ScalarDB Benchmarking Tools
This tutorial describes how to run benchmarking tools for ScalarDB. Database benchmarking is helpful for evaluating how databases perform against a set of standards.
Benchmark workloads
- TPC-C
- YCSB (Workloads A, C, and F)
- Multi-storage YCSB (Workloads C and F)
- This YCSB variant is for a multi-storage environment that uses ScalarDB.
- Workers in a multi-storage YCSB execute the same number of read and write operations in two namespaces:
ycsb_primary
andycsb_secondary
.
Prerequisites
- One of the following Java Development Kits (JDKs):
- Oracle JDK LTS version 8
- OpenJDK LTS version 8
- Gradle
- Kelpie
- Kelpie is a framework for performing end-to-end testing, such as system benchmarking and verification. Get the latest version from Kelpie Releases, and unzip the archive file.
- A client to run the benchmarking tools
- A target database
- For a list of databases that ScalarDB supports, see Databases.
Currently, only JDK 8 can be used when running the benchmarking tools.
Set up the benchmarking tools
The following sections describe how to set up the benchmarking tools.
Clone the ScalarDB benchmarks repository
Open Terminal, then clone the ScalarDB benchmarks repository by running the following command:
git clone https://github.com/scalar-labs/scalardb-benchmarks
Then, go to the directory that contains the benchmarking files by running the following command:
cd scalardb-benchmarks
Build the tools
To build the benchmarking tools, run the following command:
./gradlew shadowJar
Load the schema
Before loading the initial data, the tables must be defined by using the ScalarDB Schema Loader. To apply the schema, go to the ScalarDB Releases page and download the ScalarDB Schema Loader that matches the version of ScalarDB that you are using to the scalardb-benchmarks
root folder.
In addition, you need a properties file that contains database configurations for ScalarDB. For details about configuring the ScalarDB properties file, see ScalarDB Configurations.
After applying the schema and configuring the properties file, select a benchmark and follow the instructions to create the tables.
- TPC-C
- YCSB
- Multi-storage YCSB
To create tables for TPC-C benchmarking (tpcc-schema.json
), run the following command, replacing the contents in the angle brackets as described:
java -jar scalardb-schema-loader-<VERSION>.jar --config <PATH_TO_SCALARDB_PROPERTIES_FILE> -f tpcc-schema.json --coordinator
To create tables for YCSB benchmarking (ycsb-schema.json
), run the following command, replacing the contents in the angle brackets as described:
java -jar scalardb-schema-loader-<VERSION>.jar --config <PATH_TO_SCALARDB_PROPERTIES_FILE> -f ycsb-schema.json --coordinator
To create tables for multi-storage YCSB benchmarking (ycsb-multi-storage-schema.json
), run the following command, replacing the contents in the angle brackets as described:
java -jar scalardb-schema-loader-<VERSION>.jar --config <PATH_TO_SCALARDB_PROPERTIES_FILE> -f ycsb-multi-storage-schema.json --coordinator
Prepare a benchmarking configuration file
To run a benchmark, you must first prepare a benchmarking configuration file. The configuration file requires at least the locations of the workload modules to run and the database configuration.
The following is an example configuration for running the TPC-C benchmark. The ScalarDB properties file specified for config_file
should be the properties file for the benchmarking environment that you previously set up.
Alternatively, instead of using the ScalarDB properties file, you can specify each database configuration item in the .toml
file. If config_file
is specified, all other configurations under [database_config]
will be ignored even if they are uncommented.
[modules]
[modules.preprocessor]
name = "com.scalar.db.benchmarks.tpcc.TpccLoader"
path = "./build/libs/scalardb-benchmarks-all.jar"
[modules.processor]
name = "com.scalar.db.benchmarks.tpcc.TpccBench"
path = "./build/libs/scalardb-benchmarks-all.jar"
[modules.postprocessor]
name = "com.scalar.db.benchmarks.tpcc.TpccReporter"
path = "./build/libs/scalardb-benchmarks-all.jar"
[database_config]
config_file = "<PATH_TO_SCALARDB_PROPERTIES_FILE>"
#contact_points = "localhost"
#contact_port = 9042
#username = "cassandra"
#password = "cassandra"
#storage = "cassandra"
You can define parameters to pass to modules in the configuration file. For details, see the sample configuration files below and available parameters in Common parameters:
- TPC-C:
tpcc-benchmark-config.toml
- YCSB:
ycsb-benchmark-config.toml
- Multi-storage YCSB:
ycsb-multi-storage-benchmark-config.toml
Run a benchmark
Select a benchmark, and follow the instructions to run the benchmark.
- TPC-C
- YCSB
- Multi-storage YCSB
To run the TPC-C benchmark, run the following command, replacing <PATH_TO_KELPIE>
with the path to the Kelpie directory:
/<PATH_TO_KELPIE>/bin/kelpie --config tpcc-benchmark-config.toml
To run the YCSB benchmark, run the following command, replacing <PATH_TO_KELPIE>
with the path to the Kelpie directory:
/<PATH_TO_KELPIE>/bin/kelpie --config ycsb-benchmark-config.toml
To run the multi-storage YCSB benchmark, run the following command, replacing <PATH_TO_KELPIE>
with the path to the Kelpie directory:
/<PATH_TO_KELPIE>/bin/kelpie --config ycsb-multi-storage-benchmark-config.toml
In addition, the following options are available:
--only-pre
. Only loads the data.--only-process
. Only runs the benchmark.--except-pre
Runs a job without loading the data.--except-process
. Runs a job without running the benchmark.
Common parameters
Name | Description | Default |
---|---|---|
concurrency | Number of threads for benchmarking. | 1 |
run_for_sec | Duration of benchmark (in seconds). | 60 |
ramp_for_sec | Duration of ramp-up time before benchmark (in seconds). | 0 |
Workload-specific parameters
Select a benchmark to see its available workload parameters.
- TPC-C
- YCSB and multi-storage YCSB
Name | Description | Default |
---|---|---|
num_warehouses | Number of warehouses (scale factor) for benchmarking. | 1 |
load_concurrency | Number of threads for loading. | 1 |
load_start_warehouse | Start ID of loading warehouse. This option can be useful with --skip-item-load when loading large-scale data with multiple clients or adding additional warehouses. | 1 |
load_end_warehouse | End ID of loading warehouse. You can use either --num-warehouses or --end-warehouse to specify the number of loading warehouses. | 1 |
skip_item_load | Whether or not to skip loading item table. | false |
use_table_index | Whether or not to use a generic table-based secondary index instead of ScalarDB's secondary index. | false |
np_only | Run benchmark with only new-order and payment transactions (50% each). | false |
rate_new_order | Percentage of new-order transactions. When specifying this percentage based on your needs, you must specify the percentages for all other rate parameters. In that case, the total of all rate parameters must equal 100 percent. | N/A |
rate_payment | Percentage of payment transactions. When specifying this percentage based on your needs, you must specify the percentages for all other rate parameters. In that case, the total of all rate parameters must equal 100 percent. | N/A |
rate_order_status | Percentage of order-status transactions. When specifying this percentage based on your needs, you must specify the percentages for all other rate parameters. In that case, the total of all rate parameters must equal 100 percent. | N/A |
rate_delivery | Percentage of delivery transactions. When specifying this percentage based on your needs, you must specify the percentages for all other rate parameters. In that case, the total of all rate parameters must equal 100 percent. | N/A |
rate_stock_level | Percentage of stock-level transactions. When specifying this percentage based on your needs, you must specify the percentages for all other rate parameters. In that case, the total of all rate parameters must equal 100 percent. | N/A |
backoff | Sleep time in milliseconds inserted after a transaction is aborted due to a conflict. | 0 |
Name | Description | Default |
---|---|---|
load_concurrency | Number of threads for loading. | 1 |
load_batch_size | Number of put records in a single loading transaction. | 1 |
load_overwrite | Whether or not to overwrite when loading records. | false |
ops_per_tx | Number of operations in a single transaction. | 2 (Workloads A and C) 1 (Workload F) |
record_count | Number of records in the target table. | 1000 |
use_read_modify_write | Whether or not to use read-modify-writes instead of blind writes in Workload A. | false 1 |
Footnotes
-
The default value is
false
foruse_read_modify_write
since Workload A doesn't assume that the transaction reads the original record first. However, if you're using Consensus Commit as the transaction manager, you must setuse_read_modify_write
totrue
. This is because ScalarDB doesn't allow a blind write for an existing record. ↩