Need to test read performance

[Geoff Wild]

I created a 50GB file with prealloc.

I want to test read performance - so I thought I would copy the 50GB file to /dev/null

Except /dev/null resides in a non largefiles supported file system.

Any suggestions?

I was thinking maybe just dd the lvol to /dev/null - but would that be a sufficient read test?

Thanks...Geoff

[A. Clay Stephenson · Accepted solution]

The fact that /dev/null resides in a non-largesfiles enabled filesystem should not matter.

You should be able to do something like this:

timex dd if=mybigfile bs=256k of=/dev/null


However, if you are really wantingh to determine disk read speed then I would bypass the filesystem and buffer cache and do something like:

timex dd if=/dev/vg10/rlvol1 bs=256k of=/dev/null

[OldSchool]

i don't think that /dev/null and largefiles will be an issue, you're writing to a device, and not to the file system, but I could be wrong.

for benchmarking disk performance, you might want to try iozone, downloadable from www.iozone.org

[Geoff Wild]

As Clay says - the dd is the trick.

Here's what I did:

timex dd if=/oracle/IPR/sapdata05/test.50GB of=/dev/null bs=8k


8K as that is what Oracle uses.

Rgds...Geoff

[Hein van den Heuvel]

Geoff,

[Great to have seen you in Houston two weeks ago!]

8K may be the page size for your database, and the iosize for 'sequential reads'.
However for reading big chunks of data for tablescans Oracle will read 'multi_block' pages at a time.
So maybe bs=128k may be more correct.

Ultimately you should ofcourse just have Oracle actually read the data and measure that! Using transportable tablespace it might actually be relatively easy to have a standard sample benchmark tablespace which you copy to various storage locations and read from (and or write to).

Hope this helps,
Hein van den Heuvel
HvdH Performance Consulting