"Timothy J. Kordas" <tkordas@greenplum.com> writes:
> I would expect for the same data a hash-join with a work_mem of 256MB to run
> faster than one run with 32MB; even if the inner relation is only 30MB.
Once you get to the point where each tuple is in a different bucket, it
is clearly impossible for further increases in hashtable size to improve
matters. All you can do is waste RAM and cache lines.
Now if we set NTUP_PER_BUCKET = 1 we would not be exactly at that critical
point because of uneven bucket loading and other factors ... but I
question whether there's enough incremental improvement available to
justify making the hashtable much larger than that.
regards, tom lane