Vonng/pgtpc

How fast can PostgreSQL go?

Figures, Numbers, Benchmarks about how fast can PostgreSQL go!

• https://github.com/Vonng/pgtpc

• pgbench： Meaure PostgreSQL performence on different hardwares.

• sysbench：Compare PostgreSQL performence with MySQL & other databases

• tpc-h：Measure PostgreSQL Analysis performence and compare to other AP systems

• Author: Vonng

PGBENCH

PostgreSQL have marvelous OTLP performence measured with pgbench (TPC-B Like).

Record: 2M point select QPS, 137K write TPS (1 Xact = 5 Queries / USUUI )

PGBENCH READ WRITE TPS

PGBENCH READ-ONLY TPS

SYSBENCH

PGSQL has a similar write performance and significantly better read performance to MySQL.

PGSQL & MYSQL both overwhelm several "Distributed Databases".

Sysbench: Fair Play

If we normalize QPS performence with CPU cores used, it will be like:

References

• AWS c5d.metal (NORMAL)
• Reference: TiDB 6.1
• Reference: OceanBase 3.1
• Reference: PolarDB-X

SYSBENCH: Dirty Hack

PGSQL vs MySQL: OLTP Performence

MySQL Official Site - Why MySQL - Benchmarks - MySQL 8.0 Have a blantant bench result.

Which turns off checksum, double write, binlog, PFS, and fsync and use latin-1 charset. Which makes it a toy and impossible to use in real-world production environment.

However we can using the same dirty hack in PostgreSQL and see how far it goes.

• Reference: MySQL
• AWS c5d.metal (EXTREME)

PGSQL vs MySQL: OLTP Point Select

TPC-H

PostgreSQL has pretty good OLAP performence / effect comparing to other Ad-hoc AP systems.

PostgreSQL can finish tpc-h50 in 22min, and tpc-h100 in 80min on a 10C laptop.

Comparison

We can measure relative AP effectiveness by normalize CPU core numbers, Scale Factors. Just for a rough references: Effect = Estimated TPC-H Time spent completing 1 Warehouse using 1 Core.

QPH = ( 1 / time ) * ( scale_factor / core ) * 3600

Reference

[12]: https://dl.acm.org/doi/10.1145/3514221.3526148 "Elena Milkai: "How Good is My HTAP System?",SIGMOD ’22 Session 25" [13]: https://calculator.amazonaws.cn/ "AWS Calculator"