Re: Final decision

From: Josh Berkus <josh(at)agliodbs(dot)com>
To: "Joel Fradkin" <jfradkin(at)wazagua(dot)com>
Cc: "PostgreSQL Perform" <pgsql-performance(at)postgresql(dot)org>
Subject: Re: Final decision
Date: 2005-04-27 16:13:35
Message-ID: 200504270913.35952.josh@agliodbs.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-performance

Joel,

> So I am planning on sticking with postgres fro our production database
> (going live this weekend).

Glad to have you.

> I did not find any resolutions to my issues with Commandprompt.com (we only
> worked together 2.5 hours).

BTW, your performance troubleshooting will continue to be hampered if you
can't share actual queries and data structure. I strongly suggest that you
make a confidentiality contract with a support provider so that you can give
them detailed (rather than general) problem reports.

> Most of my application is working about the same speed as MSSQL server
> (unfortunately its twice the speed box, but as many have pointed out it
> could be an issue with the 4 proc dell). I spent considerable time with
> Dell and could see my drives are delivering 40 meg per sec.

FWIW, on a v40z I get 180mb/s. So your disk array on the Dell is less than
ideal ... basically, what you have is a more expensive box, not a faster
one :-(

> Things I still have to make better are my settings in config, I have it set
> to no merge joins and no seq scans.

Yeah, I'm also finding that our estimator underestimates the real cost of
merge joins on some systems. Basically we need a sort-cost variable,
because I've found an up to 2x difference in sort cost depending on
architecture.

> I am going to have to use flattened history files for reporting (I saw huge
> difference here the view for audit cube took 10 minutes in explain analyze
> and the flattened file took under one second).
> I understand both of these practices are not desirable, but I am at a place
> where I have to get it live and these are items I could not resolve.

Flattening data for reporting is completely reasonable; I do it all the time.

> I believe that was totally IIS not postgres, but I am curious as to if
> using postgres odbc will put more stress on the IIS side then MSSQL did.

Actually, I think the problem may be ODBC. Our ODBC driver is not the best
and is currently being re-built from scratch. Is using npgsql, a much
higher-performance driver (for .NET) out of the question? According to one
company, npgsql performs better than drivers supplied by Microsoft.

> I did have a question if any folks are using two servers one for reporting
> and one for data entry what system should be the beefier?

Depends on the relative # of users. This is often a good approach, because
the requirements for DW reporting and OLTP are completely different.
Basically:
OLTP: Many slow processors, disk array set up for fast writes, moderate shared
mem, low work_mem.
DW: Few fast processors, disk array set up for fast reads, high shared mem and
work mem.

If reporting is at least 1/4 of your workload, I'd suggest spinning that off
to the 2nd machine before putting one client on that machine. That way you
can also use the 2nd machine as a failover back-up.

--
Josh Berkus
Aglio Database Solutions
San Francisco

In response to

Responses

Browse pgsql-performance by date

  From Date Subject
Next Message Dave Page 2005-04-27 16:29:09 Re: Final decision
Previous Message Richard Rowell 2005-04-27 16:07:14 Suggestions for a data-warehouse migration routine