Re: Best hardware

Поиск
Список
Период
Сортировка
От Mischa Sandberg
Тема Re: Best hardware
Дата
Msg-id 1117927417.42a237f92b0f7@webmail.telus.net
обсуждение исходный текст
Ответ на Best hardware  ("Bernd Jagla" <baj2107@columbia.edu>)
Список pgsql-performance
Quoting Bernd Jagla <baj2107@columbia.edu>:

> ... the speed of the head of the HD is actually
> limitiing. Also, I only experimented with RAID5, and heard that
> RAID10 will be good for reading but not writing.

Au contraire. RAID5 is worse than RAID10 for writing, because it has the
extra implicit read (parity stripe) for every write. I've switched all
my perftest boxes over from RAID5 to RAID10, and the smallest
performance increase was x1.6 . This is in an update-intensive system;
the WAL log's disk write rate was the controlling factor.

> Are Quad CPUs any good? I heard that the IBM quad system is supposed
to be 40%
> faster than HP or Dell???.
Check out the other threads for negative experiences with Xeon 2x2 and
perhaps quad CPU's. Me, I'm looking forward to my first Opteron box
arriving next week.

> And how much RAM should go for: are 8GB enough? Oh, of course I wanted
to run it under RedHat...

First off, you need enough RAM to hold all your connections. Run your
app, watch the RSS column of "ps". For my own simpler apps (that pump
data into the db) I allow 20MB/connection.

Next, if you are heavy on inserts, your tables will never fit in RAM,
and you really just need enough to hold the top levels of the indexes.
Look at the disk space used in your $PGDATA/base/<dboid>/<tableoid>
files, and you can work out whether holding ALL your indexes in memory
is feasible.

If you are heavy on updates, the above holds, but ymmv depending on
locality of reference, you have to run your own tests.

If you have concurrent big queries, all bets are off --- ask not how
much RAM you need, but how much you can afford :-)



В списке pgsql-performance по дате отправления:

Предыдущее
От: "Bernd Jagla"
Дата:
Сообщение: Best hardware
Следующее
От: "Qingqing Zhou"
Дата:
Сообщение: Re: Forcing use of specific index