Why generating four Poisson numbers is so slow

Roofit and RooStats
1

I have a very simple workspace, made by the join of four poisson distribution

w = ROOT.RooWorkspace()
w.factory("RooPoisson:p1(obs1[0, 1000], 10)")
w.factory("RooPoisson:p2(obs2[0, 1000], 20)")
w.factory("RooPoisson:p3(obs3[0, 1000], 30)")
w.factory("RooPoisson:p4(obs4[0, 1000], 40)")
model = w.factory("PROD:model(p1, p2, p3, p4)")
all_obs = w.allVars().selectByName('obs*')

NTOYS = 1000
for itoy in range(NTOYS):
    data = model.generate(all_obs, 1)

this took 24 seconds on my laptop, while

%%timeit
model.generate(all_obs, 1000)

10 loops, best of 3: 27.4 ms per loop

from scipy import stats
%%timeit
stats.poisson([10, 20, 30, 40]).rvs((1000, 4))

1000 loops, best of 3: 1.4 ms per loop

So: first version = 1x, second version = 870x, scipy version = 17000x

Am I doing something wrong? I would prefer the version model.generate(all_obs, 1) since I want to model a counting experiment. By the way there is a 20x between the vectorized version of RooFit and scipy.

2

Hi,
I think it is slow because you are using C style loop,
try to have something like

data = [model.generate(all_obs, 1) for itoy in xrange(NTOYS) ]

3

I get exactly the same timing. In fact list comprehension should not be very different for explicit for-loops. I guess there is an overhead for some subtask done by generate which is repeated all the times.

I am wondering if it is possible to split the generate method in the part that can be done just one time and the real random generation.