October 18, 2006
Clearing Linux Filesystem Read Cache?
Anybody know how to clear the Linux filesystem read cache? A good answer would be much appreciated.
I have some performance tests that are being adversely affected by the filesystem’s read cache. If it makes any difference, we’re using ext3. I don’t want to disable the cache, simply clear it between tests.
A chance conversation on my walk to BART this evening got me in mind to simply copy my files between each round of tests. This kind of works. My full test data is ~140GB, but I can get by with ~20GB for tests running during the day. It takes a while to copy, but that’s better than nothing.
Luckily, my data set compresses well (5.5:1 @ 1.5GB/m), so it’s possible to blow away even the full set and restore. I haven’t tried that yet – hopefully the cache is doing a simple “has this inode/block changed recently?” test, and copying the same data back over a cached block still invalidates the cache.
That’s how I’d do it. Actually, I’d implement it using IDirectConnectionToMJsBrain to obtain my intent at every stage. But I won’t fault anybody for overlooking this.
anon said,
February 19, 2007 at 8:07 am
As of 2.6.16 you can use /proc/sys/vm/drop_caches to clear cache.
echo 1 > /proc/sys/vm/drop_caches
Will clear all the fs cache. Do a ‘sync’ before.
Andrew Z said,
March 30, 2008 at 2:47 pm
Anon’s common seems dangerous. Here’s another option
http://wiki.services.openoffice.org/wiki/Cold-start-simulator
jm said,
September 30, 2008 at 12:51 am
You can create a large file which use all filesystem cache and then remove it :
dd if=/dev/zero of=file bs=1024 count=6000000
rm file
In my case I create a 6Go file, it depends on system configuration :
more /proc/meminfo
Sandor said,
November 25, 2008 at 9:03 am
Writing to /proc/sys/vm/drop_caches should not be dangerous. The reason for doing a sync first is so more caches can get dropped.