|From:||Tomas Vondra <tv(at)fuzzy(dot)cz>|
|Subject:||Re: 9.5: Better memory accounting, towards memory-bounded HashAgg|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
On 2.8.2014 22:40, Jeff Davis wrote:
> Attached is a patch that explicitly tracks allocated memory (the blocks,
> not the chunks) for each memory context, as well as its children.
> This is a prerequisite for memory-bounded HashAgg, which I intend to
Anyway, I'm really looking forward to the memory-bounded hashagg, and
I'm willing to spend some time testing it.
> submit for the next CF. Hashjoin tracks the tuple sizes that it adds to
> the hash table, which is a good estimate for Hashjoin. But I don't think
Actually, even for HashJoin the estimate is pretty bad, and varies a lot
depending on the tuple width. With "narrow" tuples (e.g. ~40B of data),
which is actually quite common case, you easily get ~100% palloc overhead.
I managed to address that by using a custom allocator. See this:
I wonder whether something like that would be possible for the hashagg?
That would make the memory accounting accurate with 0% overhead (because
it's not messing with the memory context at all), but it only for the
one node (but maybe that's OK?).
> it's as easy for Hashagg, for which we need to track transition values,
> etc. (also, for HashAgg, I expect that the overhead will be more
> significant than for Hashjoin). If we track the space used by the memory
> contexts directly, it's easier and more accurate.
I don't think that's comparable - I can easily think about cases leading
to extreme palloc overhead with HashAgg (think of an aggregate
implementing COUNT DISTINCT - that effectively needs to store all the
values, and with short values the palloc overhead will be terrible).
Actually, I was looking at HashAgg (it's a somehow natural direction
after messing with Hash Join), and my plan was to use a similar dense
allocation approach. The trickiest part would probably me making this
available from the custom aggregates.
|Next Message||Robert Haas||2014-08-06 18:56:57||Re: posix_fadvise() and pg_receivexlog|
|Previous Message||Peter Geoghegan||2014-08-06 17:52:36||Re: A worst case for qsort|