Kenneth Marshall <ktm@is.rice.edu> writes:
> If the heads of the disk are in the right location, you could easily do
> more than 1 commit per disk revolution so the values over 2 seconds could
> actually be valid. 9 seconds would be worst case of 1 commit per revolution.
No, because a commit in PG involves appending to the WAL log, which
means that in the normal case a commit will have to rewrite the same
sector of WAL that the previous commit did. Barring some weird
remap-on-the-fly scheme, you're going to wait for the disk to come all
the way round for that. Hence, any reported sustained average much
larger than one commit per revolution has to be regarded as probably
phony.
Sometimes you can improve on this using the commit_delay parameters to
gang multiple commits into one physical write, but it's tough to do,
and we already know that this person didn't do any parameter-tuning
whatsoever, and in any case there's no improvement unless you are
committing multiple transactions concurrently.
BTW, the *real* knock against this test methodology is that it's testing
a single serial transaction stream, which is not what PG is designed to
shine at. Someday I'd like to see one of these "I can write a database
benchmark" guys actually test concurrent updates...
regards, tom lane
