sysbench / fileio - Linux vs. FreeBSD
Matthew Dillon
dillon at apollo.backplane.com
Sun Jun 6 21:18:04 UTC 2010
:> It would be interesting to see a blogbench comparison between UFS
:> and ZFS on the same hw/disk.
:
:
:I'll do it, just tell me how do you want to run the tests.
:
:The system params are:
:
:8GB Memory
:2x72GB SCSI HDD
:2x3.4Ghz Xeon
:Overall: Dell Poweredge 1850. With no raid installed.
:
:I'm waiting the benchmark options to run.
:
:- --
:Adam PAPAI
With 8G of ram blogbench should blow out the system caches at around
blog 1000-1600, though it also depends on the maximum number of
vnodes supported by the system. One of the two (VM pages or vnode
limit) will be hit.
All you need to do is run blogbench with enough iterations to ensure
that the run eventually blows out the system caches. 200 or 300 should
do the job. It's easy to tell when the system cache gets blown out from
looking at the output.
Run something like the following script for a few hours. You want to
get at least four full runs under your belt for each filesystem to
factor out edge cases.
For the filesystem setup it would be cool to test both the single-drive
case and a simple non-redundant interleaved or mirrored setup (double
the read bandwidth). With UFS use default parameters with softupdates
turned on (I'd say also without SUJ). With ZFS I don't know how best
to tune it, try to find a ZFS setup that performs decently.
However, be sure to turn off compression and dedup (if those fs options
are available), because blogbench basically just writes all-zeros which
is highly compressable/collapsable and would skew the results badly.
-Matt
#!/bin/csh
#
# /build is the filesystem under test.
set i = 0
while(1)
set name = `printf "bench%05d" $i`
echo $name
if ( ! -d /build/blogs/$name ) then
mkdir -p /build/blogs/$name
blogbench --iterations=200 -d /build/blogs/$name
sleep 120
endif
@ i = $i + 1
end
More information about the freebsd-hackers
mailing list