End Point's Liquid Galaxy platform is a cluster of computers distributing the rendering load of Google Earth and other visualization applications. Each of these "display nodes" boots the same disk image distributed over the local network. Our disk image is based on the Ubuntu 12.04 LiveCD, and uses the same overlayfs to combine a read-only ISO with a writeable ramdisk. This Union mount uses Copy-on-write to copy files from the read-only "lower" filesystem to the virtual "upper" filesystem whenever those files or directories are opened for writing.
We often allocate 4GB of system memory to the ramdisk containing the "upper" filesystem. This allows 4GB of changed files on the / filesystem, most of which are often Google Earth cache files. But sometimes the ramdisk fills up with other changes and it's difficult to track down which files have changed unexpectedly.
The df command properly displays total usage for the overlayfs "upper" filesystem mounted at /.
$ df -h Filesystem Size Used Avail Use% Mounted on /cow 3.9G 2.2G 1.8G 55% / /dev/loop0 833M 833M 0 100% /cdrom /dev/loop1 808M 808M 0 100% /rofs
But how can we identify which files are consuming that space? Because the root device has been pivoted by the casper scripts at boot, the /cow device is not readily available. We often use the du tool to estimate disk usage, but in this case it cannot tell the difference between files in the "upper" ramdisk and the "lower" read-only filesystem. To find the files filling our /cow device, we need a way to enumerate only the files in the "upper" filesystem, and then estimate only their disk usage.
The `mount` command shows the / filesystem is type "overlayfs".
$ mount /cow on / type overlayfs (rw) /dev/loop0 on /cdrom type iso9660 (ro,noatime) /dev/loop1 on /rofs type squashfs (ro,noatime)
The find command does indicate that most directories exist in the filesystem of type "overlayfs", and most unmodified files are on the "lower" filesystem, in this case "squashfs".
$ sudo find / -printf '%F\t%D\t%p\n' | head -n 7 ### fstype, st_dev, filename overlayfs 17 / overlayfs 17 /bin squashfs 1793 /bin/bash squashfs 1793 /bin/bunzip2 squashfs 1793 /bin/bzcat squashfs 1793 /bin/bzcmp squashfs 1793 /bin/bzdiff
However, the modified files are reported to be on an "unknown" filesystem on a device 16. These are the files that have been copied to the "upper" filesystem upon writing.
$ find /home/lg/ -printf '%F\t%D\t%p\n' | head -n 33 overlayfs 17 /home/lg/ squashfs 1793 /home/lg/.Xresources unknown 16 /home/lg/.bash_history squashfs 1793 /home/lg/.bash_logout squashfs 1793 /home/lg/.bashrc overlayfs 17 /home/lg/.config overlayfs 17 /home/lg/.config/Google unknown 16 /home/lg/.config/Google/GoogleEarthPlus.conf unknown 16 /home/lg/.config/Google/GECommonSettings.conf
I couldn't quickly discern how find is identifying the filesystem type, but it can use the -fstype test to reliably identify files that have been modified and copied. (Unfortunately find does not have a test for device number so if you have more than one overlayfs filesystem this solution may not work for you.)
Now that we have a reliable list of which files have been written to and copied, we can see which are consuming the most disk space by piping that list to du. We'll pass it a null-terminated list of files to accommodate any special characters, then we'll sort the output to identify the largest disk space hogs.
$ sudo find / -fstype unknown -print0 | du --files0-from=- --total | sort -nr | head -n 4 2214228 total 38600 /var/cache/apt/pkgcache.bin 38576 /var/cache/apt/srcpkgcache.bin 17060 /var/log/daemon.log
We included a total in this output, and notice that total is exactly the 2.2GB indicated by the df output above, so I believe this is measuring what we intend.
Of course five hundred 2MB Google Earth cache files consume more space than a single 38MB apt cache file, so we'd like to list the directories whose files are consuming the most ramdisk space. Unfortunately giving find and du depth arguments won't work: the "unknown" filesystem doesn't have any directories that we can see. We'll have to parse the output, and that's left as an exercise for the reader for now.
I just realized I could've simply looked for files modified after boot-time and gotten very similar results, but that's not nearly as fun. There may be a way to mount only the "upper" filesystem, but I was disappointed by the lack of documentation around overlayfs, which will likely be included in the mainline 3.10 Linux kernel.