f



Raw device VS cooked file - any benchmark recently done?

Hi everybody,

a question about raw VS cooked once again (sorry about that), hopefully
this time a bit different.

I am using IDS 10.0 on Red Hat Enterprise Linux version 4, as well as on Solaris
8 and 9. I use Informix since years, and always have installed it making use
of raw devices, since there always have been a clear statement in the IBM
documentation telling that on UNIX raw devices are the suggested choice for
the best I/O performance. The IBM IDS 10 Server Administrator Guide,
http://publib.boulder.ibm.com/epubs/pdf/25122672.pdf, at page 241 says
infact:
---------------------------------------------------------------------------
<<On UNIX, you should use raw disk devices to store data whenever
performance is important.>>
---------------------------------------------------------------------------

However, looking further in the same document, at page 242, I have noticed that
for the first time there is also this new and interesting statement:
---------------------------------------------------------------------------
<<Important: While you should use raw disk devices on UNIX to achieve better
performance, recent advances in I/O caching for cooked writes can provide
similar if not better performance. To determine the best device performance,
perform benchmark testing on the system with both types of devices for the
dbspace and table layout.>>
---------------------------------------------------------------------------
This last sentence was not included in the previous versions of the
documentation (checked for IDS 9.4).

I have been googling around, finding out on the IIUG repository that the
measured performance of raw VS cooked VS cooked on filesystem is roughly X,
X-20%, X-35% respectively. Source of these figures:
http://www.iiug.org/resources/cdi_archive/2005.05/cdii.137120

I was wondering if anyone have recently done a benchmark using IDS 10 on UNIX
configured with raw areas versus the same engine configured to use cooked files
over filesystem, where the cooked file is stored on a device where "I/O caching
for cooked writes can provide similar if not better performance".

In other words, I would like to understand if that is really true and
what feature should the disk device or SAN have in order to meet the
requirements of being good in "I/O caching for cooked writes".

One last thing I don't understand is if, in case there is a device capable
to be good in "I/O caching for cooked writes", it would be equally good in
cooked READS.

Any feedback would be appreciated.

Thanx in advance,
   Rupan3rd (from Italy)
0
Rupan3rd
1/10/2006 2:50:16 PM
comp.databases.informix 16081 articles. 0 followers. Post Follow

3 Replies
615 Views

Similar Articles

[PageSpeed] 53

At the Southeast Informix User Group, I sat through a presentation that
covered Linux. The presenter said that with the newest Kernel it no
longer mattered if you used raw. The presentation is at this website:

http://www.iiug.org/waiug/present/Forum2005/Forum2005Sessions.html

It was called:

IDS in a Linux Environment

The presenter might even have some benchmarks.

0
bozon
1/10/2006 3:18:23 PM
I would really like to see any evidence that people have on this issue.  =
I
have a number of systems that are running with acceptable performance on =
raw
disks.  But could we be doing better and at what cost?

But from almost 20 years of working with this disk architecture and with
lots of other architectures I have discovered that the performance that =
you
can get out of a disk system is in almost direct proportion to the =
amount of
time that you, the administrator, put into it.  If you want the ultimate
performance then spend a lot of time doing the systems engineering - but =
if
you can get acceptable performance without doing much then so be it.  =
Many
modern disk architectures can give you acceptable performance without =
raw
disk - now.  But will that continue when the UNIX file system gets more
fragmented?  And what are the other reasons for using raw disk - think =
about
backup and recovery as well.

Regards

Malcolm

-----Original Message-----
From: informix-list-bounces@iiug.org =
[mailto:informix-list-bounces@iiug.org]
On Behalf Of Rupan3rd
Sent: 10 January 2006 14:50
To: informix-list@iiug.org
Subject: Raw device VS cooked file - any benchmark recently done?


Hi everybody,

a question about raw VS cooked once again (sorry about that), hopefully =
this
time a bit different.

I am using IDS 10.0 on Red Hat Enterprise Linux version 4, as well as on
Solaris 8 and 9. I use Informix since years, and always have installed =
it
making use of raw devices, since there always have been a clear =
statement in
the IBM documentation telling that on UNIX raw devices are the suggested
choice for the best I/O performance. The IBM IDS 10 Server Administrator
Guide, http://publib.boulder.ibm.com/epubs/pdf/25122672.pdf, at page 241
says
infact:
-------------------------------------------------------------------------=
--
<<On UNIX, you should use raw disk devices to store data whenever
performance is important.>>
-------------------------------------------------------------------------=
--

However, looking further in the same document, at page 242, I have =
noticed
that for the first time there is also this new and interesting =
statement:
-------------------------------------------------------------------------=
--
<<Important: While you should use raw disk devices on UNIX to achieve =
better
performance, recent advances in I/O caching for cooked writes can =
provide
similar if not better performance. To determine the best device =
performance,
perform benchmark testing on the system with both types of devices for =
the
dbspace and table layout.>>
-------------------------------------------------------------------------=
--
This last sentence was not included in the previous versions of the
documentation (checked for IDS 9.4).

I have been googling around, finding out on the IIUG repository that the
measured performance of raw VS cooked VS cooked on filesystem is roughly =
X,
X-20%, X-35% respectively. Source of these figures:
http://www.iiug.org/resources/cdi_archive/2005.05/cdii.137120

I was wondering if anyone have recently done a benchmark using IDS 10 on
UNIX configured with raw areas versus the same engine configured to use
cooked files over filesystem, where the cooked file is stored on a =
device
where "I/O caching for cooked writes can provide similar if not better
performance".

In other words, I would like to understand if that is really true and =
what
feature should the disk device or SAN have in order to meet the =
requirements
of being good in "I/O caching for cooked writes".

One last thing I don't understand is if, in case there is a device =
capable
to be good in "I/O caching for cooked writes", it would be equally good =
in
cooked READS.

Any feedback would be appreciated.

Thanx in advance,
   Rupan3rd (from Italy) _______________________________________________
Informix-list mailing list
Informix-list@iiug.org =
http://www.iiug.org/mailman/listinfo/informix-list

0
malcolm
1/10/2006 3:38:15 PM
I know what pros and cons of using raw might be. My need is to have the best
performance for my OLTP system, therefore I was wordering if and in what cases
it might be actually true that there are disk devices able to achieve "I/O
caching for cooked writes" that "can provide similar if not better performance"
than raw areas.

Still the documentation does not explicitly say if the cooked READS would be
that good as well nor how to benchmark nor the effect of the fragmentation.

Regards
   Rupan3rd


malcolm weallans wrote:
> I would really like to see any evidence that people have on this issue.  I
> have a number of systems that are running with acceptable performance on raw
> disks.  But could we be doing better and at what cost?
> 
> But from almost 20 years of working with this disk architecture and with
> lots of other architectures I have discovered that the performance that you
> can get out of a disk system is in almost direct proportion to the amount of
> time that you, the administrator, put into it.  If you want the ultimate
> performance then spend a lot of time doing the systems engineering - but if
> you can get acceptable performance without doing much then so be it.  Many
> modern disk architectures can give you acceptable performance without raw
> disk - now.  But will that continue when the UNIX file system gets more
> fragmented?  And what are the other reasons for using raw disk - think about
> backup and recovery as well.
> 
> Regards
> 
> Malcolm
> 
> -----Original Message-----
> From: informix-list-bounces@iiug.org [mailto:informix-list-bounces@iiug.org]
> On Behalf Of Rupan3rd
> Sent: 10 January 2006 14:50
> To: informix-list@iiug.org
> Subject: Raw device VS cooked file - any benchmark recently done?
> 
> 
> Hi everybody,
> 
> a question about raw VS cooked once again (sorry about that), hopefully this
> time a bit different.
> 
> I am using IDS 10.0 on Red Hat Enterprise Linux version 4, as well as on
> Solaris 8 and 9. I use Informix since years, and always have installed it
> making use of raw devices, since there always have been a clear statement in
> the IBM documentation telling that on UNIX raw devices are the suggested
> choice for the best I/O performance. The IBM IDS 10 Server Administrator
> Guide, http://publib.boulder.ibm.com/epubs/pdf/25122672.pdf, at page 241
> says
> infact:
> ---------------------------------------------------------------------------
> <<On UNIX, you should use raw disk devices to store data whenever
> performance is important.>>
> ---------------------------------------------------------------------------
> 
> However, looking further in the same document, at page 242, I have noticed
> that for the first time there is also this new and interesting statement:
> ---------------------------------------------------------------------------
> <<Important: While you should use raw disk devices on UNIX to achieve better
> performance, recent advances in I/O caching for cooked writes can provide
> similar if not better performance. To determine the best device performance,
> perform benchmark testing on the system with both types of devices for the
> dbspace and table layout.>>
> ---------------------------------------------------------------------------
> This last sentence was not included in the previous versions of the
> documentation (checked for IDS 9.4).
> 
> I have been googling around, finding out on the IIUG repository that the
> measured performance of raw VS cooked VS cooked on filesystem is roughly X,
> X-20%, X-35% respectively. Source of these figures:
> http://www.iiug.org/resources/cdi_archive/2005.05/cdii.137120
> 
> I was wondering if anyone have recently done a benchmark using IDS 10 on
> UNIX configured with raw areas versus the same engine configured to use
> cooked files over filesystem, where the cooked file is stored on a device
> where "I/O caching for cooked writes can provide similar if not better
> performance".
> 
> In other words, I would like to understand if that is really true and what
> feature should the disk device or SAN have in order to meet the requirements
> of being good in "I/O caching for cooked writes".
> 
> One last thing I don't understand is if, in case there is a device capable
> to be good in "I/O caching for cooked writes", it would be equally good in
> cooked READS.
> 
> Any feedback would be appreciated.
> 
> Thanx in advance,
>    Rupan3rd (from Italy) _______________________________________________
> Informix-list mailing list
> Informix-list@iiug.org http://www.iiug.org/mailman/listinfo/informix-list
> 
0
Rupan3rd
1/12/2006 4:18:03 PM
Reply: