DragonFly kernel List (threaded) for 2011-07
Re: Blogbench RAID benchmarks
On Mon, Jul 11, 2011 at 10:23:33AM -0700, Matthew Dillon wrote:
> :The results were taken for one run only, they were not averaged.
> :Howewer, I repeated many times the tests for the RAID10 and RAID0 cases, and
> :the numbers were always consistent (7% variation at worst).
> :No fancy disk scheduler was used. Besides newfs, defaults were used for
> :all commands.
> Also, what blogbench command were you running?
The default: "blogbench -d /mnt/blogbench"
> To get real results with
> blogbench you have to remember that blogbench creates an ever-growing
> data set and if the run isn't long enough the data set may not blow out
> the buffer cache in one test, and blow it out in another, producing
> radically different results.
There was some variation but not that much; Sascha suggested I also test
FreeBSD on the same hardware and it was a damn good idea.
The FreeBSD write performance curve follows pretty well the theoretical
performance increase of the different RAID sets.
> I usually set --iterations=100 to force the blogbench data set to be
> large enough to actually blow out the buffer cache.
Ok. If I get enough time, I'll do a third run with this parameter.
> You should include the raw blogbench output as well. It's a wall of text
> but it is important because read activity can be very deceptive if a
> filesystem bogs down on writes (because only writes expand the size of
> the data set being read). This makes the final numbers a bit problematic,
> requiring additional analysis to really understand what is going on.
I'll keep the numbers in the future.