We added undo segments to Postgres, and got 20% higher OLTP throughput

With the database size of 24G, three InnoDB buffer pool sizes were tested, each representing a distinct workload characteristic:

Additionally, the WAL log space was adjusted according to the actual shared_buffers size. This ensures that checkpoints consistently trigger at fixed time intervals.

The schema contains 20 tables with 5M rows in each one adding up to 100M rows.

The size of the database files on the disk is 24G.

A common configuration was applied to all engines to ensure fairness. The buffer pool specifications were layered on top

Create the test database “bench”:

createdb bench

Use the sysbench command to load test data into the “bench” database:

sysbench oltp_read_write
--tables=20
--table_size=5000000
--pgsql-user=postgres
--pgsql-db=bench
prepare

After loading the test data, vacuum the “bench” database:

vacuumdb bench

Run the performance benchmark test:

sysbench oltp_read_write
--pgsql-user=postgres
--pgsql-db=bench
--tables=20
--table_size=5000000
--threads=128
--time=600
--report-interval=10
run

Note: The first run is a warmup phase and is not recorded as formal test result data.

The TPS (Transactions Per Second) metrics collected below represent the number of transactions processed by the database per second. In the Sysbench read-write testing model, a single transaction contains 20 query statements, so the Queries Per Second (QPS) processed by the database is approximately 20 times the TPS.

Instance Type: c9.8xlarge.4 (32vCPU 128GB)