How to Run TPC-C Test on TiDB

This document describes how to test TiDB using TPC-C.

TPC-C is an online transaction processing (OLTP) benchmark. It tests the OLTP system by using a commodity sales model that involves the following five transactions of different types:

  • NewOrder
  • Payment
  • OrderStatus
  • Delivery
  • StockLevel


Before testing, TPC-C Benchmark specifies the initial state of the database, which is the rule for data generation in the database. The ITEM table contains a fixed number of 100,000 items, while the number of warehouses can be adjusted. If there are W records in the WAREHOUSE table, then:

  • The STOCK table has W * 100,000 records (Each warehouse corresponds to the stock data of 100,000 items)
  • The DISTRICT table has W * 10 records (Each warehouse provides services to 10 districts)
  • The CUSTOMER table has W * 10 * 3,000 records (Each district has 3,000 customers)
  • The HISTORY table has W * 10 * 3,000 records (Each customer has one transaction history)
  • The ORDER table has W * 10 * 3,000 records (Each district has 3,000 orders and the last 900 orders generated are added to the NEW-ORDER table. Each order randomly generates 5 ~ 15 ORDER-LINE records.

In this document, the testing uses 1,000 warehouses as an example to test TiDB.

TPC-C uses tpmC (transactions per minute) to measure the maximum qualified throughput (MQTh, Max Qualified Throughput). The transactions are the NewOrder transactions and the final unit of measure is the number of new orders processed per minute.

This testing uses the open-source BenchmarkSQL 5.0 as the TPC-C testing tool and adds the support for the MySQL protocol. You can download the testing program by using the following command:

git clone -b 5.0-mysql-support-opt-2.1

To install Java and Ant, run the following command (take CentOS as an example):

sudo yum install -y java ant

Enter the benchmarksql directory and run ant to build:

cd benchmarksql && \

Deploy the TiDB cluster

For 1,000 warehouses, the TiDB cluster can be deployed on three machines, with one TiDB instance, one PD instance, and one TiKV instance on each machine.

For example, the hardware configuration is as follows:

Type Name
OS Linux (CentOS 7.3.1611)
CPU 40 vCPUs, Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz
DISK Optane 500GB SSD
  1. Because this type of CPU has an NUMA architecture, it is recommended to bind the core using taskset and view the NUMA node using lscpu. For example:

    NUMA node0 CPU(s):     0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38
    NUMA node1 CPU(s):     1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39
  2. Start TiDB using the following command:

    nohup taskset -c 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38 bin/tidb-server &&
    nohup taskset -c 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39 bin/tidb-server
  3. You can deploy a HAproxy to balance the load of multiple TiDB nodes. It is recommended to configure nbproc as the number of CPU cores.

Edit the configuration

Configure TiDB

level = "error"

# Sets the maximum number of CPU cores to use for a single TiDB instance according to NUMA configuration.
max-procs = 20

# Enables the prepared plan cache in TiDB configuration to reduce the overhead of optimizing your execution plan.
enabled = true

Configure TiKV

You can use the basic configuration at the beginning. Then after the test is run, you can adjust it based on the metrics on Grafana and the TiKV Tuning Instructions.

Configure BenchmarkSQL

Edit the benchmarksql/run/props.mysql file:

warehouses=1000 # Uses 1,000 warehouses.
terminals=500   # Uses 500 terminals.
loadWorkers=32  # The number of concurrent workers that load data.

Load data

  1. Use a MySQL client to connect to the TiDB server and run the following command:

    create database tpcc;
  2. Run the following BenchmarkSQL script in shell to create tables:

    cd run && \
    ./ props.mysql sql.mysql/tableCreates.sql && \
    ./ props.mysql sql.mysql/indexCreates.sql
  3. Run the following script to load data:

    ./ props.mysql

This process might last for several hours depending on the machine configuration.

After importing data, you can run sql.common/test.sql to validate the correctness of the data. If all SQL statements return an empty result, then the data is correctly imported.

Run the test

Run the following BenchmarkSQL test script:

nohup ./ props.mysql &> test.log &

After the execution is finished, view the result using test.log:

07:09:53,455 [Thread-351] INFO   jTPCC : Term-00, Measured tpmC (NewOrders) = 77373.25
07:09:53,455 [Thread-351] INFO   jTPCC : Term-00, Measured tpmTOTAL = 171959.88
07:09:53,455 [Thread-351] INFO   jTPCC : Term-00, Session Start     = 2019-03-21 07:07:52
07:09:53,456 [Thread-351] INFO   jTPCC : Term-00, Session End       = 2019-03-21 07:09:53
07:09:53,456 [Thread-351] INFO   jTPCC : Term-00, Transaction Count = 345240

The value in the tpmC section is the testing result.

After the test completes, you can also run sql.common/test.sql to validate the correctness of the data. If all SQL statements return an empty result, then the testing of the data is correctly performed.