[atlas] Feature-suggestion: "round-robin" probes

On Wed, 09 May 2012 11:20:38 +0200
Philip Homburg <philip.homburg at ripe.net> wrote:

> >>    * TTM shutdown, Atlas is expected to provide functionality
> >> similar (but not identical) to what TTM provides
> > You plan to change the measurement control protocol used in TTM ?
> > I do not know which protocol TTM is using but if it's OWAMP ou
> > TWAMP, that will not fit the needs for large scale measurement
> > campaign runt from network edge.
> TTM boxes have GPS devices for time synchronization. That allows them
> to perform accurate one-way measurements. This capability will be
> lost.

Maybe not. There's two way to cope with the lost of GPS device :

* The smart one : 

If you use a Linux or FreeBSD kernel, you can use : 

http://www.cubinlab.ee.unimelb.edu.au/radclock/ which is very
accurate (less than a micro-second) and cheap. 

* The dumb one : 

The time keeping feature is done by the control-configuration server,
which mean that no time-stamp will be sent to "probe", only the the
time difference between the current time and the measurement
time-stamp calculated on the server. It's not very accurate but it
permit to do measurement without the need to have the "probe"
synchronized. The main problem here is that the protocol is so dumb
that the time elapsed by packet travel is never counted anywhere.
If seconds accuracy  is enough, that the simplest solution. 

> 
> The TTM network is relatively small and static. Atlas can easily
> handle that, except you will be limited to two-way measurements.
> > In what Atlas call a "probe" (and what I call the measurement
> > agent) ?
> >
> Yes. Except that an Atlas probe tends to be a physical device as well.

I hope the host functionalities and the measurement functionalities
have been thought with some kind of separation :p 

> 
> >>    * Roll out of Atlas Anchor boxes (regular PCs at well connected
> >>      locations that can serve as the target of measurements and as
> >> a more powerful Atlas probe)
> > Sound like a good idea :)
> >
> > You then should add a tag to the measurement result that will permit
> > to distinguish the type of box running the measurement agent, like
> >
> > "generated": atlas-probe
> > "generated": atlas-box
> >
> > for example.
> We still have to figure out where we want to document meta-data. It 
> doesn't make much sense to put all data about a probe in each and
> every measurement result.

In the some kind of "hello" command that describe the probe by itself
only.  

In your system, the measurement agent will probably expose two
orthogonal API : 

* One to write a measurement; 
* One to gather information on the measurement host (OS type, packets
  in flight, and so on). 

I say probably because we also have discussed a pipelined architecture
of measurements that will not make this kind of difference : 

measurement 1 -OK-> measurement 2 -OK-> measurement 3
               |
	       KO -> stop measurement  

which permit to mimic a decision tree with not must complexity
added : measurement 2 wait for a result that might be conditional to
be runt. 

The latter is probably the better.  


> >>    * Better UDM interface
> >>    * UDM for all RIPE members (instead of just probe hosts and
> >> sponsors)
> > Eye candies.
> 
> No, it is not eye candy. UDM allows users of the Atlas system to
> measure their own targets using remote probes.

I see. Interface is then the protocol used by users to define their own
measurement. The work flow should define is some way a "central
authority" that will permit the measurement to be runt. As far as I
see it, UDM is a moderated measurement campaign generator that can
be hosted anywhere as long as the configuration approved go to the
configuration server and are distributed to "probe". 

> 
> We have this as internal documentation, but it should be published
> some time.
> 

Let me known when the dust has settled and RIPE publish them. 

> 
> >
> >
> > We do not have securities problem because of OCaml choice
> > on the implementation side. Securities mechanism will probably be
> > the same as of the one you can find on most "web services" (shared
> > secret salted and hashed) to ensure that a REST transaction is
> > legit. I have to think about it some more ...
> Our security policy goes further than just protecting the probe. We
> also try to avoid getting the probe hosts in trouble. For example,
> having a probe visit certain web sites may be a bad idea.

The policy is enforced on the configuration server ? (that what we
intend to implement via some kind of automation when possible)


I should write down the big picture after some talks by the end of
may.  

> >
> >
> > For API and JSON syntax standardisation, the first step is to write
> > the specifications we(grenouille.com) plan to use and Atlas use and
> > plan to use, then discuss and factor out the best of each. We have
> > some writings but most of them are in French :)
> Yes.

Great. We're going to translate some already written and document the
whole architecture in details. 

Cheers. 

-- 
Jérôme Benoit aka fraggle
La Météo du Net - http://grenouille.com
OpenPGP Key ID : 9FE9161D
Key fingerprint : 9CA4 0249 AF57 A35B 34B3 AC15 FAA0 CB50 9FE9 161D
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: </ripe/mail/archives/ripe-atlas/attachments/20120509/a7404783/attachment.sig>