Re: Different memory allocation strategy in Postgres 11?

Thomas Munro schrieb am 26.10.2018 um 22:13:
>>> I typically configure "shared_buffers = 4096MB" on my 16GB system as sometimes when testing, it pays off to have a bigger cache.
>>>
>>> With Postgres 10 and earlier, the Postgres process(es) would only allocate that memory from the operating system when needed.
>>> So right after startup, it would only consume several hundred MB, not the entire 4GB
>>>
>>> However with Postgres 11 I noticed that it immediately grabs the complete memory configured for shared_buffers during startup.
>>>
>>> It's not really a big deal, but I wonder if that is an intentional change or a result from something else?
>>>
>>>
>>> Do you have pg_prewarm in shared_preload_libraries?
>>
>> No. The only shared libraries are those for pg_stat_statemens
>
> Does your user have "Lock Pages in Memory" privilege? One thing that
> is new in 11 is huge AKA large page support, and the default is
> huge_pages=try. Not a Windows person myself but I believe that should
> succeed if you have that privilege and enough contiguous chunks of
> physical memory are available. If you set huge_pages=off does it
> revert to the old behaviour?

Turns out this was an "optimization" in Windows 10, and completely unrelated to Postgres.

Windows 10 has a feature called "Fast Boot" (or something along the lines).

When that is activated (which it is by default), a proper shutdown of the system does not seem to really shut it down. This is especially noteworthy with services: they don't get a shutdown event (which e.g. means even a service marked as "manual start", will still be running after a reboot if it did before)

In case of Postgres this is visible e.g. in the logfile, because there will no shutdown or startup messages.

So when I booted my laptop, Postgres continued where it was before the reboot - and the memory usage was caused caused by myself generating test data using generate_series() but I expected a "clean" state after the reboot.

When manually restarting the service everything works as expected.

Sorry for the noise.