Decision Support Benchmark for HashData Database

A comprehensive tool for running TPC-DS benchmarks on Cloudberry / HashData / Greenplum / PostgreSQL. Originally derived from Pivotal TPC-DS.

Overview

This tool provides:

Automated TPC-DS benchmark execution
Support for both local and cloud deployments
Configurable data generation (1GB to 100TB)
Customizable query execution parameters
Detailed performance reporting

Quick Start

# 1. Clone the repository
git clone git@github.com:HashDataFE/TPC-DS-Toolkit.git
cd TPC-DS-Toolkit

# 2. Configure your environment
cp tpcds_variables.sh.template tpcds_variables.sh
vim tpcds_variables.sh

# 3. Run the benchmark
./run.sh

Supported TPC-DS Versions

Version	Date	Specification
3.2.0	2021/06/15	PDF
2.1.0	2015/11/12	PDF
1.3.1	2015/02/19	PDF

This tool uses TPC-DS 3.2.0 as of version 1.2.

Prerequisites

This tool is built with shell scripts and has been tested primarily on CentOS-based operating systems. To accommodate different products, various options are available to choose storage types, partitions, optimizer settings, and distribution policies. Please review the tpcds_variables.sh for detailed configuration options for different models and products.

Tested Products

Cloudberry 1.x / Cloudberry 2.X
HashData Enterprise / HashData Lightning
Greenplum 4.x / Greenplum 5.x / Greenplum 6.x / Greenplum 7.x
PostgreSQL 17.X

Local Cluster Setup

For running tests on the coordinator host:

This mode leverages the MPP architecture to use data directories of segment nodes to generate data and load data using the 'gpfdist' protocol. More resources will be utilized for data generation and loading to accelerate the test process.

Set RUN_MODEL="local" in tpcds_variables.sh
Ensure you have a running HashData Database with gpadmin access
Create a gpadmin database
Configure password-less ssh between mdw (master) and segment nodes (sdw1..n)

Remote Client Setup

For running tests from a remote client:

With this mode, all data will be generated on the client machine, and data will be imported into the database using the copy command. This mode works for HashData Cloud, Cloudberry, Greenplum, HashData Lightning, and should work for other PostgreSQL-compatible products. However, it is recommended to use local mode for non-Cloud MPP products.

Set RUN_MODEL="cloud" in tpcds_variables.sh
Install psql client with passwordless access (.pgpass)

Create gpadmin database with:

ALTER ROLE gpadmin SET warehouse=testforcloud;

Configure required variables in tpcds_variables.sh:

export RANDOM_DISTRIBUTION="true"
export TABLE_STORAGE_OPTIONS="compresstype=zstd, compresslevel=5"
export CLIENT_GEN_PATH="/tmp/dsbenchmark" 
export CLIENT_GEN_PARALLEL="2"

All examples in this documentation use the standard host name convention of HashData with mdw for the master node and sdw1..n for the segment nodes.

TPC-DS Tools Dependencies

Install the dependencies on mdw for compiling the dsdgen (data generation) and dsqgen (query generation) tools:

ssh root@mdw
yum -y install gcc make byacc

The original source code is from the TPC website.

Installation

Simply clone the repository with Git or download the source code from GitHub:

ssh gpadmin@mdw
git clone git@github.com:HashDataFE/TPC-DS-Toolkit.git

Place the folder under /home/gpadmin/ and change ownership to gpadmin:

chown -R gpadmin:gpadmin TPC-DS-Toolkit

Usage

To run the benchmark, login as gpadmin on the master node (mdw):

ssh gpadmin@mdw
cd ~/TPC-DS-Toolkit
./run.sh

By default, this will run a scale 1 (1GB) benchmark with 1 concurrent user, from data generation through to score computation, in the background. Logs will be stored with the name tpcds_<timestamp>.log in the ~/TPC-DS-Toolkit directory.

Configuration

The benchmark is controlled through the tpcds_variables.sh file. Here are the key configuration sections:

Environment Options

# Core settings
export ADMIN_USER="gpadmin"
export BENCH_ROLE="dsbench" 
export DB_SCHEMA_NAME="tpcds"  # Database schema to use for all TPC-DS data tables  
export RUN_MODEL="cloud"    # "local" or "cloud"

# Remote cluster connection
export PSQL_OPTIONS="-h <host> -p <port>"
export CLIENT_GEN_PATH="/tmp/dsbenchmark"  # Location for data generation
export CLIENT_GEN_PARALLEL="2"             # Number of parallel data generation processes

Benchmark Options

# Scale and concurrency 
export GEN_DATA_SCALE="1"    # 1 = 1GB, 1000 = 1TB, 3000 = 3TB
export MULTI_USER_COUNT="2"  # Number of concurrent users during throughput tests

# For large scale tests, consider:
# - 3TB: GEN_DATA_SCALE="3000" with MULTI_USER_COUNT="5"
# - 10TB: GEN_DATA_SCALE="10000" with MULTI_USER_COUNT="7"
# - 30TB: GEN_DATA_SCALE="30000" with MULTI_USER_COUNT="10"

Storage Options

# Table format and compression options
export TABLE_ACCESS_METHOD="USING ao_column"  # Available options:
                                       # - heap: Classic row storage
                                       # - ao_row: Append-optimized row storage
                                       # - ao_column: Append-optimized columnar storage
                                       # - pax: PAX storage format (Cloudberry 2.0/HashData Lightning only)

export TABLE_STORAGE_OPTIONS="WITH (compresstype=zstd, compresslevel=5)"  # Compression settings:
                                                                         # - zstd: Best compression ratio
                                                                         # - compresslevel: 1-19 (higher=better compression)

# Table partitioning for 7 large tables:
# catalog_returns, catalog_sales, inventory, store_returns, store_sales, web_returns, web_sales
export TABLE_USE_PARTITION="true"

Note:

TABLE_ACCESS_METHOD: Default to non-value to be compatible with HashData Cloud and early Greenplum versions. Should be set to USING ao_column for Cloudberry or Greenplum. USING PAX is available for Cloudberry 2.0 and HashData Lightning.
For earlier Greenplum products without TABLE_ACCESS_METHOD support, use full options: appendoptimized=true, orientation=column, compresstype=zlib, compresslevel=5, blocksize=1048576
Distribution policies are defined in TPC-DS-Toolkit/03_ddl/distribution.txt. With products supporting REPLICATED policy, 14 tables use REPLICATED distribution by default. For early Greenplum products without REPLICATED policy support, see TPC-DS-Toolkit/03_ddl/distribution_original.txt.
Table partition definitions are in TPC-DS-Toolkit/03_ddl/*.sql.partition. Note: When using table partitioning along with column-oriented tables, if the block size is set to a large value, it might cause high memory consumption and result in out-of-memory errors. In that case, reduce the block size or the number of partitions.

Step Control Options

# Benchmark execution steps
# 1. Setup and compilation
export RUN_COMPILE_TPCDS="true"  # Compile data/query generators (one-time)
export RUN_INIT="true"           # Initialize cluster settings

# 2. Data generation and loading
export RUN_GEN_DATA="true"       # Generate test data
export RUN_DDL="true"            # Create database schemas/tables
export RUN_LOAD="true"           # Load generated data

# 3. Query execution
export RUN_SQL="true"                 # Run power test queries
export RUN_SINGLE_USER_REPORTS="true" # Upload single-user test results
export RUN_MULTI_USER="false"         # Run throughput test queries
export RUN_MULTI_USER_REPORTS="false" # Upload multi-user test results
export RUN_SCORE="false"              # Compute final benchmark score

There are multiple steps in running the benchmark, controlled by these variables:

Variable	Default	Description
`RUN_COMPILE_TPCDS`	`true`	Compiles `dsdgen` and `dsqgen`. Usually only needs to be done once.
`RUN_GEN_DATA`	`true`	Generates flat files for the benchmark in parallel on all segment nodes. Files are stored under the `${PGDATA}/dsbenchmark` directory.
`RUN_INIT`	`true`	Sets up GUCs for the database and records segment configurations. Only required if the cluster is reconfigured.
`RUN_DDL`	`true`	Recreates all schemas and tables (including external tables for loading). Set to `false` to keep existing data.
`RUN_LOAD`	`true`	Loads data from flat files into tables and computes statistics.
`RUN_SQL`	`true`	Runs the power test of the benchmark.
`RUN_SINGLE_USER_REPORTS`	`true`	Uploads results to the database under the schema `tpcds_reports`. Required for the `RUN_SCORE` step.
`RUN_MULTI_USER`	`true`	Runs the throughput test of the benchmark. This generates multiple query streams using `dsqgen`, which samples the database to find proper filters. For very large databases with many streams, this process can take hours just to generate the queries.
`RUN_MULTI_USER_REPORTS`	`true`	Uploads multi-user results to the database.
`RUN_SCORE`	`true`	Computes the final `QphDS` score based on the benchmark standard.

WARNING: TPC-DS does not rely on the log folder to determine which steps to run or skip. It will only run the steps that are explicitly set to true in the tpcds_variables.sh file. If any necessary step is set to false but has never been executed before, the script will abort when it tries to access data that doesn't exist.

Miscellaneous Options

# Misc options
export SINGLE_USER_ITERATIONS="1"      # Number of times to run the power test
export EXPLAIN_ANALYZE="false"         # Set to true for query plan analysis
export RANDOM_DISTRIBUTION="false"     # Use random distribution for fact tables
export ENABLE_VECTORIZATION="off"      # Set to on/off to enable vectorization
export STATEMENT_MEM="2GB"             # Memory per statement for single-user test
export STATEMENT_MEM_MULTI_USER="1GB"  # Memory per statement for multi-user test
export GPFDIST_LOCATION="p"            # Where gpfdist will run: p (primary) or m (mirror)
export OSVERSION=$(uname)
export ADMIN_USER=$(whoami)
export ADMIN_HOME=$(eval echo ${HOME}/${ADMIN_USER})
export MASTER_HOST=$(hostname -s)

Key options explained:

EXPLAIN_ANALYZE: When set to true, executes queries with EXPLAIN ANALYZE to see query plans, costs, and memory usage. For debugging only, as it affects benchmark results.
RANDOM_DISTRIBUTION: When set to true, fact tables are distributed randomly rather than using pre-defined distribution columns. Recommended for Cloud products.
SINGLE_USER_ITERATION: Controls how many times the power test runs. The fastest query time from multiple runs is used for final scoring.
STATEMENT_MEM: Sets memory per statement for single-user tests (should be less than gp_vmem_protect_limit).
STATEMENT_MEM_MULTI_USER: Sets memory per statement for multi-user tests (note: STATEMENT_MEM_MULTI_USER × MULTI_USER_COUNT should be less than gp_vmem_protect_limit).
ENABLE_VECTORIZATION: Set to on to enable vectorized computing for better performance (supported in Lightning 1.5.3+). Only works with AO column and PAX table formats.

Performance Tuning

For optimal performance:

Memory Settings

# Recommended for 100GB+ RAM systems
export STATEMENT_MEM="8GB"
export STATEMENT_MEM_MULTI_USER="4GB"

Storage Optimization

# For best compression ratio
export TABLE_ACCESS_METHOD="USING ao_column"
export TABLE_STORAGE_OPTIONS="WITH (compresstype=zstd, compresslevel=9)"
export TABLE_USE_PARTITION="true"

Concurrency Tuning

# Adjust based on available CPU cores
export CLIENT_GEN_PARALLEL="$(nproc)"
export MULTI_USER_COUNT="$(( $(nproc) / 2 ))"

Enable Vectorization (for supported systems)
```
export ENABLE_VECTORIZATION="on"
```

Benchmark Modifications

The TPC-DS queries were modified in the following ways to ensure compatibility:

1. Date Interval Syntax Changes

Changed date addition syntax from:

and (cast('2000-02-28' as date) + 30 days)

To:

and (cast('2000-02-28' as date) + '30 days'::interval)

Affected queries: 5, 12, 16, 20, 21, 32, 37, 40, 77, 80, 82, 92, 94, 95, and 98.

2. ORDER BY Column Alias Fixes

Added subqueries for ORDER BY clauses with column aliases:

-- New version with subquery
select * from (
  -- Original query
) AS sub
order by
  lochierarchy desc
  ,case when lochierarchy = 0 then s_state end
  ,rank_within_parent
limit 100;

Affected queries: 36 and 70.

3. Column Reference Corrections

Modified query templates to exclude columns not found in the query, specifically in common table expressions where alias columns were used in dynamic filters. Affected query: 86.

4. Table Alias Additions

Added table aliases to improve query parser compatibility. Affected queries: 2, 14, and 23.

5. Result Limiting

Added LIMIT 100 to queries that could produce very large result sets. Affected queries: 64, 34, and 71.

Troubleshooting

Common Issues and Solutions

Missing or Invalid Environment Variables
Ensure all required environment variables in tpcds_variables.sh are set correctly. If any variable is missing or invalid, the script will abort and display the problematic variable name. Double-check the following key variables:
- RUN_MODEL
- GEN_DATA_SCALE
- TABLE_ACCESS_METHOD
- PSQL_OPTIONS
Permission Errors
- Verify ownership: chown -R gpadmin:gpadmin /home/gpadmin/TPC-DS-Toolkit
- Ensure gpadmin has proper database access permissions
Data Generation Failures
- Confirm successful compilation of dsdgen
- Verify CLIENT_GEN_PATH points to a valid, writable directory
- Check available disk space
Query Execution Errors
- Ensure tables and schemas exist (set RUN_DDL=true on first run)
- Look for syntax errors in modified queries
- Verify database connectivity
Performance Issues
- Adjust memory settings based on system resources
- Enable vectorization if supported
- Use appropriate storage options for your workload
- Consider partitioning for large tables

Logs and Diagnostics

For detailed diagnostics, examine:

Main log file: tpcds_<timestamp>.log in ~/TPC-DS-Toolkit
Database server logs
System resource utilization during test runs

Contributing

Fork the repository
Create your feature branch (git checkout -b feature/amazing-feature)
Commit your changes (git commit -m 'Add amazing feature')
Push to the branch (git push origin feature/amazing-feature)
Open a Pull Request

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Name		Name	Last commit message	Last commit date
Latest commit History 191 Commits
00_compile_tpcds		00_compile_tpcds
01_gen_data		01_gen_data
02_init		02_init
03_ddl		03_ddl
04_load		04_load
05_sql		05_sql
06_single_user_reports		06_single_user_reports
07_multi_user		07_multi_user
08_multi_user_reports		08_multi_user_reports
09_score		09_score
log		log
tpcds_tools		tpcds_tools
.gitignore		.gitignore
Makefile		Makefile
README.md		README.md
functions.sh		functions.sh
rollout.sh		rollout.sh
run.sh		run.sh
tpcds.sh		tpcds.sh
tpcds_variables.sh		tpcds_variables.sh

HashDataFE/TPC-DS-Toolkit-Achived

Folders and files

Latest commit

History

Repository files navigation