Pg_upgrade by default (without --link) copies heap/index files from the
old to new cluster. This patch implements parallel heap/index file
copying in pg_upgrade using the --jobs option. It uses the same
infrastructure used for pg_upgrade parallel dump/restore. Here are the
performance results:
--- seconds ---
GB git patched
2 62.09 63.75
4 95.93 107.22
8 194.96 195.29
16 494.38 348.93
32 983.28 644.23
64 2227.73 1244.08
128 4735.83 2547.09
Because of the kernel cache, you only see a big win when the amount of
copy data exceeds the kernel cache. For testing, I used a 24GB, 16-core
machine with two magnetic disks with one tablespace on each. Using more
tablespaces would yield larger improvements. My test script is
attached.
I consider this patch ready for application. This is the last
pg_upgrade performance improvement idea I am considering.
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://enterprisedb.com
+ It's impossible for everything to be true. +