Linux – Drop a specific file from the Linux filesystem cache

cachefilesfilesystemslinux

I know I can drop everything from the Linux filesystem cache, but is there a way to drop just one specific file? Or prevent a file from being cached? Or tell a process not to cache any files it writes?

I have a process that reads a lot of little files and writes a big file. I want to keep the little files in cache to avoid disk seeks, and I don't care about caching the big file.

Best Answer

Potential Method #1 - F_DROP_CACHES

I found a method from 2012 that discusses a proposed patch to the Linux kernel in this mail thread titled: Re: [RFC Patch] fs: implement per-file drop caches.

excerpt

Cong> This is a draft patch of implementing per-file drop caches.

Interesting. So can I do this from outside a process? I'm a SysAdmin, so my POV is from noticing, finding and fixing performance problems when the system is under pressure.
Cong> It introduces a new fcntl command  F_DROP_CACHES to drop  
Cong> file caches of a specific file. The reason is that currently  
Cong> we only have a system-wide drop caches interface, it could  
Cong> cause system-wide performance down if we drop all page caches  
Cong> when we actually want to drop the caches of some huge file.
How can I tell how much cache is used by a file? And what is the performance impact of this when run on a busy system? And what does this patch buy us since I figure the VM should already be dropping caches once the system comes under mem pressure...

Cong> Below is small test case for this patch:

The thread includes both a testcase and the actual patch to several files within the Linux kernel which adds an additional function to fs/drop_caches.c called drop_pagecache_file(struct file *filp). This function is then accessible through the frontend tool, fnctl.c via the command F_DROP_CACHES. This case calls this function:

file_drop_caches(filp, arg);

Which handles the dropping of all the caches associated with the given file. From the file include/linux/mm.h:

void file_drop_caches(struct file *filp, unsigned long which);

So this can be used?

I found no evidence that this patch ever made its way into the main Linux kernel code repository, so this option would appear to be available, only if you're willing to recompile the Linux kernel yourself.

Potential Method #2 - Using dd

In that same thread, another user mentions a completely different methodology that makes use of dd.

The following is excerpt from that email

This is useful functionality. Though isn't it already provided with POSIX_FADV_DONTNEED? This functionality was added to GNU dd (8.11) a year ago.
Here are the examples from that patch:
Advise to drop cache for whole file
 $ dd if=ifile iflag=nocache count=0
Ensure drop cache for the whole file
 $ dd of=ofile oflag=nocache conv=notrunc,fdatasync count=0
Drop cache for part of file
 $ dd if=ifile iflag=nocache skip=10 count=10 of=/dev/null
Stream data using just the read-ahead cache
 $ dd if=ifile of=ofile iflag=nocache oflag=nocache

Testing it out

I wasn't 100% positive how to test this out but I came up with the following approach.

make a 100MB file

$ dd if=/dev/urandom of=sample.txt bs=100M count=1

trace file accesses using fatrace
```
$ sudo fatrace | grep sample.txt
```
run top so we can monitor memory usage, note amount free.
```
$ top
```
open file, note amount of free memory now. Note the fatrace of the file sample.txt.
```
$ cat sample.txt > /dev/null
```

drop the file from memory, note amount of free memory now. Note the output of fatrace.

$ sudo dd of=/home/saml/tst/162600/sample.txt \
    oflag=nocache conv=notrunc,fdatasync count=0

Example

In terminal #1:

$ dd if=/dev/urandom of=sample.txt bs=100M count=1
1+0 records in
1+0 records out
104857600 bytes (105 MB) copied, 7.37996 s, 14.2 MB/s

$ ls -l sample.txt 
-rw-rw-r--. 1 saml saml 104857600 Oct 17 22:54 sample.txt

In terminal #2:

$ top
...
KiB Mem:   7968336 total,  6900956 used,  1067380 free,   267080 buffers
...

In terminal #3:

$ sudo fatrace | grep sample.txt

Now open the file, sample.txt, and note the amount of RAM. In terminal #1.

$ cat sample.txt > /dev/null

In terminal #2:

KiB Mem:   7968336 total,  7011896 used,   956440 free,   267336 buffers

Notice the output of fatrace in terminal #3:

cat(25940): R /home/saml/tst/162600/sample.txt
cat(25940): R /home/saml/tst/162600/sample.txt
cat(25940): RC /home/saml/tst/162600/sample.txt

Now remove the file from RAM, in terminal #4:

$ sudo dd of=/home/saml/tst/162600/sample.txt \
    oflag=nocache conv=notrunc,fdatasync count=0

Note the output of fatrace in terminal #2:

dd(26229): O /home/saml/tst/162600/sample.txt
dd(26229): CW /home/saml/tst/162600/sample.txt

Note the RAM in terminal #3:

KiB Mem:   7968336 total,  6908364 used,  1059972 free,   267364 buffers

So it would seem that all of the that was consumed by the file in RAM is freed.

Potential Method #3 - python-fadvise

Thanks to a comment by @frostchutz, there's another tool, a Python script, named [pyadvise][4] which provides a much simpler interface than the above dd methods. This script makes use of the same posix_fadvise(2) interface.

Example

$ sudo pyadvise --help
Usage: 
    pyadvise [options] [FILE]..

Options:
  -h, --help        show this help message and exit
  -w, --willneed    The specified files will be accessed in the near future
  -s, --sequential  The application expects to access the specified files
                    sequentially (with lower offsets read before higher ones)
  -d, --dontneed    The specified files will not be accessed in the near
                    future
  -r, --random      The specified files will be accessed in random order
  -o, --noreuse     The specified files will be accessed only once. Under
                    Linux, this operation is a no-op; see contrib/copyfileobj-
                    fadvise.py in the python-fadvise source tree for an
                    example on how to achieve approximately the same effect
  -n, --normal      Indicates that the application has no advice to give about
                    its access pattern for the specified files. If no advice
                    is given for an open file, this is the default assumption
  -v, --verbose     Explain what is being done

And if we repeat the above test and use pyadvise in place of dd:

$ pyadvise -d /home/saml/tst/162600/sample.txt

I noticed an identical drop in the RAM being consumed as before when I used dd.

USB thumbdrive raw performance

For starters I'd take a look at the USB thumb drive just to make sure that it's operating at an acceptable level. You can use the following 2 commands to measure it's performance:

$ hdparm -t /dev/sdb

$ dd count=100 bs=1M if=/dev/zero of=/media/disk/test oflag=sync

Substitute in the appropriate HDD info for your given setup.

File system

Also take a look at this article on increasing performance of USB thumbdrives from the Ubuntu help site.

References

Limit the size of the linux file cache

Not really an answer, but too long for a comment.

Linux's memory management has been carefully tuned over its long lifetime by some very smart people and it normally does a pretty good job of making the right decision when choosing what to keep in memory and what to drop.

Unfortunately, it looks like your workload isn't very compatible with its decisions :-( Still, I am quite surprised about your report that it it is actually frequently preferring to force dirty memory out to swap rather than drop something from cache. A decision between dropping thing A from cache versus dropping thing B from cache might be a toss-up, but a decision between dropping thing C from cache versus writing out dirty memory D to swap should be heavily weighted toward dropping thing C because that's much less expensive!

There is a way to advise Linux that some piece of cached memory won't be needed in the future, and that is the madvise() system call with MADV_DONTNEED but I think it would be difficult for you to invoke that system call from a MATLAB script...

In any case, I don't think that reducing the size of the file cache is really what you want to do here. Remember that the executable files and libraries and run time files and plugins etc... for MATLAB itself, your script, your GUI environment, and other system software, also all live in the file cache, and you'd be forcing those things out of the file cache (in favour of either other contents of the file cache or in favour of heap memory and other non-file-backed mappings). That will cause your system to become slow just as surely as swapping will.