parallel simulation for radio medium

[ypan...@icloud.com]

Hi I have question regarding parallel simulation.

For those objects like radio medium, where there is only one instance in the simulation scenario, how should I partition the module so that I can take the benefit of parallel simulation?

Thanks!

[Rudolf Hornig]

On Friday, 19 May 2017 00:50:43 UTC+2, ypan...@icloud.com wrote:

Hi I have question regarding parallel simulation.

For those objects like radio medium, where there is only one instance in the simulation scenario, how should I partition the module so that I can take the benefit of parallel simulation?

Well, to be honest, you can't. Wireless simulation is the prime example, where paralelizing a simulation using a (classical) computer is not feasible. The reason is that the delay between the information passing between the nodes is absolutely the minimum (being the speed of light) so the different partitions in the simulation cannot proceed independently (at least long enough to be feasible). 

Mother Nature solves this problem by limiting the propagation speed of the information to a constant value (speed of light) and at the same time employing a massively parallel, extremely fast quantum computer... Sadly we are just not there yet... :( It's fun to think, but even the universe would need infinite computing power to compute herself if the speed of light would not be limited, so that was a reasonable engineering limit during the creation of the universe ;)

And one additional tip, just for the sake of record so someone else will find it later via search:

Parallel simulation is always a bit over-hyped to speed up the simulation. In some corner cases it can and it is very effective. In certain situations it can give super-linear speedup (i.e. you can achieve 10x times speedup with using 8 CPUs), but it is rare.

On the other hand, while you are developing the simulation model you need to debug/track/inspect it and mostly you are running debug (un-optimized) code anyway in a GUI, so parallel execution is in the way. Once you run your final simulation (you should compile it in release mode!), probably exploring parameter spaces etc., you NEVER execute a single run. Even if you are not exploring a parameter space, you MUST run the same simulation several times with different RNG seeds to gain statistical confidence in the final numbers. A single run will NEVER provide your solution.

Now, at this point you can safely use all of your CPUs in your machine to execute several *independent* runs parallel (see the opp_run_all command and the batch running modes in the IDE). This way you can utilize your CPU resources at near 100% without any tricks inside your code. And even better, you can deploy it on a cloud service if you are in a hurry. Only money is the limiting factor :)

 

[ypan...@icloud.com]

Thanks for the explanation, Rudolf.

On Monday, May 22, 2017 at 3:21:38 AM UTC-6, Rudolf Hornig wrote:

On Friday, 19 May 2017 00:50:43 UTC+2, ypan...@icloud.com wrote:

Hi I have question regarding parallel simulation.

For those objects like radio medium, where there is only one instance in the simulation scenario, how should I partition the module so that I can take the benefit of parallel simulation?

Well, to be honest, you can't. Wireless simulation is the prime example, where paralelizing a simulation using a (classical) computer is not feasible. The reason is that the delay between the information passing between the nodes is absolutely the minimum (being the speed of light) so the different partitions in the simulation cannot proceed independently (at least long enough to be feasible). 

Mother Nature solves this problem by limiting the propagation speed of the information to a constant value (speed of light) and at the same time employing a massively parallel, extremely fast quantum computer... Sadly we are just not there yet... :( It's fun to think, but even the universe would need infinite computing power to compute herself if the speed of light would not be limited, so that was a reasonable engineering limit during the creation of the universe ;)

 

Thanks!

[Rudolf Hornig]

Oh and one additional tip, just for the sake of record so someone else will find it later via search:

Parallel simulation is always a bit over-hyped to speed up the simulation. In some corner cases it can and it is very effective. In certain situations it can give super-linear speedup (i.e. you can achieve 10x times speedup with using 8 CPUs), but it is rare.

On the other hand, while you are developing the simulation model you need to debug/track/inspect it and mostly you are running debug (un-optimized) code anyway in a GUI, so parallel execution is in the way. Once you run your final simulation (you should compile it in release mode!), probably exploring parameter spaces etc., you NEVER execute a single run. Even if you are not exploring a parameter space, you MUST run the same simulation several times with different RND seeds to gain statistical confidence in the final numbers. A single run will NEVER provide your solution.

Now, at this point you can safely use all of your CPUs in your machine to execute several *independent* runs parallel. This way you can utilize your CPU resources at near 100% without any tricks inside your code. And even better, you can deploy it on a cloud service if you in a hurry. Only money is the limiting factor :)