On Mon, May 17, 2010 at 3:52 AM, Piotr Legiecki <piotrlg@ams.edu.pl> wrote:
> Scott Marlowe pisze:
>
>>>> So still I don't get this: select * from table; on old server takes 0,5
>>>> sec,
>>>> on new one takes 6sec. Why there is so big difference? And it does not
>>>> matter how good or bad select is to measure performance, because I don't
>>>> measure the performance, I measure the relative difference. Somwhere
>>>> there
>>>> is a bottleneck.
>>>
>>> Yep, the network I'd say. How fast are things like scp between the
>>> various machines?
>
> Sure it is, but not in a way one could expect:
> - scp from 1000Gbit laptop to old server 27MB/sec
> - scp from the same laptop to new server 70MB/sec
> Both servers have 1000Gbit connection. So it is still mysterious why old
> server makes 9x faster select?
> I don't claim that something is slow on new (or even older) server. Not at
> all. the application works fine (still on older machine). I only wonder
> about those differences.
Is one connecting via SSL? Is this a simple flat switched network, or
are these machines on different segments connected via routers?
>>>> 4. Machine. The new server has 5 SAS disks (+ 1 spare), but I don't
>>>> remember
>>>> how they are set up now (looks like mirror for system '/' and RAID5 for
>>>> rest
>>>> - including DB). size of the DB is 405MB
>>>
>>> Get off of RAID-5 if possible. A 3 Disk RAID-5 is the slowest
>>> possible combination for RAID-5 and RAID-5 is generally the poorest
>>> choice for a db server.
>
> Sure I know that RAID-5 is slower than mirror but anyway how much slower?
> And for sure not as much as single ATA disk.
Actually, given the amount of read read / write write RAID5 does, it
can be slower than a single drive, by quite a bit. A mirror set only
reads twice as fast, it writes the same speed as a single disk.
RAID-5 is antithetical to good db performance (unless you hardly ever
write).
>
>> I refer you to this classic post on the subject:
>> http://www.mail-archive.com/pgsql-general@postgresql.org/msg93043.html
>
> Well, this thread is about benchmarking databases (or even worse, comparison
> between two RDBMS). I'm not benchmarking anything, just compare one factor.
That was a mis-post...