parameters with the driver/create-validation-parameters.sh script.
- Inputs:
  - The query substitution parameters are taken from the directory set in ldbc.snb.interactive.parameters_dir configuration property.
  - The update streams are the updateStream_0_0_{forum,person}.csv files from the location set in the ldbc.snb.interactive.updates_dir configuration property.
  - For this mode, the query frequencies are set to a uniform 1 value to ensure the best average test coverage.
- Output: The results will be stored in the validation parameters file (e.g. validation_params.csv) file set in the create_validation_parameters configuration property.
- Parallelism: The execution must be single-threaded to ensure a deterministic order of operations.
Validate against existing validation parameters with the driver/validate.sh script.
- Input:
  - The query substitution parameters are taken from the validation parameters file (e.g. validation_params.csv) file set in the validate_database configuration property.
  - The update operations are also based on the content of the validation parameters file.
- Output:
  - The validation either passes of fails.
  - The per query results of the validation are printed to the console.
  - If the validation failed, the results are saved to the validation_params-failed-expected.json and validation_params-failed-actual.json files.
- Parallelism: The execution must be single-threaded to ensure a deterministic order of operations.
Run the benchmark with the driver/benchmark.sh script.
- Inputs:
  - The query substitution parameters are taken from the directory set in ldbc.snb.interactive.parameters_dir configuration property.
  - The update streams are the updateStream_*_{forum,person}.csv files from the location set in the ldbc.snb.interactive.updates_dir configuration property.
    - To get 2n write threads, the framework requires n updateStream_*_forum.csv and n updateStream_*_person.csv files.
    - If you are generating the data sets from scratch, set ldbc.snb.datagen.serializer.numUpdatePartitions to n in the data generator to get produce these.
  - The goal of the benchmark is the achieve the best (lowest possible) time_compression_ratio value while ensuring that the 95% on-time requirement is kept (i.e. 95% of the queries can be started within 1 second of their scheduled time). If your benchmark run returns "failed schedule audit", increase this number (which lowers the time compression rate) until it passes.
  - Set the thread_count property to the size of the thread pool for read operations.
  - Different scale factors require different configurations for the ldbc.snb.interactive.update_interleave value and the query frequencies. Make sure you use the correct properties based on the files provided in the sf-properties/ directory.
  - For audited benchmarks, ensure that the warmup and operation_count properties are set so that the warmup and benchmark phases last for 30+ minutes and 2+ hours, respectively.
- Output:
  - Passed or failed the "schedule audit" (the 95% on-time requirement).
  - The throughput achieved in the run (operations/second).
  - The detailed results of the benchmark are printed to the console and saved in the results/ directory.
- Parallelism: Multi-threaded execution is recommended to achieve the best result.

For all scripts, configure the properties file (driver/${MODE}.properties) to match your setup and the scale factor of the data set used.

For more details on validating and benchmarking, visit the driver wiki.

Developer's guide

To create a new implementation, it is recommended to use one of the existing ones: the Neo4j implementation for graph database management systems and the PostgreSQL implementation for RDBMSs.

The implementation process looks roughly as follows:

Create a bulk loader which loads the initial data set to the database.
Implement the complex and short reads queries (22 in total).
Implement the 7 update queries.
Test the implementation against the reference implementations using various scale factors.
Optimize the implementation.

Data sets

Benchmark data sets

To generate the benchmark data sets, use the Hadoop-based LDBC SNB Datagen.

The key configurations are the following:

ldbc.snb.datagen.generator.scaleFactor: set this to snb.interactive.${SCALE_FACTOR} where ${SCALE_FACTOR} is the desired scale factor
ldbc.snb.datagen.serializer.numUpdatePartitions: set this to the number of write threads used in the benchmark runs
serializers: set these to the required format, e.g. the ones starting with CsvMergeForeign or CsvComposite
- ldbc.snb.datagen.serializer.dynamicActivitySerializer
- ldbc.snb.datagen.serializer.dynamicPersonSerializer
- ldbc.snb.datagen.serializer.staticSerializer

Pre-generated data sets

Producing large-scale data sets requires non-trivial amounts of memory and computing resources (e.g. SF100 requires 24GB memory and takes about 4 hours to generate on a single machine). To mitigate this, we have pregenerated data sets using 9 different serializers and the update streams using 17 different partition numbers:

Serializers: csv_basic, csv_basic-longdateformatter, csv_composite, csv_composite-longdateformatter, csv_composite_merge_foreign, csv_composite_merge_foreign-longdateformatter, csv_merge_foreign, csv_merge_foreign-longdateformatter, ttl
Partition numbers: 2^k (1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024) and 6×2^k (24, 48, 96, 192, 384, 768).

The data sets are available at the SURF/CWI data repository.

Test data set

The test data sets are placed in the cypher/test-data/ directory for Neo4j and in the postgres/test-data/ for the SQL systems.

To generate a data set with the same characteristics, see the documentation on generating the test data set.

Preparing for an audited run

Implementations of the Interactive workload can be audited by a certified LDBC auditor. The Auditing Policies chapter of the specification describes the auditing process and the required artifacts.

Determining the best TCR

Select a scale factor and configure the driver/benchmark.properties file as described in the Driver modes section.
Load the data set with scripts/load-in-one-step.sh.
Create a backup with scripts/backup-database.sh.
Run the driver/determine-best-tcr.sh.
Once the "best TCR" value has been determined, test it with a full workload (at least 0.5h for warmup operation and at least 2h of benchmark time), and make further adjustments if necessary.

Recommendations

We have a few recommendations for creating audited implementations. (These are not requirements – implementations are allowed to deviate from these recommendations.)

The implementation should target a popular Linux distribution (e.g. Ubuntu LTS, CentOS, Fedora).
Use a containerized setup, where the DBMS is running in a Docker container.
Instead of a specific hardware, target a cloud virtual machine instance (e.g. AWS r5d.12xlarge). Both bare-metal and regular instances can be used for audited runs.

arvindshmicrosoft/ldbc_snb_interactive_impls