Agfdhyk

I just started to use numba to improve performance of my programs. I have reduce the case that I will present

import numba as nb

import numpy as np

from time import time





def dt_max(U,f, eps=1e-5):

    return np.min( np.abs( U ) / ( np.abs( f ) + eps ) )



@nb.njit(cache=True)

def fast_dt_max(U,f, eps=1e-5):      

    m=U[0]

    if m<0 : m=-U[0]

    for i in range(len(U)) :

        v = abs(U[i]) / ( abs(f[i]) + eps )

        if v < m : m = v

    return m





N=100

Niter = int(1e5)



x=np.linspace(-50,50,N)



t0 = time()

for i in range(Niter):

    dt_max(x,x)

print('numpy',time()-t0)



t0 = time()

for i in range(Niter):

    fast_dt_max(x,x)

print('numba' ,time()-t0)

I execute this entire file in spyder under the following distribution :

Python 3.5.5 |Anaconda custom (64-bit)| (default, Apr  7 2018, 04:52:34) [MSC v.1900 64 bit (AMD64)]

Type "copyright", "credits" or "license" for more information.

The problem is the following. When I execute for the first time this program, it looks work good :

runfile('E:/02-Codes/TestCode/Python_numba/bug_second_execution.py', wdir='E:/02-Codes/TestCode/Python_numba')

numpy 0.45239996910095215

numba 0.2964000701904297

But when I repeat the execution of all file :

runfile('E:/02-Codes/TestCode/Python_numba/bug_second_execution.py', wdir='E:/02-Codes/TestCode/Python_numba')

numpy 0.45239996910095215

numba 3.5879998207092285



runfile('E:/02-Codes/TestCode/Python_numba/bug_second_execution.py', wdir='E:/02-Codes/TestCode/Python_numba')

numpy 0.4679999351501465

numba 3.5734000205993652

The numba's performance are not the same at all. If I restart Python kernel in my spyder environment, the problemn happens again : first execution is good, all the followings not.

So the first question is : why ? and the second is : how to avoid that ?

Thanks for your help

Thanks a lot Max for the answer.

Indeed, the default value for eps is the reason of the problem. If I switch in debug mode (thanks Max for this tips!) :

import os

os.environ['NUMBA_DEBUG_CACHE'] = '1'

import numba as nb

import numpy as np

from time import time



@nb.njit(cache=True)

def fast_dt_max(U,f, eps=1e-5):

    m=U[0]

    if m<0 : m=-U[0]

    for i in range(len(U)) :

        v = abs(U[i]) / ( abs(f[i]) + eps )

        if v < m : m = v

    return m

This code shows that at each execution, 4 operations is done (2 loads, and 2 saving). Now if I remove the default value for eps, by keeping debug mode :

@nb.njit(cache=True)

def fast_dt_max(U,f, eps):

    m=U[0]

    if m<0 : m=-U[0]

    for i in range(len(U)) :

        v = abs(U[i]) / ( abs(f[i]) + eps )

        if v < m : m = v

    return m

the successive call to the function, let appear only 2 operations per execution (1 load + 1 saving)

CONCLUSION : Not use default parameter values in numba function.

Thanks !