Re: In-Memory Columnar Store - Mailing list pgsql-hackers

From knizhnik
Subject Re: In-Memory Columnar Store
Date
Msg-id 52A9FDF0.6000604@garret.ru
Whole thread Raw
In response to Re: In-Memory Columnar Store  (Merlin Moncure <mmoncure@gmail.com>)
Responses Re: In-Memory Columnar Store
List pgsql-hackers
On 12/12/2013 07:03 PM, Merlin Moncure wrote:
> On Thu, Dec 12, 2013 at 4:02 AM, knizhnik <knizhnik@garret.ru> wrote:
> Yeah. It's not fair to compare vs an implementation that is 
> constrained to use only 1mb. For analytics work huge work mem is 
> pretty typical setting. 10x improvement is believable considering 
> you've removed all MVCC overhead, locking, buffer management, etc. and 
> have a simplified data structure. merlin 
I agree that it is not fair comparison. As an excuse I can say that I am 
not an experienced PostgreSQL user, so I thought that setting 
shared_buffers is enough to avoid disk access by PostgreSQL. Only after 
getting such strange results I started investigation of how to properly 
tune P{ostgreSQL parameters.

IMHO it is strange to see such small default values in postgresql 
configuration - PostgreSQL is not an embedded database and now even 
mobile devices have several gigs of memory...
Also it will be nice to have one single switch - how much physical 
memory can PostgreSQL use. And let PostgreSQL spit it in optimal way. 
For example I have no idea how to optimally split memory between 
""shared_buffers", "temp_buffers", "work_mem", "maintenance_work_mem". 
PostgreSQL itself should do this work much better than unexperienced 
administrator.

And one of the possible values of such parameter can be "auto": make it 
possible to automatically determine available memory (it is not a big 
deal to check amount of available RAM in the system). I know that 
vendors of big databases never tries to simplify configuration and 
tuning of their products: just because most of the profit them get from 
consulting. But I think that it is not true for PostgreSQL.




pgsql-hackers by date:

Previous
From: Tom Lane
Date:
Subject: Re: should we add a XLogRecPtr/LSN SQL type?
Next
From: Merlin Moncure
Date:
Subject: Re: In-Memory Columnar Store