Hmmm... why does pl/pgsql code parallelise so badly when queries parallelise fine? Anyone else seen this?

Hi everyone,

I've written a new open source tool for easily parallelising SQL scripts in postgres. [obligatory plug: https://github.com/gbb/par_psql ]

Using it, I'm seeing a problem that I've also seen in other postgres projects involving high degrees of parallelisation in the last 12 months.

Basically:

- I have machines here with up to 16 CPU cores and 128GB memory, very fast SSDs and controller etc, carefully configured kernel/postgresql.conf for high performance.

- Ordinary queries parallelise nearly perfectly (e.g. SELECT some_stuff ...), e.g. almost up to 16x performance improvement.

- Non-DB stuff like GDAL, python etc. parallelise nearly perfectly.

- HOWEVER calls to CPU-intensive user-defined pl/pgsql functions (e.g. SELECT myfunction(some_stuff)) do not parallelise well, even when they are independently defined functions, or accessing tables in a read-only way. They hit a limit of 2.5x performance improvement relative to single-CPU performance (pg9.4) and merely 2x performance (pg9.3) regardless of how many CPU cores I throw at them. This is about 6 times slower than I'm expecting.

I can't see what would be locking. It seems like it's the pl/pgsql environment itself that is somehow locking or incurring some huge frictional costs. Whether I use independently defined functions, independent source tables, independent output tables, makes no difference whatsoever, so it doesn't feel 'lock-related'. It also doesn't seem to be WAL/synchronisation related, as the machines I'm using can hit absurdly high pgbench rates, and I'm using unlogged tables for output.

Take a quick peek here: https://github.com/gbb/par_psql/blob/master/BENCHMARKS.md

I'm wondering what I'm missing here. Any ideas?

Graeme.