Partitioning benchmark for PostgreSQL 10

This benchmark is based on YCSB.

Overview

This benchmark runs all kinds of workloads (a, b, etc) for several minutes. The table is partitioned by range (using pg_pathman or vanilla), partition key is the 1st column of type varchar(255). The amount of partitions is always the same (500), and the number of worker threads (sessions) changes from 1 to 128 with a step size of 16.

Step-by-step guide

Perform the following steps:

1. Install your favorite Java SDK (e.g. OpenJDK) and Maven.
2. Build YCSB using Maven: mvn package.
3. Install or build PostgreSQL 10.
4. Install pg_pathman.
5. Change logs_dir and pg_bin_dir in run.sh.
6. (OPTIONAL) Adjust port in run.sh and *.conf files.
7. Create a cluster using initdb and adjust its postgresql.conf using postgresql.add.
8. Start the PostgreSQL cluster.
9. Finally, run the benchmark (e.g. ./run.sh pathman). Available options: pathman, vanilla.

What should I do next?

Examine the logs in logs_dir. There should be "workload-X-T" + "insert-X-T" files, where X is a workload type (a, b, c etc) and T is the number of threads. Note that we always use T=FILLER_THREADS for ycsb load to finish it as quickly as possible. The most interesting lines contain words "Throughput" and "RunTime", so grep is your best friend.