So, I was to the P2P conference and am now back since 3 days in France. It's time to do some sort of reporting on this.

The conference was incredibly useful. Not because of my presentation, actually. I only had 10 little minutes to present SimGrid, were I was used to be given 3 hours, as it was the case at the PDCAT conference in New-Zealand last december. This 10mn story realy sounds like a joke, even more because it was an invited talk. I got the feeling that they invited me but didn't really want to hear what I had to tell. Nevermind, the main interest came from the side discussions. I saw bright people, and I'm sure that several interesting collaboration will come up in the future.

First, I discussed quite a bit with Charles Kilian, from MACE project. The main strengh of SimGrid is its simulation kernel, both fast and accurate, and our main weakness (that I see) is our interfaces. It's a mess to explain why MSG and GRAS both exist and it's a mess to not being able to mix them together. More fundamentally, these are only interfaces to regular programming languages (mainly C but also Java for MSG). The main strengh of MACE is that they come up with a Domain Specific Language (DSL) for distributed applications. One hope would for example be to plug the SimGrid simulation kernel into the MACE framework. It would give the best of both worlds. Also, Mace guys are a bit more advanced with regard to model-checking of distributed applications, but there is no strict overlap of the efforts. Likewise, I guess that good things could come out of a collaboration on this topic.

Then, my invited talk was in a "evaluation tools" session, with people from the main P2P simulators and emulators invited. Several of them come from europe, so I played with the idea of federating our efforts to come up with the final simulator. We'd take SimGrid kernel, OverSim UI, ProtoPeer ability to live deploy java code (plus GRAS for C code), the statistics module of another project, and so on. But I'm really unsure that this will lead to something because almost every other tools presented were done by PhD students in groups studying P2P protocols. So, they did these tools because they were a need, but it's still to be proved that they are willing to go for a european project. Moreover, the timeframe then is something like 5 years (one year to setup the project one year to get it evaluated, three years of project). That's not very compatible with the time frame of a PhD about to write his dissertation. I'd love to see something spreading in this direction however. Will see.

Finally, I discussed with potential users, and I'm sure that their feedback would be incredibly usefull to us. Ensuring that SimGrid fit their needs would definitely improve the tool. The first group is the Skype company. I'm not sure of what I can tell about what we are thinking of because they are so secret-oriented. I prefer not to say anything, but the fact that they want to use SimGrid somehow. The second group is lead by Ken Birman at Cornell University. I'm completely excited about this too. SimGrid is not usable as is for what they want: they need multiple interfaces per machine. Like the fact that for my laptop, I can use the wireless or the wire connection. It's similar to a problem we already faced in HPC where clusters have 2 networks (a regular ethernet one, and one high-speed one, such as infiniband or myrinet). But not completely. We already started discussing how we could allow such things. I'd like to go fast on this, make the tool usable quickly to Cornell people. First because I have the feeling that them using it is exactly the kind of advertisement we need, but also because I'm sure that several of them can turn into power-users providing patches from time to time, or even to co-developers of the tool one day.