[pve-devel] Default cache mode for VM hard drives

[Stanislav German-Evtushenko]

On Tue, Nov 18, 2014 at 9:28 PM, Dietmar Maurer <dietmar at proxmox.com> wrote:

>
> > Sorry, but I cannot find an explaination there?
> >
> > In short:
> > 1) DATA from write buffer commits locally
> > 2) DATA in write buffer changes (because write cache is enabled)
> > 3) DATA from write buffer sends to another DRBD node
> > And DRBD can't catch this until integrity-alg not enabled. When enabled
> - we get
> > split brain for each occurrence for master-master or just secondary
> > reconnection for master-slave mode.
>
> So this is a DRBD bug?
>

Lars doesn't see this as a bug. See also says that software raid behaves in
the same way. I've asked him twice if it is possible to copy buffer before
committing but haven't got any answer.

> If DRBD requires O_DSYNC, it is incredible slow?

No difference to cache=none when using enterprise level hardware raid with
write cache enabled (BBU must be in place) because hardware raid ignores
O_DSYNC and reports data being committed as it reached raid cache. There
could be huge difference for low end raid controllers and other cases when
O_DSYNC is not ignored.

Stanislav
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.proxmox.com/pipermail/pve-devel/attachments/20141118/5898e65c/attachment.htm>