by Christos Tsotskas on 2017-02-13 13:41:09

---

Dear All,

I would like to run MATSim to simulate a large geographical area. However, for integration and scalabillity purposes, a single instance of MATSim is not enough (currently I am using a very high-power machine, with 30+ cores and 160+ GB of RAM). So, the natural thought would be to separate parts of the network, add extra code to handle "boundary conditions" between adjacent areas and implement a communication protocol that would control ownership. How could I do this by using MATSim? I have found a few scientific references by Pieter Fourie

and

http://www.futurecities.ethz.ch/assets/abm.workshop.html#/

but technical details are not present, so as to enable me to achieve the above.

Any ideas and thoughts are more than welcome and highly appreciated.

PS. to some point I would like to run all this in a cluster with many cores, for a truly big simulation.

Many Thanks in advance for your time and your effort,

XT

---

---

by Johan W. Joubert on 2017-02-13 13:59:24

Hi Christos, although it probably has not been officially published, I recall from a number of discussions at MATSim meetings that there is little to no benefit beyond something like 12 CPUs.

So if you want to get bang for your CPU buck I suggest you rather run multiple instances, each with a different random seed. That would be better from a simulation methodology point of view, I think.

But if you want to play with big toys, I suggest contact Pieter directly. 

---

by Christos Tsotskas on 2017-02-13 15:21:01

Hello Sir,

Thank you very much for your prompt reply. I was thinking something along the lines of the following split:

, 

where each section could be separately simulated by a different isntance of MATSim. Then, via any appropriate interface demand/supply by B could be transferred across to C, and vice-versa. Each letter, depending of their size would use different resources.

Is this feasible? If yes, could you please point out any appropriate directions?

Many Thanks,

XT