1. Tomas Vondra
  2. kernel-version-i5



PostgreSQL vs. kernel versions

This repository contains tools (well, a bunch of shell scripts) for running a suite of PostgreSQL benchmarks using pgbench, and results from multiple Linux kernel versions.

benchmark scripts

The following two scripts are the primary benchmarking tools:

  • run-pgbench-short.sh - reinitializes the cluster only once for each scale, runs a suite of read-only/read-write pgbench runs for multiple scales

  • run-pgbench-long.sh - only runs a single long read-write benchmark for each scale, with a single client count (no read-only runs etc.)

Running the scripts is fairly simple - both scripts accept the directory for results as a single parameter, and print a simple progress log. So create a directory 'results' and run the script (storing the results in the directory) like this:

$ mkdir results
$ ./run-pgbench-short.sh ./results > ./results/bench.log 2>&1

directory structure

The repository contains a number of directories matching kernel version, with a variety of information shared by all the runs:

  • PostgreSQL log (pg.log)
  • cpuinfo
  • sysctl info
  • list of mounts
  • pg_settings data
  • postgresql.conf
  • dum of sar data
  • kernel config

and also results for all the runs, orgainzed in a simple directory structure, with $scale/$run directories. For each run the directory contains log files for each type of benchmark (read-only/read-write) and client count. It also includes pgbench transaction log files (aggregated per second) etc.

result processing

To make analysis easier, the repository also includes shell scripts extracting results as CSV (printed on stdout).

  • build-long.sh - throughput info (tps, avg latency) for long runs
  • build-short.sh - throughput info (tps, avg latency) for short runs
  • build-histograms-long.sh - histogram/percentiles from aggregated data
  • build-histograms-short.sh - histogram/percentiles from aggregated data

So for example:

$ ./build-short.sh > short-results.csv

should do the trick.