Our Latest Posts: -  Passive Income Smooths the Flow  -  Hitting the Google Sweet Spot  -  Vista Boomers - Microsoft's Power Failure  -  Book Review: Programming in Objective-C 2.0 (Second Edition)

Translators wanted!

---

Best of the Newsgroups: flushing linux buffer cache

---

What is this stuff?

Main Index

From: "Tony  Lawrence" <pcu...@gmail.com>
Subject: Re: Is there a way to flush the file cache?
Date: 29 Dec 2006 05:26:25 -0800
Message-ID: <1167398785.047893.37730@73g2000cwn.googlegroups.com> 
References: <4vig1lF1b4vglU18@mid.individual.net> 


Jean-David Beyer wrote:
> Tony Lawrence wrote:
> > General Schvantzkoph wrote:
> >> On Thu, 28 Dec 2006 13:29:24 -0800, Tony  Lawrence wrote:

> >>> General Schvantzkoph wrote:
> >>>> I'm writing a system exerciser that checks memory and disk I/O. Is there a
> >>>> C call that flushes the file cache so that I can be assured that my freads
> >>>> are coming from the disk and not the RAM. I'm using a sync() but I don't
> >>>> think that does the whole job, it guarantees that the disk writes happen
> >>>> but I don't think it clears out the file cache.

> >>> Well, of course it can't guarantee that some random app hasn't written
> >>> data, if that's what you mean..

> >> What I want to do is the following,

> >> 1) Create a file and write it to the disk.
> >> 2) Read that file off of the disk and then check it.

> >> I'm trying to check the SATA interfaces so it's important that the file go
> >> to and from the disks. The file cache interferes with this because if the
> >> file is present in the cache the program will get the cached copy not the
> >> copy that is on the disk. The sync() call forces the disk writes to happen
> >> immediately which is half of the battle. The other half is to purge the
> >> file cache so that all subsequent file reads come from the disk. Is there
> >> a way to do this?

> > O, I see now:  you are checking disk performance or maybe even checking
> > disk integrity.

> > No, sync won't help with that (because the read can still come from
> > cache).  What you need to do is overwrite the cache with new data
> > unrelated to the blocks you want to read.  That could be as simple as a
> > well placed "ls -lR".. or much more complex.  See
> > http://www.linuxjournal.com/article/7105

> On my machine, I have 8 GBytes RAM, and usually over 3/4 of it are used by
> IO buffers and cache. It would take a lot to flush all that. In fact, sync
> does not do it.

Right.  A "sync" is a request, it doesn't mean "stop everything else
you are doing, write all your buffers to disk and get back to me when
you are done".

> >
> > Apparently newer kernels have made this easier:
> > http://linux-mm.org/Drop_Caches

And of course that would require a sync first.  I don't know how this
is implemented, but I would guess that it doesn't actually throw away
the cache until sync or a cache overwrite allows it; it probably just
marks the cache as invalid so that subsequent reads would have to go to
the source.  I  assume that something somewhere ties this all together
so that if you have synced and then dropped, your reads will get
correct data, but I don't know that.  It would be simple enough to test
in a VM with artificially constrained memory (or somebody with more
energy than I have could paw through the source and see how it works).

-- 
Tony Lawrence
Unix/Linux/Mac OS X Resources
http://aplawrence.com