On Wed, Apr 22, 2009 at 7:22 AM, Jens Axboe <jens.axboe@xxxxxxxxxx> wrote:
> Hi,
>
> The sample job files shipped with fio are (generally) pretty weak, and
> I'd really love for the selection to be better. In my experience, that
> is the first place you look when trying out something like fio. It
> really helps you getting a (previously) unknown job format going
> quickly.
>
> So if any of you have "interesting" job files that you use for testing
> or performance analysis, please do send them to me so I can include them
> with fio.
Jens,
I normally use scripts to run I/O benchmarks, and pretty much use fio
exclusively.
Hopefully, in sharing the scripts, you can see usage, and feeback
anything I may be doing wrong.
In one incarnation, I put all the devices to be tested on the script's
command line, then concatenate a fio-ready list of these devices along
with a sum of 10% of all the disks with:
filesize=0
fiolist=""
for i in $*
do fiolist=$fiolist" --filename="$i
t=`basename $i`
let filesize=$filesize+`cat /proc/partitions | grep $t | awk
'{ printf "%d\n", $3*1024/10 }'`
done
Rather than a "job file", In this case I do everything on the command
line for power of 2 block sizes from 1MB down to 512B:
for i in 1m 512k 256k 128k 64k 32k 16k 8k 4k 2k 1k 512
do
for k in 0 25 50 75 100
do
fio --rw=randrw --bs=$i --rwmixread=$k --numjobs=32
--iodepth=64 --sync=0 --direct=1 --randrepeat=0 --softrandommap=1 \
--ioengine=libaio $fiolist --name=test --loops=10000
--size=$filesize --runtime=$runtime
done
done
So the above "fiolist" is going to look like "--filename=/dev/sda
--filename=/dev/sdb", and the "filesize" is going to be the sum of 10%
of each disks size. I only use this with disks of the same size, and
assume that fio will exercise 10% of each disk. That assumption seems
to pan out in the resulting data, but I've never traced the code to
verify that this is what it will do.
Then I moved to a process-pinning strategy that has some number of
pinned fio threads running per disk. I still calculate the
"filesize", but just uses 10% of one disk, and assume they are all the
same. Much of the affinity settings have to do with specific bus-CPU
affinity, but for a simple example, lets say I just round-robin the
files on the command line to the available processors, and create
arrays "files" and "pl" consisting of block devices and processor
numbers:
totproc=`cat /proc/cpuinfo | grep processor | wc -l`
p=0
for i in $*
do
files[$p]="filename="$i
pl[$p]=$p
let p=$p+1
if [ $p -eq $totproc ]
then break
fi
done
let totproc=$p-1
Then generate "job files" and run fio with:
for i in 1m 512k 256k 128k 64k 32k 16k 8k 4k 2k 1k 512
do
for k in 0 25 50 75 100
do echo "" >fio-rand-script.$$
for p in `seq 0 $totproc`
do
echo -e
"[cpu${p}]\ncpus_allowed=${pl[$p]}\nnumjobs=$jobsperproc\n${files[$p]}\ngroup_reporting\nbs=$i
\nrw=randrw\nrwmixread=$k \nsoftrandommap=1\nruntime=$runtime
\nsync=0\ndirect=1\niodepth=64\nioengine=libaio\nloops=10000\nexitall\nsize=$filesi
e \n" >>fio-rand-script.$$
done
fio fio-rand-script.$$
done
done
The scripts look like:
# cat fio-rand-script.8625
[cpu0]
cpus_allowed=0
numjobs=8
filename=/dev/sda
group_reporting
bs=4k
rw=randrw
rwmixread=0
softrandommap=1
runtime=600
sync=0
direct=1
iodepth=64
ioengine=libaio
loops=10000
exitall
size=16091503001
[cpu1]
cpus_allowed=1
numjobs=8
filename=/dev/sdb
group_reporting
bs=4k
rw=randrw
rwmixread=0
softrandommap=1
runtime=600
sync=0
direct=1
iodepth=64
ioengine=libaio
loops=10000
exitall
size=16091503001
[cpu2]
cpus_allowed=2
numjobs=8
filename=/dev/sdc
group_reporting
bs=4k
rw=randrw
rwmixread=0
softrandommap=1
runtime=600
sync=0
direct=1
iodepth=64
ioengine=libaio
loops=10000
exitall
size=16091503001
[cpu3]
cpus_allowed=3
numjobs=8
filename=/dev/sdd
group_reporting
bs=4k
rw=randrw
rwmixread=0
softrandommap=1
runtime=600
sync=0
direct=1
iodepth=64
ioengine=libaio
loops=10000
exitall
size=16091503001
I would sure rather do that on the command line and not create a file,
but the groups never worked out for me on the command line... hints
would be appreciated.
Thanks,
Chris
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
