Re: btrfs write behavior on idle system

[Chris Mason <chris.mason@oracle.com>]

On Mon, Jan 18, 2010 at 08:36:37PM +0100, Carlos R. Mafra wrote:
> On Mo 18.Jan'10 at 12:14:43 -0600, Ravi Pinjala wrote:
> > On 01/18/10 11:17, Carlos R. Mafra wrote:
> > >Hi everyone,
> > >
> > >I am using btrfs for my /home partition since I upgraded my slow
> > >laptop hdd for an ssd 3 weeks ago. I am always in sync with Linus'
> > >tree of the day (plus a btrfs patch which is not in there yet) and
> > >so far I haven't lost any data, so all is good.
> > >
> > >I have a question about the write behavior of the various [btrfs- ]
> > >kernel threads, as I've been monitoring what is writing to the ssd
> > >just in case.
> > >
> > >So what I've been observing with 'iostat', 'iotop' and 'blktrace'
> > >is the following. If my laptop is almost absolutely idle (just
> > >a plain Window Maker and a few xterms and a couple dockapps open)
> > >there is nothing writing to the disk (which is OK).
> > >
> > >But as soon as I leave an open tab in chrome (or firefox) the various
> > >[btrfs- ] threads start writing in my /home, and I don't know what.
> > >For testing purposes, I mounted the config dir of chrome (~/.config/google-chrome)
> > >in my SD card (at /dev/mmcblk0p1) to exclude the possibility of maybe chrome
> > >trying to update its history or something, so that it does not write
> > >anything in my /home partition with btrfs.
> > >
> > >But I see this in the output of 'iotop' from a 60 sec interval, showing
> > >only the processes which wrote something:
> > >
> > >Total DISK READ: 0 B/s | Total DISK WRITE: 10.26 K/s
> > >   PID USER      DISK READ  DISK WRITE   SWAPIN    IO    COMMAND
> > >   485 root           0 B/s    5.19 K/s  0.00 %  0.02 % [btrfs-transacti]
> > >  3792 root           0 B/s       0 B/s  0.00 %  0.01 % [flush-btrfs-1]
> > >   476 root           0 B/s    0.13 K/s  0.00 %  0.00 % [btrfs-delalloc-]
> > >   481 root           0 B/s    4.93 K/s  0.00 %  0.00 % [btrfs-endio-wri]
> > >
> > >and there are more instances like this. Is there a way to avoid (or reduce)
> > >the writings of these threads?
> > >
> > >And when I start opening some pages in chrome and use it some more I
> > >get many many writes on my /home partition from these threads (and swapper,
> > >see below) even though I mounted the .config/google-chrome dir under
> > >/dev/mmcblk0p1 which uses ext4.
> > >
> > >>From another experiment where chrome was showing a blank tab a ~7 minutes
> > >run of 'blktrace -a write /dev/sda3' (sda3 is my /home) ends like this
> > >(from 'blkparse -s sda3.blktrace.0'):
> > >
> > - snip -
> > 
> > Don't forget cache - should be under ~/.cache/google-chrome. That
> > would probably explain the disk activity you're seeing.
> 
> Oh my... what a shame! That was it; I moved it to the SD card and
> now my ssd is not suffering anymore.

Web browsers can write a lot (really a lot) to disk.  There are caches
of the images, caches of the web pages, your history, book mark updates,
and then whatever tracking and bank account numbers google is holding on
to for you.

Does anyone know if chrome share's firefox's affinity to use fsync
heavily?  You could test with strace.

Either way, lots of writing when surfing the web isn't a big surprise.

-chris