# Sieve of Sundaram for Project Euler 7: Python implementation slower than C++ and R

A friend of mine recently started learning R, and she complained to me that R is very slow. She was working on Project Euler Problem 7, which asks for the value of the 10001st prime number. For comparison, I decided to write the exact same program in C++ (which I use most frequently), R (which I do not use regularly, but used for comparison), and Python (because I wanted to advise her to switch her learning from R to Python).

The program follows these steps:

1. Find the upper bound on the Nth prime.
2. Fill a prime number sieve (Sieve of Sundaram) large enough to capture the upper bound.
3. Count the primes in the prime number sieve until the Nth prime is found.

To my surprise, the Python solution was significantly slower than the R version. Here are the results from time for the three implementations:

• C++: real 0.002s, user 0.001s, sys 0.001s
• R: real 2.239s, user 2.065s, sys 0.033s
• Python: real 13.588s, user 13.566s, sys 0.025s

I tried using numpy arrays instead of lists, but to little effect.

I would appreciate an explanation of why the Python implementation is so slow relative to the other two. I would also appreciate tips on how to use lists in R and Python more efficiently. Is there a better way to do this?

C++:

#include <cmath>
#include <iostream>
#include <vector>
using namespace std;

const int N = 10001;

int main()
{
int max = floor(N*(log(N) + log(log(N))));
vector<bool> marked(max/2, false);
for (size_t i = 1; i < marked.size(); i++) {
for (size_t j = 1; j <= i; j++) {
size_t m = i + j + 2*i*j;
if (m < marked.size()) {
marked[m] = true;
} else {
break;
}
}
}
int count = 1;
for (size_t m = 1; m < marked.size(); m++) {
if (not marked[m]) count++;
if (count == N) {
cout << 2*m + 1 << endl;
break;
}
}
return 0;
}


R:

n <- 10001
max <- n*(log(n) + log(log(n)))
marked <- vector(mode="logical", length=max/2)
for (i in 1:length(marked)) {
for (j in 1:i) {
m <- i + j + 2*i*j
if (m > length(marked)) break
marked[m] <- T
}
}
count <- 1
for (i in 1:length(marked)) {
if (!marked[i]) count <- count + 1
if (count == n) {
print(2*i + 1)
break
}
}


Python:

from math import log

n = 10001
max = n*(log(n) + log(log(n)))
marked = [False] * int(max/2)
for i in range(1, len(marked)):
for j in range(1, i + 1):
m = i + j + 2*i*j
if m > len(marked):
break
marked[m] = True
count = 0
for i in range(0, len(marked)):
if not marked[i]:
count += 1
if count == n:
print 2*i + 1
break

• Rather than trying to do three code reviews in on question, I have edited the title and tags to focus on the Python performance question. You may wish to ask separate questions about the C++ and R implementations. – 200_success Nov 29 '14 at 3:15
• In the C++ version, j goes from 1 to i, while in the R version j goes from 1 to length (marked). – mjolka Nov 29 '14 at 3:54
• @200_success, Thanks, that is a much better title. – castle-bravo Nov 29 '14 at 4:21
• @mjolka, I fixed the sloppy mistake in the R version. The code now runs in about half the time. I'm still curious as to why R beats Python here. – castle-bravo Nov 29 '14 at 4:24
• Was it entirely unexpected that python is slower than C++? – dtech Nov 29 '14 at 19:14

TL;DR: Just use PyPy; it gets you to about 10x the time of C++. If you really want to use CPython, a lot of clever optimizations (not algorithm changes) gets you as fast as PyPy and then using Numpy gets you close to C++ (2x the time).

The first thing of note is that your Python code is broken:

    if m > len(marked):
break


Remember that Python is 0-indexed. What about when m == len(marked)?

So that's the first thing to fix. Going from the top, I'd do the translation so:

from math import log

n = 10001                              n <- 10001
maximum = n * (log(n) + log(log(n)))   max <- n*(log(n) + log(log(n)))
marked = [False] * int(maximum // 2)   marked <- vector(mode="logical", length=max/2)
for i in range(1, len(marked)+1):      for (i in 1:length(marked)) {
for j in range(1, i+1):              for (j in 1:i) {
m = i + j + 2*i*j                  m <- i + j + 2*i*j
if m > len(marked): break          if (m > length(marked)) break
marked[m-1] = True                 marked[m] <- T
}
}
count = 1                              count <- 1
for i in range(1, len(marked)+1):      for (i in 1:length(marked)) {
if not marked[i-1]: count += 1       if (!marked[i]) count <- count + 1
if count == n:                       if (count == n) {
print(2*i + 1)                     print(2*i + 1)
break                              break
}
}


This is as direct a mapping as possible; instead of changing the comparison I just shifted the index when indexing. This isn't idiomatic, but it's direct. It's largely the same as your code, but it's correct. This matters when we go to larger N, where your code fails. It's also Python 3 compatible simply by using print with brackets.

Let's time them:

$time Rscript r.r [1] 104743 Rscript r.r 1.59s user 0.78s system 99% cpu 2.375 total$ time python2 p.py
104743
python2 p.py  12.88s user 0.00s system 100% cpu 12.873 total

$time python3 p.py 104743 python3 p.py 0.16s user 0.00s system 98% cpu 0.163 total$ # A faster, very compatible Python interpreter
$time pypy p.py 104743 pypy p.py 0.04s user 0.01s system 98% cpu 0.051 total$ time pypy3 p.py
104743
pypy3 p.py  0.05s user 0.01s system 99% cpu 0.054 total


So there we have it. Python is over an order of magnitude faster on 75% of interpreters, and under an order of magnitude slower in the worst case...

But why is it so slow with python2? line_profiler is a good utility:

import line_profiler
profiler = line_profiler.LineProfiler()

def main():
# ... the code ...

profiler.enable()
main()
profiler.print_stats()


Giving:

Line #      Hits         Time  Per Hit   % Time  Line Contents
==============================================================
4                                           def main():
5         1          271    271.0      0.0      from math import log
6
7         1            1      1.0      0.0      n = 10001
8         1           22     22.0      0.0      maximum = n * (log(n) + log(log(n)))
9         1          373    373.0      0.0      marked = [False] * int(maximum // 2)
10     57160        37122      0.6      0.3      for i in range(1, len(marked)+1):
11    188523      8444200     44.8     62.9          for j in range(1, i+1):
12    188355       118260      0.6      0.9              m = i + j + 2*i*j
13    188355      4672922     24.8     34.8              if m > len(marked): break
14    131364        75962      0.6      0.6              marked[m-1] = True
15
16
17         1            1      1.0      0.0      count = 1
18     52371        25618      0.5      0.2      for i in range(1, len(marked)+1):
19     52371        28393      0.5      0.2          if not marked[i-1]: count += 1
20     52371        26003      0.5      0.2          if count == n:
21         1           60     60.0      0.0              print(2*i + 1)
22         1          329    329.0      0.0              break


So our most likely offending line is:

11    188523      8444200     44.8     62.9          for j in range(1, i+1):


On Python 2 there is both range and xrange. Using

for i in range(...):


will generate a list of numbers and then loop over them, whereas

for i in xrange(...):


will just increment i as wanted each time. This is often much faster. So let's add:

# Set range to xrange on Python 2
try:
range = xrange
except NameError:
pass


to the top of the script and retime:

$time python2 p.py 104743 python2 p.py 0.09s user 0.00s system 96% cpu 0.090 total$ time python3 p.py
104743
python3 p.py  0.16s user 0.00s system 99% cpu 0.161 total

$time pypy p.py 104743 pypy p.py 0.04s user 0.00s system 93% cpu 0.050 total$ time pypy3 p.py
104743
pypy3 p.py  0.04s user 0.01s system 99% cpu 0.054 total


Yup. CPython is much improved.

So how would we make the code good? First you can transform the i+j+2ij calculation to directly use a stepping range:

step = 2*i + 1
for m in range(i+step-1, i+i*step, step):
if m >= len(marked):
break

marked[m] = True


Then you can just do a slice assignment:

for i in range(1, len(marked)+1):
step = 2*i + 1
marked[i+step-1::step] = [True] * len(marked[i+step-1::step])


and you can immediately see that you can stop when...

    if i+step-1 >= len(marked):
break


This leads to calculating the bound in the for loop itself. step = 2*i + 1, so i+step-1 = i+(2*i+1)-1 = 3*i, so we should stop before len(marked)//3:

for i in range(1, len(marked)//3):
step = 2*i + 1
marked[i+step-1::step] = [True] * len(marked[i+step-1::step])


The count part can use enumerate:

count = 1
for i, mark in enumerate(marked, 1):
if not mark:
count += 1

if count == n:
print(2*i + 1)
break


I suggest inverting the True/False in the array to make this simpler, and then you can just do count += mark. Why? Well, it lets us use itertools.accumulate on Python 3:

for i, count in enumerate(accumulate(marked), 1):
if count+1 == n:
print(2*i + 1)
break


You might also think len(marked[i+step-1::step]) is wasteful, since it's generating a new array each time: it is! It's still really fast, but Jaime's solution is better... except it involves risky math steps. How could we get the best of both?

Well, remember how in Python 3 range is a fancy "pseudo-list" and doesn't build a real list? Python 3's range can be sliced, too. We just need to build a range the same length as the list and slice that instead. This is actually slower on CPython (the one invoked as python3) for normal-sized lists because CPython's list slices are really fast, but on PyPy the overheads of the range version are optimized out:

from itertools import accumulate
from math import log

n = 1000001
maximum = n * (log(n) + log(log(n)))
marked = [True] * int(maximum // 2)

slice_proxy = range(len(marked))
for i in range(1, len(marked)//3):
step = 2*i + 1
marked[i+step-1::step] = [False] * len(slice_proxy[i+step-1::step])

for i, count in enumerate(accumulate(marked), 1):
if count+1 == n:
print(2*i + 1)
break


(Note the much larger n)

Times without slice_proxy:

$time python3 p.py 15485867 python3 p.py 5.57s user 0.01s system 100% cpu 5.576 total$ time pypy3 p.py
15485867
pypy3 p.py  3.58s user 0.08s system 100% cpu 3.654 total


Times with slice_proxy:

$time python3 p.py 15485867 python3 p.py 5.94s user 0.02s system 100% cpu 5.949 total$ time pypy3 p.py
15485867
pypy3 p.py  2.52s user 0.06s system 100% cpu 2.585 total


Then just wrap the whole thing in a main function, because local variables are faster than global ones:

from itertools import accumulate
from math import log

def main():
# ...

main()


Times:

$time python3 p.py 15485867 python3 p.py 4.74s user 0.02s system 100% cpu 4.754 total$ time pypy3 p.py
15485867
pypy3 p.py  2.40s user 0.06s system 100% cpu 2.456 total


But wait! We dropped the end point!

for i in range(1, len(marked)//3):
step  = 2*i + 1
start = i+step-1
stop  = 2*i*(i+1)
marked[start:stop:step] = [False] * len(slice_proxy[start:stop:step])


This is important on PyPy; on CPython it's about the same speed:

$time python3 p.py 15485867 python3 p.py 4.88s user 0.01s system 100% cpu 4.890 total$ time pypy3 p.py
15485867
pypy3 p.py  1.44s user 0.02s system 99% cpu 1.461 total


It's not really faster on CPython because the number of things you do tends to be more important than what you do, simply from interpreter overhead.

We know it's way faster on CPython, but how does this compare to the original version on PyPy (once stuck inside a main function)?

from math import log

def main():
n = 1000001
maximum = n * (log(n) + log(log(n)))
marked = [False] * int(maximum // 2)

for i in range(1, len(marked)+1):
for j in range(1, i+1):
m = i + j + 2*i*j

if m > len(marked):
break

marked[m-1] = True

count = 1
for i in range(1, len(marked)+1):
if not marked[i-1]:
count += 1

if count == n:
print(2*i + 1)
break

main()


Times:

$time pypy p.py 15485867 pypy p.py 0.83s user 0.02s system 99% cpu 0.851 total$ time pypy3 p.py
15485867
pypy3 p.py  0.78s user 0.01s system 99% cpu 0.790 total


Interestingly, the original is significantly faster on PyPy. This implies that unless you want a code base optimized for both CPython and PyPy, some of these optimizations (like slice_proxy) just aren't needed.

But, eh, what if this isn't fast enough?

Eh?

Well, let's quickly draw the things we're setting. T are the parts we set. . is the part we would set if not for the :end: optimization.

index:  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
i=1         T        .        .        .        .        .        .        .        .        .
i=2                  T              T              .              .              .              .
i=3                           T                    T                    T                    .
i=4                                    T                          T                          T
i=5                                             T                                T
i=6                                                      T                                      T
i=7                                                               T
i=8                                                                        T
i=9                                                                                 T
i=10                                                                                         T
i=11                                                                                                  T


We're doing i=1 then i=2 then i=3, etc. From i² ≈ len(marked) onwards, though, there aren't many things set each run. We end up moving back-and-forth in the array a lot, spending more time "seeking" in memory than actually setting items. Consider, though, this decomposition:

index:  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
i=1         1        2        3        4        5        6        7        8        9       10
i=2                  1              2              3              4              5              6
i=3                           1                    2                    3                    4
i=4                                    1                          2                          3
i=5                                             1                                2
i=6                                                      1                                      2
i=7                                                               1
i=8                                                                        1
i=9                                                                                 1
i=10                                                                                         1
i=11                                                                                                  1


For low numbers, 1, 2, 3, ... this hits a lot of those later values. We can then split the work into an upper part:

  i=1         T        .        .        .        .        .        .        .        .        .
i=2                  T              T              .              .              .              .
i=3                           T                    T                    T                    .
i=4                                    T                          T                          T
i=5                                             T                                T


and a lower part:

  i=6                                                      1                                      2
i=7                                                               1
i=8                                                                        1
i=9                                                                                 1
i=10                                                                                         1
i=11                                                                                                  1


This sets the same values, but does so in a different order.

Here's the code:

from itertools import accumulate
from math import log

def main():
n = 1000001
maximum = n * (log(n) + log(log(n)))
marked = [True] * int(maximum // 2)

slice_proxy = range(len(marked))
for i in range(1, int(len(marked)**0.5)):
step  = 2*i + 1
start = 3*i
stop  = 2*i*(i+1)
marked[start:stop:step] = [False] * len(slice_proxy[start:stop:step])

i += 1
start_step  = 2*i + 1
start_start = 3*i
step = 3
for start in range(start_start, len(marked), start_step):
marked[start::step] = [False] * len(slice_proxy[start::step])
step += 2

for i, count in enumerate(accumulate(marked), 1):
if count+1 == n:
print(2*i + 1)
break

main()


How fast is it?

$time python3 p.py 15485867 python3 p.py 1.53s user 0.01s system 99% cpu 1.538 total$ time pypy3 p.py
15485867
pypy3 p.py  1.26s user 0.06s system 100% cpu 1.319 total


CPython improves a lot with this, although PyPy doesn't really improve. This makes sense: you do a lot fewer operations but each operation is more complicated. Since CPython has a really fast implementation under the hood but a really slow interpreter, fewer operations is important. PyPy, however, is bound by the size of the operations themselves.

Of course, for PyPy you'd want to modify the "original" code:

from math import log

def main():
n = 1000001
maximum = n * (log(n) + log(log(n)))
marked = [True] * int(maximum // 2)

for i in range(1, int(len(marked)**0.5)):
step  = 2*i + 1
start = 3*i
stop  = 2*i*(i+1)

if stop > len(marked):
stop = len(marked)

for m in range(start, stop, step):
marked[m] = False

i += 1
start_step  = 2*i + 1
start_start = 3*i
step = 3
for start in range(start_start, len(marked), start_step):
for m in range(start, len(marked), step):
marked[m] = False

step += 2

count = 1
for i in range(1, len(marked)+1):
if marked[i-1]:
count += 1

if count == n:
print(2*i + 1)
break

main()


Which gives

$time pypy p.py 15485867 pypy p.py 0.65s user 0.02s system 99% cpu 0.672 total$ time pypy3 p.py
15485867
pypy3 p.py  0.61s user 0.01s system 99% cpu 0.625 total


The difference is obviously much less, but it exists. This is probably due to improvements in memory locality. It actually turns out that just terminating the range with len(marked)//3 gets you almost to this point, so the optimization, IMHO, isn't worth it.

Sadly, CPython still trails the first version on PyPy. Janne Karila gave a good suggestion in the comments for speeding up CPython. The key part is to use itertools (or filter) to apply a filter and islice to do the counting. The large advantage comes from the iteration and counting being done inside C routines.

I went with compress from Python 3.1 and Janne Karila's use of islice to get the wanted element:

primes = compress(range(3, int(maximum), 2), marked)
print(next(islice(primes, n-2, None)))


Timings are really good:

$time python3 p.py 15485867 python3 p.py 0.58s user 0.01s system 99% cpu 0.598 total  This is (negligibly) faster than PyPy! But, eh, but... but... What if this isn't fast enough? Eh? Well, that's what Numpy is for. Numpy is Python's fast array library. from math import log import numpy def main(): n = 1000001 maximum = n * (log(n) + log(log(n))) marked = numpy.ones(int(maximum // 2), dtype=bool) for i in range(1, int(len(marked)**0.5)): step = 2*i + 1 start = 3*i stop = 2*i*(i+1) marked[start:stop:step] = False i += 1 start_step = 2*i + 1 start_start = 3*i step = 3 for start in range(start_start, len(marked), start_step): marked[start::step] = False step += 2 accumulated = numpy.add.accumulate(marked) [where] = numpy.where(accumulated+1 == n) print(2 * where[0] + 3) main()  This also works on Python 2. Unfortunately PyPy doesn't (yet) have an easy Numpy port. Even so, the times: $ time python3 p.py
15485867
python3 p.py  0.23s user 0.02s system 99% cpu 0.252 total

$time python2 p.py 15485867 python2 p.py 0.17s user 0.02s system 98% cpu 0.196 total  C++ with this N takes: $ g++ -O3 c.cpp -o c; time ./c
15485867
./c  0.06s user 0.00s system 96% cpu 0.065 total


Neat, but we're still not done. One slow part is

accumulated = numpy.add.accumulate(marked)
[where] = numpy.where(accumulated+1 == n)
print(2 * where[0] + 3)


It turns out that Numpy's count_nonzero is really fast, so let's try capitalizing on that with what amounts to a binary search. There are similarities to Quickselect:

# I can't think of a good name...
def count_nonzero_sorted(array, count):
# We always assume the count is achievable
if array.shape[0] == 1:
return 0

# Split the array in two
midpoint = len(array) // 2
first_half = array[:midpoint]
second_half = array[midpoint:]

first_counts = numpy.count_nonzero(first_half)

# If the counts in the first half is greater
# than the wanted index, check that half
if first_counts >= count:
return count_nonzero_sorted(first_half, count)

# else check the other half and offset the result
return midpoint + count_nonzero_sorted(second_half, count-first_counts)


This gives

# compress sieve
where = count_nonzero_sorted(marked, n-1)
print(2 * where + 3)


and timings of just

$time python2 p.py 15485867 python2 p.py 0.10s user 0.01s system 97% cpu 0.116 total$ time python3 p.py
15485867
python3 p.py  0.16s user 0.01s system 99% cpu 0.167 total


On Python 2 that's half the speed of C++! Vectorization is a wonderful thing.

It would be interesting how many of these optimizations can apply to R. A preliminary guess puts the time required for the original R code to run as about 7 hours.

What should you take away from this?

1. PyPy is really, really fast. The original code (after the bug is fixed) takes only 0.7-0.9 seconds, or a tenth the speed of C++.

PyPy thus ends up, by my guess, about 10000 times as fast as the implementation of R that I have available (for this task).

2. Optimizing PyPy is possible but it's hard; the optimized PyPy was only a little faster than the original PyPy code at 0.6-0.7s. This is because the JIT gives you most optimizations for free.

3. CPython has sharp corners (range vs xrange) but Python 3 takes some of these problems away by default. Once you handle this, CPython is an order of magnitude faster than R on this code.

4. CPython can be forced to go fast if you know what you are doing. When forced it became a touch faster than PyPy.

5. CPython is ridiculously fast when Numpy is used; getting to half the speed of C++.

I thought to check how a traditional sieve compares against this:

import sys
from itertools import compress, islice
from math import log

def main():
n = 1000001
maximum = int(n * (log(n) + log(log(n))))
maxidx = maximum//2
sieve = [True] * maxidx # 2, 3, 5, 7... might be prime

j = 0
for i in range(1, int(maxidx**0.5)+1):
j += 4*i

if sieve[i]:
step = 2*i + 1
sieve[j::step] = [False] * -(-(maxidx-j) // step)

# compress sieve
print(next(islice(compress(range(1, maximum, 2), sieve), n-1, None)))

main()


and the PyPy version:

import sys
from itertools import compress, islice
from math import log

def main():
n = 1000001
maximum = int(n * (log(n) + log(log(n))))
maxidx = maximum//2
sieve = [True] * maxidx # 2, 3, 5, 7... might be prime

j = 0
for i in range(1, int(maxidx**0.5)+1):
j += 4*i

if sieve[i]:
for k in range(j, maxidx, 2*i + 1):
sieve[k] = False

# compress sieve
count = 1
for i in range(1, maxidx):
count += sieve[i]

if count == n:
print(2*i + 1)
break

main()


and the Numpy version:

import numpy
from math import log

def count_nonzero_sorted(array, count):
# We always assume the count is achievable
if array.shape[0] == 1:
return 0

midpoint = len(array) // 2
first_half = array[:midpoint]
second_half = array[midpoint:]

first_counts = numpy.count_nonzero(first_half)

if first_counts >= count:
return count_nonzero_sorted(first_half, count)

return midpoint + count_nonzero_sorted(second_half, count-first_counts)

def main():
n = 1000001
maximum = int(n * (log(n) + log(log(n))))
maxidx = maximum//2
sieve = numpy.ones(maxidx, dtype=bool) # 2, 3, 5, 7... might be prime

j = 0
for i in range(1, int(maxidx**0.5)+1):
j += 4*i

if sieve[i]:
sieve[j::2*i + 1] = False

# compress sieve
where = count_nonzero_sorted(sieve, n)
print(2 * where + 1)

main()


Timings:

$time python3 p.py 15485867 python3 p.py 0.40s user 0.01s system 99% cpu 0.415 total$ time pypy3 p.py
15485867
pypy3 p.py  0.36s user 0.01s system 99% cpu 0.368 total

$# Numpy versions$ time python2 p.py
15485867
python2 p.py  0.07s user 0.01s system 97% cpu 0.075 total

\$ time python3 p.py
15485867
python3 p.py  0.12s user 0.00s system 99% cpu 0.124 total


Sorry, Sieve of Sundaram, you're not that great after all. :P

And, heck, that Python 2 version is way too fast.

Note: I made a mistake previously in comparing vector<bool> to vector<int>; I should have compared it to vector<uint8_t> instead. It turns out that the difference is actually relatively minor (30-40%) and although significant it isn't nearly the factor-of-4 difference from before.

• This is the best response that I've seen on CodeReview so far. I have learned a great deal about how to improve my Python style and use of lists. All of this is new to me. Thank you for the Herculean effort that you put into writing this. – castle-bravo Nov 29 '14 at 22:57
• I get a nice speedup on Python 3, starting from the "here's the code" version, by initializing marked = list(range(1, int(maximum // 2) + 1)) and replacing the last loop by i = next(islice(filter(None, marked), n-2, None)) – Janne Karila Dec 1 '14 at 14:42
• @JanneKarila Thanks; I've incorporated (and built on) this suggestion. – Veedrac Dec 2 '14 at 12:39
• Nice answer. As far as I can tell, len(mark) doesn't change at all. Why not put in in a variable instead of calculating it at every loop? – Eric Duminil Jan 21 '18 at 18:38
• @EricDuminil len on an array is effectively just an attribute lookup, and is unlikely to be expensive unless called inside the innermost loop. – Veedrac Jan 21 '18 at 19:08

Explicit looping in Python is (very) slow, and you have a double loop, which isn't helping. Your inner loop is effectively setting to True a slice of marked of the form marked[3*i+1::2*i+1], and you can take advantage of that to speed things up considerably. The only tricky part is figuring out how many items you have to set, to properly build the right hand side of the assignment. It is not hard to figure out that if you have a list withlen_ items, and you set every step-th item starting with start, you will in total set ceil((len_-start) / step) elements, which can be computed directly as (len_-start-1)//step + 1. Putting it all together:

from math import log
from itertools import repeat

n = 10001
max_ = n*(log(n) + log(log(n)))
len_ = int(max_/2)
marked = [False] * len_
for i in range(1, len_):
step = 2*i + 1
start = step + i
marked[start::step] = [True] * ((len_-start-1)//step+1)

count = 0
for i in range(0, len(marked)):
if not marked[i]:
count += 1
if count == n:
print 2*i + 1
break


Which prints out the exact same value in a fraction of a second.

The Python version isn't exactly the same as the C++ one.

C++:

  if (m < marked.size()) {
marked[m] = true;
} else {
break;
}


Python:

if m > len(marked):
break
marked[m] = True


So we want to change part of the inner loop to this:

if m < len(marked):
marked[m] = True
else:
break


To speed it up, just change range to xrange.

When you call range, it creates a list containing so many number (int, long, or float) objects. All of those objects are created at once, and all of them exist at the same time. This can be a pain when the number of numbers is large.

xrange, on the other hand, creates no numbers immediately - only the range object itself. Number objects are created only when you pull on the generator, e.g. by looping through it.

from math import log

n = 10001
max = n*(log(n) + log(log(n)))
marked = [False] * int(max/2)
for i in xrange(1, len(marked)):
for j in xrange(1, i + 1):
m = i + j + 2*i*j
if m < len(marked):
marked[m] = True
else:
break
count = 0
for i in xrange(0, len(marked)):
if not marked[i]:
count += 1
if count == n:
print 2*i + 1
break


On my machine: 0.15s user 0.01s system 97% cpu 0.162 total.

Although your main question has been answered, all of these implementations have a common shortcoming: no need to do the count == n check if count hasn't changed.

  if (!marked[i]) count <- count + 1
if (count == n) {
print(2*i + 1)
break
}


This is better:

  if (!marked[i]) {
count <- count + 1
if (count == n) {
print(2*i + 1)
break
}
}


• using namespace std is considered bad practice
• !marked[i] is more common (and thus more natural) than not marked[i]
• Since the size of marked is decided in advance, it might be interesting to try a boolean array for it instead of a vector
• vector<bool> is special cased to a bit vector; I'd assume (naïvely perhaps) that this will end up more important than any gains from having a "raw" array. – Veedrac Nov 29 '14 at 8:05