views:

125

answers:

2

Just to clarify, this is not a homework problem :)

I wanted to find primes for a math application I am building & came across Sieve of Eratosthenes approach.

I have written an implementation of it in Python. But it's terribly slow. For say, if I want to find all primes less than 2 million. It takes > 20 mins. (I stopped it at this point). How can I speed this up?

def primes_sieve(limit):
    limitn = limit+1
    primes = range(2, limitn)

    for i in primes:
        factors = range(i, limitn, i)
        for f in factors[1:]:
            if f in primes:
                primes.remove(f)
    return primes

print primes_sieve(2000)

UPDATE: I ended up doing profiling on this code & found that quite a lot of time was spent on removing an element from the list. Quite understandable considering it has to traverse the entire list (worst-case) to find the element & then remove it and then readjust the list (maybe some copy goes on?). Anyway, I chucked out list for dictionary. My new implementation -

def primes_sieve1(limit):
    limitn = limit+1
    primes = dict()
    for i in range(2, limitn): primes[i] = True

    for i in primes:
        factors = range(i,limitn, i)
        for f in factors[1:]:
            primes[f] = False
    return [i for i in primes if primes[i]==True]

print primes_sieve1(2000000)
+2  A: 

Removing from the beginning of an array (list) requires moving all of the items after it down. That means that removing every element from a list in this way starting from the front is an O(n^2) operation.

You can do this much more efficiently with sets:

def primes_sieve(limit):
    limitn = limit+1
    not_prime = set()
    primes = []

    for i in range(2, limitn):
        if i in not_prime:
            continue

        for f in range(i*2, limitn, i):
            not_prime.add(f)

        primes.append(i)

    return primes

print primes_sieve(1000000)

... or alternatively, avoid having to rearrange the list:

def primes_sieve(limit):
    limitn = limit+1
    not_prime = [False] * limitn
    primes = []

    for i in range(2, limitn):
        if not_prime[i]:
            continue
        for f in xrange(i*2, limitn, i):
            not_prime[f] = True

        primes.append(i)

    return primes
Glenn Maynard
See @Piet Delport answer below for an optimization: replace `i*2` above with `i*i`.
GregS
+1  A: 

You're not quite implementing the correct algorithm:

In your first example, primes_sieve doesn't maintain a list of primality flags to strike/unset (as in the algorithm), but instead resizes a list of integers continuously, which is very expensive: removing an item from a list requires shifting all subsequent items down by one.

In the second example, primes_sieve1 maintains a dictionary of primality flags, which is a step in the right direction, but it iterates over the dictionary in undefined order, and redundantly strikes out factors of factors (instead of only factors of primes, as in the algorithm). You could fix this by sorting the keys, and skipping non-primes (which already makes it an order of magnitude faster), but it's still much more efficient to just use a list directly.

The correct algorithm (with a list instead of a dictionary) looks something like:

def primes_sieve2(limit):
    a = [True] * limit                          # Initialize the primality list
    a[0] = a[1] = False

    for (i, isprime) in enumerate(a):
        if isprime:
            yield i
            for n in xrange(i*i, limit, i):     # Mark factors non-prime
                a[n] = False

(Note that this also includes the algorithmic optimization of starting the non-prime marking at the prime's square (i*i) instead of its double.)

Piet Delport
another optimization, the step size of your `xrange(i*i,limit,i)` can be made `2*i`
st0le
Well spotted! If i'm not mistaken, that is a step in the direction of Euler's version of the sieve, right?
Piet Delport