build out of a server advice

Nicholas Leippe nick at leippe.com
Sat Jun 23 10:27:08 MDT 2012


IMO http://www.baarf.com/
it links this article from Oracle talking about "S.A.M.E.":
http://www.miracleas.com/BAARF/oow2000_same.pdf

Disk is cheap. Raid 10 in software or hardware will outperform Raid5
and let you sleep easier at night. The main tradeoffs between 6 and
1+0 are:

Raid6:
- can withstand any combination of double disk fault
- much worse healthy write performance than 1+0
- huge performance penalty in degraded and rebuild mode

Raid1+0:
- can only withstand specific combinations of double disk fault (must
not be two pairs of the same mirror)
- minimal healthy write performance penalty
- performance hardly affected in degraded mode, rebuild mode is a
simple copy--no CPU calculations, just I/O

The other issue with large drives (where large >= 4TB), is which file
system you are going to use and what are its fault tolerances. You do
not want to do fsck on a 4TB volume--a quick (slow) way to make five
9s go to one.

If you're using linux sw raid, the md raid10 driver can achieve raid0
read performance in some cases (with layout=far), this is at the cost
of a higher penalty while degraded and rebuilding.

If you're using linux sw raid with SSD storage, the raid10 driver will
achieve raid0 performance (w/any layout), and no penalties. Of course,
1TB SSDs are a bit pricey still--but for one or two DB servers in many
cases throwing some money at SSDs can get you better performance gains
than scaling horizontally. I'm personally a fan of OCZ's revodrive
PCIe products--over 100,000 IOPs each, and at a half to a quarter of
the price of fusionio's stuff.


More information about the PLUG mailing list