IOBench

[stephane]

Hello girls (at least Marion) and guys (at least Dimitri ;-)

Has anyone already used the IObench script delivered with STAT_Service on linux server?
It gives me no datas.

I would like to compare benchmarks carried out on a physical server and a VM.
For network I/O, i was thinking about iperf solution.
Do you know any benchmark tools for stressing disk, and why not CPU, RAM, postgres db?


Cheers
Stephane
 

[Dimitri]

Hi Stephane :-)

The IObench is shipped within dim_STAT since.. (don't remember which
version ;-)) - you may start it directly from the command line ( it'll
print you many options) - this tool was made to generate an I/O load
according various needs. The most simple way to use it is first to
generate a test scenario via web interface (the link is placed on the
very first welcome page) - it'll generate you a shell script to
execute. And then start this script when you already collecting
IObench stats from the same server - you'll be able to graph the
results live and correlate them to the "iostat" and other system
metrics. The shell script is simply executing IObench commands ( and
you may always re-edit it if needed).

For PgSQL (and databases generally) - there is dbSTRESS. Working same
way via web interface and shell script. Details about both tools are
available directly from http://finitely.free.fr - let me know if you
need more info. BTW, all MySQL and PgSQL benchmark results presented
on my site were obtained with dbSTRESS ;-))

Rgds,
-Dimitri

> --
> You received this message because you are subscribed to the Google Groups "dim_STAT" group.
> To post to this group, send email to dim...@googlegroups.com.
> To unsubscribe from this group, send email to dimstat+u...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/dimstat?hl=en.
>

[stephane]

Thanks a lot, Dimitri

Http link doesn't work, but i will try your directives asap.

I find another way to use your software each week, that 's nice ;-))

Thanks again
Stef

[Matthieu Bordonne]

I think he meant http://dimitrik.free.fr where you can find the
db_STRESS docs http://dimitrik.free.fr/db_STRESS.html and the IObench
ones http://dimitrik.free.fr/IObench.html

Matthieu

[stephane]

Cheer Matthieu

I had looked into the doc delivered with the distro (user guide actually), but not on the web site.
It will be very helpfull

Stef

[Dimitri]

Thanks Matthieu! :-)

seems "dimitrik" in the http link was auto-replaced by check speller
and I did not pay attention :-)

Rgds,
-Dimitri

[stephane]

Hi Dimitri

It is me again.

I would have some more questions about usage of IOBench.
I am bit lost with this benchmark.

Actually, i am looking for the maximum I/O that can handle a disk.
How to do that? For writing indicator, is it the write/s or the TotalWs?
Moreover, it shoudl depend on I/O size and number of I/O processes, but i could not retrive theses parametes on the GUI.

Thanks again for your support
Stephane

[Dimitri]

Hi Stephane,

it's the main rule for every benchmark - first of all you have to
understand what exactly you want to test :-))

Regarding I/O performance:
- if your goal is to reach the max possible I/O operations/sec: you
have to use small block size (similar to transactional workload on
databases)

- if your goal to reach the max possible throughput: you have to use
big block sizes (1MB for ex., similar to DWH on databases)

- if your storage is able to process several I/O operations in
parallel: for sure you'll not reach its max performance with a single
I/O process activity, you'll need more than one :-)

Then according your goals, you're preparing your test scenario and
looking for Op/sec or MB/sec levels..

The main goal of IObench is not to execute a "fixed" workload and give
you a result.. - its goal to help to simulate the I/O activity you're
expecting to see or looking for limitations, etc. - you may even
prepare several scenarios and run them in parallel just to see if
there is any performance impact when one kind of I/O activity is
meeting another one.. In most of cases you're limited only by your
imagination ;-))

Rgds,
-Dimitri

On 6/15/11, stephane <stephane....@gmail.com> wrote:
> Hi Dimitri
>
> It is me again.
>
> I would have some more questions about usage of IOBench.
> I am bit lost with this benchmark.
>
> Actually, i am looking for the maximum I/O that can handle a disk.
> How to do that? For writing indicator, is it the write/s or the TotalWs?
> Moreover, it shoudl depend on I/O size and number of I/O processes, but i
> could not retrive theses parametes on the GUI.
>
> Thanks again for your support
> Stephane
>
>

[stephane]

Thanks Dimitri.

It is getting clearer. Your mail confirms me that i started to understand yestarerday, that is great ;-)

Moreover, I was a bit confused to see all these test cases in the GUI. But it seems the first one (for instance: RWrnd:RW-001) is the only one to take into account.

See you
Stephane

[stephane]

Hi Dimitry

I am still with my I/O disk questions ;-)
Sorry to ask you this question, but i could not find any response anywhere else:
so here is my question:
Is I/O size used within the I/O Bench test the same than the block size of ext fs?
If so, i guess the bench doesn't acces directly to the FS to be able to test different I/O size.

Moreover, i find a little bug in the iostat delivered within Stat service. Indeed, for RAID disk, it shows me a Percent busy > 100% (the OS embeded iostat is ok)

Thank you for your support
Stephane

[Dimitri]

Hi Stephane,

don't worry, no problem for questions :-)

* for block size: if you're trying to simulate a real application you
may use a block size which is used by this application.. If no - you
may try different block sizes to see which one will give you a better
performance. If you're using buffering I/O (not sync) - a "smart"
filesystem group your blocks in bigger chunks, or even use Direct I/O
bypassing cache buffer in some cases (like VxFS). If you're using EXTn
(ext3? ext4?) on Linux - you have also pay attention for the IO
scheduler you're using (cfq, noop, deadline) - depending on your linux
distro performance may vary a lot..

as well, when you use a bigger block size than FS block size,
different FS will manage it different way - some will split it by FS
blocks, some will keep initial I/O block size and manage it as a
single I/O operation.. - and this is one of the reasons to test I/O
level to understand how the things are working in your case :-))

* regarding "iostat" - it's not a bug, it's just one more example of
the famous "linux compatibility" :-)) as you may imagine, the
existing "iostat" program worked well until now on other distros, etc.
- but seems on you linux something was "improved" on the system level
making a previous stuff uncompatible.. - before it was already
happened to "mpstat", and you may see 800% CPU usage in "top" as well
:-))

if "embedded iostat" prints the same columns - just edit
/etc/STATsrv/access file and point it to the right (embedded) command.
Otherwise, you may always add it as a new Add-On and share it with
others :-))

Rgds,
-Dimitri

>>> >> >> On Fri, Jun 10, 2011 at 08:29, Dimitri <dimit...@gmail.com>

[stephane]

Thank you Dim
I need to digest all these informations now ;-)

About AddOn, i will do that asap. Moreover, indeed, i can see the same issue with % CPU on multi core server.

Cheers
Stephane

>>> >> >> On Fri, Jun 10, 2011 at 08:29, Dimitri <dimitrik.fr@gmail.com>