On Thu, 23 Oct 2003, Mark Kirkwood wrote:
>
> scott.marlowe wrote:
>
> >
> >OK, but here's the real test. As the postgres user, run 'pgbench -i',
> >then after that runs, run 'pgbench -c 50 -t 1000000'. While it's running
> >and settled (pg aux|grep postgres|wc -l should show a number of ~54 or
> >so.) pull the plug. Wait for the hard drives to spin down, then plug it
> >back in and power it one. With SCSI you will still have a coherent
> >database.
> >
> >
> Agreed in principle - pgbench is the most interesting test... for this
> mailing list anyway :-).
> However s = 1 makes a tiny database that fits into the file buffer cache
> on most machines, which is not a very realistic situation.
>
> e.g. the Dell gets tps = 250 for s = 1 c = 5 t = 1000. This number
> looks great but its not too much to do with IO....
>
> I am happier about s = 10 - 50 for machines with 512+ Mb of RAM.
>
> From memory the Dell gets tps = 36 for s = 10 c = 5 t = 100000. This
> result seems more believable!
You missed my point there. I wasn't CARING what kind of numbers you get
back at all. My point was that if you place the database under fairly
high transactional load, and pull the plug, is the database still coherent
when it comes back up.
I generally test with -s10 through -s50, but for this test it makes no
difference I can see, i.e. if the thing is gonna get scrammed at -s50,
it'll get scrammed at -s1 as well, and take less time to test.
> >If you want a coherent database on IDE drives under postgresql you will
> >need to issue this command: 'hdparm -W0 /dev/hdx' where x is the letter of
> >the drives under the RAID array to turn off write caching. This will slow
> >them to a crawl on writes.
> >
> >
> I should have said that I was using Freebsd 4.8 with write caching off.
> The question of whether the disk *actually* turned it off is the
> significant issue, so yes, "use with care" should preface any comments
> about IDE usage!
-- NOTE in a correction Mark stated that caching was on, not off --
Assuming that the caching was on, I'm betting your database won't survive
a power plug pull in the middle of transactions like the test I put up
above.