Re: [PERFORMANCE] Buying hardware - Mailing list pgsql-performance

From david@lang.hm
Subject Re: [PERFORMANCE] Buying hardware
Date
Msg-id alpine.DEB.1.10.0901261929100.16162@asgard.lang.hm
Whole thread Raw
In response to Re: [PERFORMANCE] Buying hardware  (Matthew Wakeling <matthew@flymine.org>)
List pgsql-performance
On Mon, 26 Jan 2009, Matthew Wakeling wrote:

> On Sun, 25 Jan 2009, Scott Marlowe wrote:
>> RAID-10 is almost always the right choice.
>
> Agreed. Unless you don't care about the data and need the space, where RAID 0
> might be useful, or if you really don't need the space, where RAID 1 might be
> okay. If your controller supports it.

if you can reproduce your data at will (would not mind loosing it) you can
do a single disk or raid 0 (and you can disable fsync for even more
performance)

   for this you need N disks (where N is the number needed to hold your data)


if you cannot reproduce your data at will (or it takes too long)

if you need the capacity of a single disk do raid 1

if you need the capacity of a small number of disks do raid 10 (raid 1
combined with raid 0 to stripe redundant copies of data across multiple
disks)

if you need the capacity of a large number of disks you need to seriously
think about your performance needs. the fact that raid 10 can loose data
if the wrong 2 disks fail and requires buying 2 disks for every disk worth
of capacity that you need are both factors. (note that some raid 10
implimentations let you have more than 2 copies of your data, but your
disk requirements go up # copies * capacity)

   for these you need N*M disks (where N is the number needed to hold your
data and M-1 is the number of disks you can loose without loosing any
data)

at some point it may make sense to use raid 6 for some data. It is
definantly slower, but you can loose two drives and not loose any data
while only needing N+2 drives

to further complicate matters, some parts of your database are more
sensitive to performance than others.

the fsync throughput of the device you have the WAL on will determine the
max transactions/sec of your system (and any seeking that this disk needs
to do for other purposes will hurt this)

two other performance sensitive areas are temporary table space and
indexes
   question, are these more seneitive to random or sequential performance?

see the recent discussions and links to performance ratings of different
drive types in the thread 'SSD performance' unfortunantly the SSD drive
types so overwelm the normal drives that it's hard to see the differences
in the graphs between the 15K rpm SCSI/SAS drives and the 7200 rpm SATA
drives, but they are there.

David Lang



pgsql-performance by date:

Previous
From: Craig Ringer
Date:
Subject: Re: [PERFORMANCE] Buying hardware
Next
From: david@lang.hm
Date:
Subject: Re: [PERFORMANCE] Buying hardware