Re: [pgsql-novice] Daily digest v1.2311 (11 messages)

At 06:46 AM 10/17/2008, Jasmin Dizdarevic wrote:
>Hi again,
>
>we like to build a data warehouse and i would like to read your
>thoughts about the hardware we plan to use.
>for us the most important thing is performance.

Really? Is performance more important than the safety and integrity
of your data? Is performance more important than the reliability and
robustness of your system? It would be a _very_ rare situation in
the DB world if you answered "yes" to either question.

>IBM x3655 2 x AMD Opteron QuadCore 2.3 Ghz L2: 2 MB
>64 GB of ram

One can not make the best HW decisions until one has analyzed what
the application the HW is going to support is. Characterize your DB
app before considering potential HW. Also, you need to let us in on
the expected usage scenario before we can give you responsible advice
on what the best HW is for you and your DB app. :-)

For instance, currently which CPU is best for a DB server is very
dependent on the DB app under consideration. For some, AMD's better
memory IO wins. For others, Intel's (much) larger on-die caches and
higher available CPU clock rates win. Considering which platform is
going to be better for you can only be done if we know some details
about the expected behavior of the DB app in question.

Another thing. In the DB world, the general rule-of-thumb is "the
more RAM, the better". But there really is an optimal amount given a
specific DB. What happens with too little RAM is obvious. But
unless you can buy enough to fit your entire DB into RAM, there is
some point where buying more RAM is less effective than spending
money on other resources. Is 64GB a good amount for _your_ DB
server? We can't tell without some details about the DB in question.

>IBM System Storage EXP3000
>with 7 x 146 GB 15k SAS HDDs (8 MB Cache)
>
Why 15K rpm SAS disks? These make sense if your are doing a lot of
transaction processing (because random IO rate, IOW seek time, really
matters for such applications). OTOH, for anything but the most
transaction heavy applications, acceptable performance can often be
had by using 10K or even 7.2K rpm HDs.

Often one can buy more HDs for the same money if one buys something
other than the 15Ks... ...and many apps will perform better with more
spindles, even slower ones, than with fewer 15K spindles.

Often you can literally buy 3x - 4x the 7.2K SATA HDs that you can
15K SAS or FC HDs of the same capacity for the same money.
Many, many apps are going to perform better on a 22 or 28 7.2K rpm HD
subsystem than they will on a 7 15K HD subsystem.

If performance matters to you, get as much battery backed cache as
you possibly can. The price/performance ratio on this is well worth it.

For example, Areca cards can be configured with gigabytes of battery
backed cache.

>Do you see any problems with the L2-Cache or the HDD-Cache? Are they
>important for pg?
>Have you any suggestions for the RAID-Controller?
>
>We plan to set up the whole thing on gentoo. A good idea?
>
Potential maintenance nightmare. You better have a real linux guru
on staff and feel comfortable with the idea of compiling your entire
OS from source from scratch. OTOH, if done right, a gentoo distro
custom compiled for your HW will likely be one of the best performing
distros. ?IF? you do EVERYTHING right.

For most circumstances, there are better open source distros for a DB server.

>Thank you very much for you'r replies.
>
>Jasmin

Hope this helps,
Ron Peacetree