Re: Optimize planner memory consumption for huge arrays

On 9/8/23 07:11, Lepikhov Andrei wrote:
>
>
> On Wed, Sep 6, 2023, at 8:09 PM, Ashutosh Bapat wrote:
>> Hi Lepikhov,
>>
>> Thanks for using my patch and I am glad that you found it useful.
>>
>> On Mon, Sep 4, 2023 at 10:56 AM Lepikhov Andrei
>> <a(dot)lepikhov(at)postgrespro(dot)ru> wrote:
>>>
>>> Hi, hackers,
>>>
>>> Looking at the planner behaviour with the memory consumption patch [1], I figured out that arrays increase memory consumption by the optimizer significantly. See init.sql in attachment.
>>> The point here is that the planner does small memory allocations for each element during estimation. As a result, it looks like the planner consumes about 250 bytes for each integer element.
>>
>> I guess the numbers you mentioned in init.sql are total memory used by
>> the planner (as reported by the patch in the thread) when planning
>> that query and not memory consumed by Const nodes themselves. Am I
>> right? I think the measurements need to be explained better and also
>> the realistic scenario you are trying to oprimize.
>
> Yes, it is the total memory consumed by the planner - I used the numbers generated by your patch [1]. I had been increasing the number of elements in the array to exclude the memory consumed by the planner for other purposes. As you can see, the array with 1 element consumes 12kB of memory, 1E4 elements - 2.6 MB. All of that memory increment is related to the only enlargement of this array. (2600-12)/10 = 260 bytes. So, I make a conclusion: each 4-byte element produces a consumption of 260 bytes of memory.
> This scenario I obtained from the user complaint - they had strict restrictions on memory usage and were stuck in this unusual memory usage case.
>
>> I guess, the reason you think that partitioning will increase the
>> memory consumed is because each partition will have the clause
>> translated for it. Selectivity estimation for each partition will
>> create those many Const nodes and hence consume memory. Am I right?
>
> Yes.
>
>> Can you please measure the memory consumed with and without your
>> patch.
>
> Done. See test case and results in 'init_parts.sql' in attachment. Short summary below. I varied a number of elements from 1 to 10000 and partitions from 1 to 100. As you can see, partitioning adds a lot of memory consumption by itself. But we see an effect from patch also.
>
> master:
> elems 1 1E1 1E2 1E3 1E4
> parts
> 1 28kB 50kB 0.3MB 2.5MB 25MB
> 10 45kB 143kB 0.6MB 4.8MB 47MB
> 100 208kB 125kB 3.3MB 27MB 274MB
>
> patched:
> elems 1 1E1 1E2 1E3 1E4
> parts
> 1 28kB 48kB 0.25MB 2.2MB 22.8MB
> 10 44kB 100kB 313kB 2.4MB 23.7MB
> 100 208kB 101kB 0.9MB 3.7MB 32.4MB
>
> Just for comparison, without partitioning:
> elems 1 1E1 1E2 1E3 1E4
> master: 12kB 14kB 37kB 266kB 2.5MB
> patched: 12kB 11.5kB 13kB 24kB 141kB
>

These improvements look pretty nice, considering how simple the patch
seems to be. I can't even imagine how much memory we'd need with even
more partitions (say, 1000) if 100 partitions means 274MB.

BTW when releasing memory in scalararraysel, wouldn't it be good to also
free the elem_values/elem_nulls? I haven't tried and maybe it's not that
significant amount.

Considering there are now multiple patches improving memory usage during
planning with partitions, perhaps it's time to take a step back and
think about how we manage (or rather not manage) memory during query
planning, and see if we could improve that instead of an infinite
sequence of ad hoc patches?

Our traditional attitude is to not manage memory, and rely on the memory
context to not be very long-lived. And that used to be fine, but
partitioning clearly changed the equation, increasing the amount of
allocated memory etc.

I don't think we want to stop relying on memory contexts for planning in
general - memory contexts are obviously very convenient etc. But maybe
we could identify "stages" in the planning and release the memory more
aggressively in those?

For example, I don't think we expect selectivity functions to allocate
long-lived objects, right? So maybe we could run them in a dedicated
memory context, and reset it aggressively (after each call).

Ofc, I'm not suggesting this patch should be responsible for doing this.

>>> It is maybe not a problem most of the time. However, in the case of partitions, memory consumption multiplies by each partition. Such a corner case looks weird, but the fix is simple. So, why not?
>>
>> With vectorized operations becoming a norm these days, it's possible
>> to have thousands of element in array of an ANY or IN clause. Also
>> will be common to have thousands of partitions. But I think what we
>> need to do here is to write a selectivity estimation function which
>> takes an const array and return selectivity without requiring to
>> create a Const node for each element.
>
> Maybe you're right. Could you show any examples of vectorized usage of postgres to understand your idea more clearly?
> Here I propose only quick simple solution. I don't think it would change the way of development.
>

I'm a big fan of SIMD and vectorization, but I don't think there's a
chance to achieve that without major reworks to how we evaluate
expressions. It's pretty fundamentally incompatible with how we handle
with user-defined functions, FunctionCall etc.

>>> The diff in the attachment is proof of concept showing how to reduce wasting of memory. Having benchmarked a bit, I didn't find any overhead.
>>>
>>
>> You might want to include your benchmarking results as well.
>
> Here is nothing interesting. pgbench TPS and planning time for the cases above doesn't change planning time.
>

Yeah, I don't think we'd expect regressions from this patch. It pretty
much just pfree-s a list + Const node.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company