Re: heavly load system spec - Mailing list pgsql-performance

From Merlin Moncure
Subject Re: heavly load system spec
Date
Msg-id CAHyXU0xQDwfCsRfUPixa-7GhzYz6mprTAKAaq-u8Poe7uVwDTQ@mail.gmail.com
Whole thread Raw
In response to heavly load system spec  (Gregg Jaskiewicz <gryzman@gmail.com>)
List pgsql-performance
On Thu, Apr 5, 2012 at 10:39 AM, Gregg Jaskiewicz <gryzman@gmail.com> wrote:
> I know this is a very general question. But if you guys had to specify
> system (could be one server or cluster), with sustainable transaction
> rate of 1.5M tps running postgresql, what configuration and hardware
> would you be looking for ?
> The transaction distribution there is 90% writes/updates and 10% reads.
> We're talking 64 linux, Intel/IBM system.
>
> I'm trying to see how that compares with Oracle system.

That's not gonna be possible with stock postgres.  You could cluster
out the reads with a (probably cascaded) HS/SR setup but we'd still
have to get 90% of 1.5M tps running in a single instance.  You'd hit
various bottlenecks trying to get write transactions up anywhere near
that figure -- the walinsert lock being the worst since it is
something of an upper bound on tps rates.  I can't speak for Oracle
but I'm skeptical it's possible there in a monolithic system; if it is
in fact possible it would cost megabucks to do it.

At the end of the day your problem needs some serious engineering and
serious engineers.  Any solution is probably going to involve a
cluster of machines to stage the data with various ETL type jobs to
move it into specific services that will do the actual processing.
This is how all solutions that sustain very high transaction rates
work; you divide tasks and develop communications protocols between
various systems.

merlin

pgsql-performance by date:

Previous
From: Stephen Frost
Date:
Subject: Re: heavly load system spec
Next
From: Claudio Freire
Date:
Subject: Re: High load average in 64-core server , no I/O wait and CPU is idle