Feature request: multi core support for benchmarking

[ompl]

Feature request: multi core support for benchmarking

[mamoll]

Original comment by Mark Moll (Bitbucket: mamoll, GitHub: mamoll).

---

This is useful to include in MachineSpecs.h to get the number of cores:
http://stackoverflow.com/questions/150355/programmatically-find-the-number-of-cores-on-a-machine

[mamoll]

Original comment by Ioan Sucan (Bitbucket: isucan, GitHub: isucan).

---

How about just using boost::thread::hardware_concurrency()?

[mamoll]

Original comment by Mark Moll (Bitbucket: mamoll, GitHub: mamoll).

---

Yes, that is the right thing to do. It's one of the suggestions in the stackoverflow article.

Is it possible to monitor memory consumption per thread? The config file allows you to specify a memory limit per run, but this seems hard to check per thread. We should also check that the number of threads <= free memory / memory limit per run.

[mamoll]

Original comment by Ioan Sucan (Bitbucket: isucan, GitHub: isucan).

---

Right now we are measuring memory per process. I am not sure how to measure memory per thread. In general, I think it is easier to reason about total memory consumption than memory per thread because there are often threads that don't do much, just wait around for an event (e.g., sleep or condition wait). Do you have a specific use case in mind?

[mamoll]

Original comment by Ioan Sucan (Bitbucket: isucan, GitHub: isucan).

---

Seems this issue has been around for a while. I am thinking that since it is pretty easy to parallelize outside OMPL, i.e., run the benchmarking process with different inputs in parallel, this feature is not something that would add much utility, but would add complexity and perhaps inaccuracies in the produced results. I propose we close this as wontfix.

[mamoll]

Original comment by Mark Moll (Bitbucket: mamoll, GitHub: mamoll).

---

Agreed. For example, with the benchmark program in omplapp it's easy enough to launch separate processes in parallel with different config files (one per planner, say).