Re: [Qemu-devel] [RFC 0/3] qemu-ga: support for sending events

[Tomáš Golembiovský]

Hi,

On Sun, 25 Jun 2017 23:25:17 +0200
Tomáš Golembiovský <address@hidden> wrote:

> Hi,
> 
> On Fri, 23 Jun 2017 13:25:34 +0000
> Marc-André Lureau <address@hidden> wrote:
> 
> > Hi
> > 
> > On Fri, Jun 23, 2017 at 3:03 PM Tomáš Golembiovský <address@hidden>
> > wrote:
> >   
> > > This is just a draft, or a request for comments if you will.
> > >
> > > This patch sets drafts the support of sending events by QEMU Guest Agent.
> > > Events can plan important role in monitoring of the guest OS behaviour. 
> > > The
> > > range of use cases ranges from events important for scheduling, e.g.
> > > memory and
> > > CPU usage statistics, to things like changes to IP addresses on network
> > > interfaces to for example changes in the list of active users.
> > >
> > > For now the patch set adds single periodic callback function to the GA 
> > > main
> > > loop that can perform checks and trigger events that have occured since
> > > previous run of the callback.
> > >
> > > We can of course take it one step further and add a general framwork for
> > > periodically running any of the already implemented commands. Add a
> > > function
> > > that would maintain a list of registered checks. Client would use some
> > > command
> > > (register-monitor-command) passing it a command name and timeout in
> > > seconds and
> > > the monitoring handler would then run the specified command and report the
> > > result... or report only if the return value changed since previous
> > > invocation.
> > > This feature would remove part of the communication overhead between
> > > client and
> > > GA.
> > >
> > > So before I invest any more time in either of these approaches, tell me.
> > > Would
> > > somethign like this be wanted or is that too controversial? Any other
> > > thoughts
> > > and ideas?
> > >
> > >    
> > It doesn't feel wrong, but Is there really too much overhead and/or latency
> > if a request is periodic from the client? ie did you do some measurements
> > before coming to this proposal?  
> 
> No, I didn't do any measurements. And it may be even true that in the
> grand scheme of things the overhead/latency may be insignificant, if we
> imagine a client that repeatedly calls about 5 to 10 commands every 5 or
> 10 seconds. Still, it just feels like a more correct approach to me. But
> that may be just my feeling, that's why I brought this to the list to
> get the opinion of others.
> 

This is not really a wild discussion I have anticipated. Or does the
silence mean I should drop the idea?

If some measurements are necessary can you suggest how to construct the
benchmark? What numbers would be convincing to support the idea?

Still, let me restate that I see it more as an architectural decision
rather than something where latency or overhead would be the main
factor.


    Tomas

> 
> > 
> > Tomáš Golembiovský (3):  
> > >   qemu-ga: add support for events
> > >   qemu-ga: add simple event reporting memory statistics
> > >   qemu-ga: add support for periodic command runner
> > >
> > >  Makefile               |  7 +++-
> > >  qga/Makefile.objs      |  2 +-
> > >  qga/channel-posix.c    |  8 +++++
> > >  qga/channel-win32.c    |  6 ++++
> > >  qga/channel.h          |  1 +
> > >  qga/guest-agent-core.h |  1 +
> > >  qga/main.c             | 98
> > > ++++++++++++++++++++++++++++++++++++++++++++++++++
> > >  qga/qapi-event.json    | 35 ++++++++++++++++++
> > >  qga/qapi-schema.json   |  2 ++
> > >  9 files changed, 158 insertions(+), 2 deletions(-)
> > >  create mode 100644 qga/qapi-event.json
> > >
> > > --
> > > 2.13.1
> > >
> > >
> > > --    
> > Marc-André Lureau  
> 
> 
> -- 
> Tomáš Golembiovský <address@hidden>


-- 
Tomáš Golembiovský <address@hidden>