[SOLVED] Performance Testing - clear files mapped into memory

[dpressm]

I am trying to replicate a series of tests that involve processing large input files into an application program (Essbase) under different conditions.

To avoid having inconsistent results from having disk based files remaining in memory from the previous test and to assure that all runs are burdened with the same read from disk requirements i am using the Windows sysInternals program RamMap to clear disk file pages from ram between runs.

What would the equivalent command be for unix?

I hope I have explained this properly - I am not speaking of explicitly mapped to ramdisk files but files that (In Windows and from what i remember from unix III many years ago) are files left in mapped memory locations in case needed again or until that space is needed.

Thank you.

[Rubberman]

Yes, Linux will cache (not map) recently read files into memory, but if the file is changed on disc, then the cache will become invalidated, so the next read will pull the actual data from disc. IE, you are worrying about nothing. This may be a problem with Windows, but Linux is not so encumbered.

[Irithori]

I believe, the input file(s) are not changed in dpressm´s usecase. So disk cache would influence the tests.

The disk cache can be cleared with

sync; echo 3 > /proc/sys/vm/drop_caches

See here
http://www.kernel.org/doc/Documentation/sysctl/vm.txt

[Rubberman]

I believe, the input file(s) are not changed in dpressm´s usecase. So disk cache would influence the tests.

The disk cache can be cleared with

sync; echo 3 > /proc/sys/vm/drop_caches

See here
http://www.kernel.org/doc/Documentation/sysctl/vm.txt

Good point Irithori. I read the post too quickly to realize that he was using the same input file for multiply tests. I was assuming he was using the same file name, but different contents of that file. If the file contents (or update timestamp) didn't change, then your point is correct.

[dpressm]

Rubberman - thanks for your original reply and for confirming Irithori's "more nuanced" reply.

Yes that is myt use case each test re-reads the file. i often find that people do not realize that all modern op systems do this and wonder why they can get such differing results from seemingly equivalent runs.

Thanks again.