[CRIU] [PATCH 0/15] Buffered images IO

Cyrill Gorcunov gorcunov at gmail.com
Tue Sep 30 08:18:36 PDT 2014


On Mon, Sep 29, 2014 at 12:46:28PM +0400, Pavel Emelyanov wrote:
> Hi
> 
> Recent experiments show, that we spend too much time in writing
> and reading images. E.g. for dump strace shows that we spend most
> of the time logging (with -v4) and writing images. On restore we
> wait for other tasks to do it (only spoils the parallel restore)
> and, yet again, reading from images.
> 
> This is because we use write() and read() for every single object.
> For on-disk FS this it too time consuming even when using the
> std Linux page cache -- small IO-s take little time to get
> memcpy-d and a lot of time for FS preparations (especially writes).
> 
> So this is a set that uses the recently added bfd engine for
> images IO. On dump this removes ~14% of time for p.haul test and
> for restore it removes ~3K read() calls our of ~4.5K (hard to
> estimate the time percentage since we're in multi-process mode 
> there).

Lets merge it and look how it goes fixing nits on top.
Acked-by: Cyrill Gorcunov <gorcunov at openvz.org>


More information about the CRIU mailing list