search

rakhimov / scram

Probabilistic Risk Analysis Tool (fault tree analysis, event tree analysis, etc.)
https://scram-pra.org
GNU General Public License v3.0
131 stars 54 forks source link

Parallelize Monte Carlo Simulations #140

Open rakhimov opened 8 years ago

rakhimov commented 7 years ago

Boost accumulators don't support parallelism via map-reduce or combine.

jto888 commented 7 years ago

Oh, I hate when that happens!

On 1/18/2017 11:57 PM, Olzhas Rakhimov wrote:

Boost accumulators don't support parallelism via map-reduce or combine.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rakhimov/scram/issues/140#issuecomment-273682269, or mute the thread https://github.com/notifications/unsubscribe-auth/AIc65MZeP3bmmFJgcDewcD80zCuM5dOeks5rTu0cgaJpZM4I4Jd1.

rakhimov commented 7 years ago

@jto888 I am not sure how bad the naive implementation would be: with one vector per thread, and push all the data in the vectors into one accumulator (the reduce step).

It would be inefficient, but sampling and calculating probabilities/values would be much more expensive in my estimates (like 200x with many cache misses) at least initially.

rakhimov commented 7 years ago

An alternative approach is to implement custom combine<tag::stat_measure>(accumulators...) per statistical measure. This is simple for mean and stddev but doesn't sound fun when statistical measures get complex.