Обсуждение: importance of fast disks with pg

Hi

During the somes I did I noticed that it does not necessarily seem to be
true that one needs the fastest disks to have a pg system that is fast.

It seems to me that its more important to:
- choose the correct methods to use for the operation
- tune the pg memory settings
- tune/disable pg xlog/wal etc

It also seems to me that fast disks are more important for db systems of
the OLTP type applications with real concurrency of both readers and
writes across many, possibly larger, tables etc.

Are the above statements close to having any truth in them?

regards

thomas

Thomas Finneid wrote:
> Hi
>
> During the somes I did I noticed that it does not necessarily seem to be
> true that one needs the fastest disks to have a pg system that is fast.
>
> It seems to me that its more important to:
> - choose the correct methods to use for the operation
> - tune the pg memory settings
> - tune/disable pg xlog/wal etc
>
> It also seems to me that fast disks are more important for db systems of
> the OLTP type applications with real concurrency of both readers and
> writes across many, possibly larger, tables etc.
>
> Are the above statements close to having any truth in them?
>
> regards
>
> thomas

I'd say that "it depends".  We run an OLAP workload on 350+ gigs of database on
a system with 64GB of RAM.  I can tell you for certain that fetching non-cached
data is very sensitive to disk throughput!

Different types of workloads will find different bottlenecks in the system..

-Dan

Thomas Finneid wrote:
> During the somes I did I noticed that it does not necessarily seem to be
> true that one needs the fastest disks to have a pg system that is fast.
>
> It seems to me that its more important to:
> - choose the correct methods to use for the operation
> - tune the pg memory settings
> - tune/disable pg xlog/wal etc
>
> It also seems to me that fast disks are more important for db systems of
> the OLTP type applications with real concurrency of both readers and
> writes across many, possibly larger, tables etc.
>
> Are the above statements close to having any truth in them?

It depends.

The key to performance is to identify the bottleneck. If your CPU is
running at 50%, and spends 50% of the time waiting for I/O, a faster
disk will help. But only up to a point. After you add enough I/O
capability that the CPU is running at 100%, getting faster disks doesn't
help anymore. At that point you need to get more CPU power.

Here's the algorithm for increasing application throughput:

while throughput is not high enough
{
  identify bottleneck
  resolve bottleneck, by faster/more hardware, or by optimizing application
}

--
  Heikki Linnakangas
  EnterpriseDB   http://www.enterprisedb.com