First of all I like to thank everyone who responded.
The original question was:
>
> Dear group,
> is there a simple way to disable in-memory caching while testing an external
> raid-system.
> When disabled I get a better impression of the raid-system performance (I
> guess).
> Now I see very high values like > 1Gig/Second.
>
> I test a 6 disk/128MBcache Infortrend Raid-controller connected to a DS10 with
> 512 MB memory with the software from http://www.iozone.org and like to get a
> good impression of the unit.
>
The answer that gave me a hint to a partial solution was the one of
alan_at_nabeth.cxo.dec.com when he mentions the SYNC-options.
This can be triggered with the -o option and has the wanted effect on writes.
The reads and many others are not affected.
However I can get a good impression now of the performance of the
Raid-controller.
With a KZPBA-CA UWSE-adapter I get a maximum throughput between 20 and 25
KB/sec. Since the adapter is specified at a max of 40 MB/Sec I called my dealer
and he explained that they saw the same performance and that it's caused by the
Raid-5 configuration and the parity calculations. The processor is only a simple
AMD 133Mhz.
The latest controllers, equiped with a PowerPC Risc processor are much faster
and say to deliver >45MB/Sec for Raid-5.
If you like to see some of the results (simple presentation) look at
http://gissrv.iend.wau.nl/~bol/iozone/results.cgi
There are no conclusions, only some graphs.
The answers I received are below.
-----------------------John P Speno <speno_at_isc.upenn.edu>
There's a 'sync' mount option for AdvFS which says:
sync
Causes all writes to be written immediately to disk as well as to the
buffer cache.
Would that work? You could try it with and without to see if it makes a
difference in performance.
------------------------"Fliguer, Miguel" <M_Fliguer_at_unifon.com.ar>
It seems there is a -p flag on iozone that does
what you want....
>From the iozone docs :
-p
This purges the processor cache before each file operation. Iozone will
allocate another
internal buffer that is aligned to the same processor cache boundary and is
of a size that
matches the processor cache. It will zero fill this alternate buffer before
beginning each test.
This will purge the processor cache and allow one to see the memory
subsystem without
the acceleration due to the processor cache
Regards,
Miguel Fliguer - UNIFON S.A.
Buenos Aires, Argentina
------------------------alan_at_nabeth.cxo.dec.com
I believe the only way to bypass the cache on reads is
to not use a file system; use the raw device instead.
On writes there are a number of ways to force synchronous
writes:
o Mount the file system with the "-o sync" option.
o Open the file with the O_DSYNC or O_SYNC flags.
o Call fsync as needed.
The usual way to bypass caching affect on file system
read test is to read very large files. However, if
your application doesn't actually read files that
large, then the caching affect may be informative.
Unmounting the file system between tests will invalidate
the cache, but that's sometime inconvient.
------------------------Niels Roetert <Niels.Roetert_at_pqr.nl>
Je kunt spelen met de ubc-borrowpercent ubc-maxpercent
en ubc-minpercent deze waardes kunnen worden aangepast
via /etc/sysconfigtab
Alle gegevens staan in Digital UNIX performance tuning manual
online op de Compaq site.
------------------------anthony.miller_at_vf.vodafone.co.uk
If your external raid system is an HSZ controller, then do a SET UNIT Dnnn
NOWRITEBACKCACHE on the units you are testing. This will give you the
actual performance of the raid array architecture by disabling its onboard
cache. I would suggest you probably should keep any unix caching enabled
for your tests.
I did a little bit of this sort of thing some while ago and drew some graphs
of performance with and without writeback cache enabled - and on raid-5 sets
with and without a failed member - with and without caching enabled etc.
Quite interesting actually.
------------------------Alex Gorbachev <alex_at_iss-integration.com>
I would use dd for IO testing. dd will not cache anything in OS memory.
-------------------------
Received on Fri Mar 24 2000 - 12:08:30 NZST