I actually have an answer to my question but it is not parallelized so I am interested in ways to improve the algorithm. Anyway it might be useful as-is for some people.
int Until = 20000000;
BitArray PrimeBits = new BitArray(Until, true);
/*
* Sieve of Eratosthenes
* PrimeBits is a simple BitArray where all bit is an integer
* and we mark composite numbers as false
*/
PrimeBits.Set(0, false); // You don't actually need this, just
PrimeBits.Set(1, false); // remindig you that 2 is the smallest prime
for (int P = 2; P < (int)Math.Sqrt(Until) + 1; P++)
if (PrimeBits.Get(P))
// These are going to be the multiples of P if it is a prime
for (int PMultiply = P * 2; PMultiply < Until; PMultiply += P)
PrimeBits.Set(PMultiply, false);
// We use this to store the actual prime numbers
List<int> Primes = new List<int>();
for (int i = 2; i < Until; i++)
if (PrimeBits.Get(i))
Primes.Add(i);
Maybe I could use multiple BitArray
s and BitArray.And() them together?
@DrPizza Profiling only really helps improve an implementation, it doesn't reveal opportunities for parallel execution, or suggest better algorithms (unless you've experience to the otherwise, in which case I'd really like to see your profiler).
I've only single core machines at home, but ran a Java equivalent of your BitArray sieve, and a single threaded version of the inversion of the sieve - holding the marking primes in an array, and using a wheel to reduce the search space by a factor of five, then marking a bit array in increments of the wheel using each marking prime. It also reduces storage to O(sqrt(N)) instead of O(N), which helps both in terms of the largest N, paging, and bandwidth.
For medium values of N (1e8 to 1e12), the primes up to sqrt(N) can be found quite quickly, and after that you should be able to parallelise the subsequent search on the CPU quite easily. On my single core machine, the wheel approach finds primes up to 1e9 in 28s, whereas your sieve (after moving the sqrt out of the loop) takes 86s - the improvement is due to the wheel; the inversion means you can handle N larger than 2^32 but makes it slower. Code can be found here. You could parallelise the output of the results from the naive sieve after you go past sqrt(N) too, as the bit array is not modified after that point; but once you are dealing with N large enough for it to matter the array size is too big for ints.
There's a very good article about the Sieve of Eratosthenes: The Genuine Sieve of Eratosthenes
It's in a functional setting, but most of the opimization do also apply to a procedural implementation in C#.
The two most important optimizations are to start crossing out at P^2 instead of 2*P and to use a wheel for the next prime numbers.
For concurrency, you can process all numbers till P^2 in parallel to P without doing any unnecessary work.
You also should consider a possible change of algorithms.
Consider that it may be cheaper to simply add the elements to your list, as you find them.
Perhaps preallocating space for your list, will make it cheaper to build/populate.
Are you trying to find new primes? This may sound stupid, but you might be able to load up some sort of a data structure with known primes. I am sure someone out there has a list. It might be a much easier problem to find existing numbers that calculate new ones.
You might also look at Microsofts Parallel FX Library for making your existing code multi-threaded to take advantage of multi-core systems. With minimal code changes you can make you for loops multi-threaded.
You might save some time by cross-referencing your bit array with a doubly-linked list, so you can more quickly advance to the next prime.
Also, in eliminating later composites once you hit a new prime p for the first time - the first composite multiple of p remaining will be p*p, since everything before that has already been eliminated. In fact, you only need to multiply p by all the remaining potential primes that are left after it in the list, stopping as soon as your product is out of range (larger than Until).
There are also some good probabilistic algorithms out there, such as the Miller-Rabin test. The wikipedia page is a good introduction.