On 2017-09-18 12:02, Stefan Hajnoczi wrote:
> On Sat, Sep 16, 2017 at 04:02:45PM +0200, Max Reitz wrote:
>> On 2017-09-14 17:42, Stefan Hajnoczi wrote:
>>> On Wed, Sep 13, 2017 at 08:18:52PM +0200, Max Reitz wrote:
>>>> There may be a couple of things to do on top of this series:
>>>> - Allow switching between active and passive mode at runtime: This
>>>>   should not be too difficult to implement, the main question is how to
>>>>   expose it to the user.
>>>>   (I seem to recall we wanted some form of block-job-set-option
>>>>   command...?)
>>>>
>>>> - Implement an asynchronous active mode: May be detrimental when it
>>>>   comes to convergence, but it might be nice to have anyway.  May or may
>>>>   not be complicated to implement.
>>>
>>> Ideally the user doesn't have to know about async vs sync.  It's an
>>> implementation detail.
>>>
>>> Async makes sense during the bulk copy phase (e.g. sync=full) because
>>> guest read/write latencies are mostly unaffected.  Once the entire
>>> device has been copied there are probably still dirty blocks left
>>> because the guest touched them while the mirror job was running.  At
>>> that point it definitely makes sense to switch to synchronous mirroring
>>> in order to converge.
>>
>> Makes sense, but I'm not sure whether it really is just an
>> implementation detail.  If you're in the bulk copy phase in active/async
>> mode and you have enough write requests with the target being slow
>> enough, I suspect you might still not get convergence then (because the
>> writes to the target yield for a long time while ever more write
>> requests pile up) -- so then you'd just shift the dirty tracking from
>> the bitmap to a list of requests in progress.
>>
>> And I think we do want the bulk copy phase to guarantee convergence,
>> too, usually (when active/foreground/synchronous mode is selected).  If
>> we don't, then that's a policy decision and would be up to libvirt, as I
>> see it.
> 
> This is a good point.  Bulk copy should converge too.
> 
> Can we measure the target write rate and guest write rate?  A heuristic
> can choose between async vs sync based on the write rates.
> 
> For example, if the guest write rate has been larger than the target
> write rate for the past 10 seconds during the bulk phase, switch to
> synchronous mirroring.

I guess we can just count how many unfinished target write requests are
piling up.

...or libvirt can simply see that the block job is not progressing and
switch the mode. :-)

Max