Re: poor performance
2006-08-21 15:47:30
>> On Mon, 21 Aug 2006 08:43:26 -0700, "Gill, Geoffrey L." <GEOFFREY.L.GILL AT
>> SAIC DOT COM> said:
> To be more clear. The 4 D40's are a mixture of database, logs and
> disk pools. One VG for database, one for mirrored database and one
> for data storage pools.
Mkay; good. :)
What size disks? were these the same disks as on your previous setup?
When I moved from 9G to 18G spindles for databases, I found that
having DB volumes fill the spindle blew my performance for ( I
inferred ) contention reasons.
You said 120GB; I'll presume that's available space. Even with 36G
SSA, that's just 4 volumes, 4 threads with which DB work can be done.
For comparison's sake, with 316GB of aggregate available databse
space, I have 35 volumes defined, mostly on 18G spindles.
Now, I'm in a very different environment, 11 servers on that piece of
hardware. But even so, you get the sense. My largest single DB,
~70G, has 8 (mirrored) volumes.
I think it possible that you will see performance improvement if you
merely cut your DB volume size. Don't go nuts, there's certainly a
performance trainwreck at the other side of the scale, too: dozens of
DB vols per spindle is Right Out. But you might consider 2 or 3 if
you're at 36G, maybe even more if you're at 72.
> Four full D40's just for a database would be quite unnecessary.
Agreed, though I wasn't visualizing that; I was visualizing one JBOD
RAID (raid-0?) with LVs carved out for whatever purposes.
> No raid anywhere, all raw disk consisting of one LV per disk no
> matter if it is database, log or disk pool related.
Digression: strongly suggest RAID for the data pools. 5-disk RAIDs
seem to be good performance for SSA, and then you have a hot spare
1/drawer.
Pedantic comment: If you're doing LVM with unraided disks, I'm not
sure JBOD applies. Isn't that a RAID term-of-art?
Gratuitous label: things come in threes.
- Allen S. Rout
|
|
|