Hello hackers,
There were many discussions concerning possible ways of speeding-up
Postgres. Different approaches were suggested:
- JIT (now we have three different prototype implementations based on LLVM)
- Chunked (vectorized) executor
- Replacing pull with push
- Columnar store (cstore_fdw, IMCS)
- Optimizing and improving current executor (reducing tuple deform
overhead, function call overhead,...)
Obviously the best result can be achieved in case of combining all this
approaches. But actually them are more or less interchangeable:
vectorized execution is not eliminating interpretation overhead, but it
is divided by vector size and becomes less critical.
I decided to write small prototype to estimate possible speed
improvement of vectorized executor. I created special types representing
"tile" and implement standard SQL operators for them. So neither
Postgres planer, nether Postgres executor, nether Postgres heap manager
are changed. But I was able to reach more than 10 times speed
improvement on TPC-H Q1/Q6 queries!
Please find more information here:
https://cdn.rawgit.com/postgrespro/vops/ddcbfbe6/vops.html
The sources of the project can be found here:
https://github.com/postgrespro/vops.git
--
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company