Re: Add the ability to limit the amount of memory that can be allocated to backends.

On 12/26/23 11:49, Anton A. Melnikov wrote:
> Hello!
>
> Earlier in this thread, the pgbench results were published, where with a
> strong memory limit of 100MB
> a significant, about 10%, decrease in TPS was observed [1].
>
> Using dedicated server with 12GB RAM and methodology described in [3], i
> performed five series
> of measurements for the patches from the [2].

Can you share some info about the hardware? For example the CPU model,
number of cores, and so on. 12GB RAM is not quite huge, so presumably it
was a small machine.

> The series were like this:
> 1) unpatched 16th version at the REL_16_BETA1 (e0b82fc8e83) as close to
> [2] in time.
> 2) patched REL_16_BETA1 at e0b82fc8e83 with undefined
> max_total_backend_memory GUC (with default value = 0).
> 3) patched REL_16_BETA1 with max_total_backend_memory = 16GB
> 4) the same with max_total_backend_memory = 8GB
> 5) and again with max_total_backend_memory = 200MB
>

OK

> Measurements with max_total_backend_memory = 100MB were not be carried out,
> with limit 100MB the server gave an error on startup:
> FATAL:  configured max_total_backend_memory 100MB is <=
> shared_memory_size 143MB
> So i used 200MB to retain all other GUCs the same.
>

I'm not very familiar with the patch yet, but this seems a bit strange.
Why should shared_buffers be included this limit?

> Pgbench gave the following results:
> 1) and 2) almost the same: ~6350 TPS. See orange and green
> distributions on the attached graph.png respectively.
> 3) and 4) identical to each other (~6315 TPS) and a bit slower than 1)
> and 2) by ~0,6%.
> See blue and yellow distributions respectively.
> 5) is slightly slower (~6285 TPS) than 3) and 4) by another 0,5%. (grey
> distribution)
> The standard error in all series was ~0.2%. There is a raw data in the
> raw_data.txt.
>

I think 6350 is a pretty terrible number, especially for scale 8, which
is maybe 150MB of data. I think that's a pretty clear sign the system
was hitting some other bottleneck, which can easily mask regressions in
the memory allocation code. AFAICS the pgbench runs were regular r/w
benchmarks, so I'd bet it was hitting I/O, and possibly even subject to
some random effects at that level.

I think what would be interesting are runs with

pgbench -M prepared -S -c $N -j $N

i.e. read-only tests (to not hit I/O), and $N being sufficiently large
to maybe also show some concurrency/locking bottlenecks, etc.

I may do some benchmarks if I happen to find a bit of time, but maybe
you could collect such numbers too?

The other benchmark that might be interesting is more OLAP, with low
concurrency but backends allocating a lot of memory.

regards

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