haitianchen / pg_tpch

TPC-H like benchmark for PostgreSQL

TPC-H PostgreSQL benchmark

This repository contains a simple implementation that runs a TPC-H-like benchmark with a PostgreSQL database. It builds on the official TPC-H benchmark available at http://tpc.org/tpch/default.asp (uses just the dbgen a qgen parts).

Preparing dbgen and qgen

The first thing you need to do is to prepare the tool that generates data and queries. This step is more thoroughly explained at my blog at

http://www.fuzzy.cz/en/articles/dss-tpc-h-benchmark-with-postgresql/

but let's briefly repeat what needs to be done.

First, download the TPC-H benchmark from http://tpc.org/tpch/default.asp and extract it to a directory

$ wget http://tpc.org/tpch/spec/tpch_2_14_3.tgz
$ mkdir tpch
$ tar -xzf tpch_2_14_3.tgz -C tpch

and then prepare the Makefile - create a copy from makefile.suite

$ cd tpch/dbgen
$ cp makefile.suite Makefile
$ nano Makefile

and modify it so that it contains this (around line 110)

CC=gcc
DATABASE=ORACLE
MACHINE=LINUX
WORKLOAD=TPCH

and compile it using make as usual. Now you should have dbgen and qgen tools that generate data and queries.

Generating data

Right, so let's generate the data using the dbgen tool - there's one important parameter 'scale' that influences the amount of data. It's roughly equal to number of GB of raw data, so to generate 10GB of data just do

$ ./dbgen -s 10

which creates a bunch of .tbl files in Oracle-like CSV format

$ ls *.tbl

and to convert them to a CSV format compatible with PostgreSQL, do this

$ for i in `ls *.tbl`; do sed 's/|$//' $i > ${i/tbl/csv}; echo $i; done;

Finally, move these data to the 'dss/data' directory or somewhere else, and create a symlink to /tmp/dss-data (that's where tpch-load.sql is looking for for the data from).

It's a good idea to place this directory on a ramdrive so that it does not influence the benchmark (e.g. it's a very bad idea to place the data on the same drive as PostgreSQL data directory).

Generating queries

Now we have to generate queries from templates specified in TPC-H benchmark. The templates provided at tpch.org are not suitable for PostgreSQL. So I have provided slightly modified queries in the 'dss/templates' directory and you should place the queries in 'dss/queries' dir.

for q in `seq 1 22`
do
    DSS_QUERY=dss/templates ./qgen $q >> dss/queries/$q.sql
    sed 's/^select/explain select/' dss/queries/$q.sql > dss/queries/$q.explain.sql
    cat dss/queries/$q.sql >> dss/queries/$q.explain.sql;
done

Now you should have 44 files in the dss/queries directory. 22 of them will actually run the queries and the other 22 will generate EXPLAIN plan of the query (without actually running it).

Running the benchmark

The actual benchmark is implemented in the 'tpch.sh' script. It expects an already prepared database and four parameters - directory where to place the results, database and user name. So to run it, do this:

$ ./tpch.sh ./results tpch-db tpch-user

and wait until the benchmark.

Processing the results

All the results are written into the output directory (first parameter). To get useful results (timing of each query, various statistics), you can use script process.php. It expects two parameters - input dir (with data collected by the tpch.sh script) and output file (in CSV format). For example like this:

$ php process.php ./results output.csv

This should give you nicely formatted CSV file.