What is the fastest way to count the unique elements in a list of billion elements?
My problem is not usual. Let's imagine few billions of strings. Strings are usually less then 15 characters. In this list I need to find out the number of the unique elements.
First of all, what object should I use? You shouldn't forget if I add a new element I have to check if it is already existing in the list. It is not a problem in the beginning, but after few millions of words it can really slow down the process.
That's why I thought that Hashtable would be the ideal for this task because checking the list is ideally only log(1). Unfortunately a single object in .net can be only 2GB.
Next step will be to implement a custom hashtable which contains a list of 2GB hashtables.
I am wondering maybe some of you know a better solution. (Computer has extremely high specification.)
I would skip the data structures exercise and just use an SQL database. Why write another custom data structure that you have to analyze and debug, just use a database. They are really good at answering queries like this.
I'd consider a Trie or a Directed acyclic word graph which should be more space-efficient than a hash table. Testing for membership of a string would be O(len) where len is the length of the input string, which is probably the same as a string hashing function.
This can be solved in worst-case O(n) time using radix sort with counting sort as a stable sort for each character position. This is theoretically better than using a hash table (O(n) expected but not guaranteed) or mergesort (O(n log n)). Using a trie would also result in a worst-case O(n)-time solution (constant-time lookup over n keys, since all strings have a bounded length that's a small constant), so this is comparable. I'm not sure how they compare in practice. Radix sort is also fairly easy to implement and there are plenty of existing implementations.
If all strings are d characters or shorter, and the number of distinct characters is k, then radix sort takes O(d (n + k)) time to sort n keys. After sorting, you can traverse the sorted list in O(n) time and increment a counter every time you get to a new string. This would be the number of distinct strings. Since d is ~15 and k is relatively small compared to n (a billion), the running time is not too bad.
This uses O(dn) space though (to hold each string), so it's less space-efficient than tries.
If the items are strings, which are comparable... then I would suggest abandoning the idea of a Hashtable and going with something more like a Binary Search Tree. There are several implementations out there in C# (none that come built into the Framework). Be sure to get one that is balanced, like a Red Black Tree or an AVL Tree.
The advantage is that each object in the tree is relatively small (only contains it's object, and a link to its parent and two leaves), so you can have a whole slew of them.
Also, because it's sorted, the retrieval and insertion time are both O log(n).
Since you specify that a single object cannot contain all of the strings, I would presume that you have the strings on disk or some other external memory. In that case I would probably go with sorting. From a sorted list it is simple to extract the unique elements. Merge sorting is popular for external sorts, and needs only an amount of extra space equal to what you have. Start by dividing the input into pieces that fit into memory, sort those and then start merging.
With a few billion strings, if even a few percent are unique, the chances of a hash collision are pretty high (.NET hash codes are 32-bit int, yielding roughly 4 billion unique hash values. If you have as few as 100 million unique strings, the risk of hash collision may be unacceptably high). Statistics isn't my strongest point, but some google research turns up that the probability of a collision for a perfectly distributed 32-bit hash is (N - 1) / 2^32, where N is the number of unique things that are hashed.
You run a MUCH lower probability of a hash collision using an algorithm that uses significantly more bits, such as SHA-1.
Assuming an adequate hash algorithm, one simple approach close to what you have already tried would be to create an array of hash tables. Divide possible hash values into enough numeric ranges so that any given block will not exceed the 2GB limit per object. Select the correct hash table based on the value of the hash, then search in that hash table. For example, you might create 256 hash tables and use (HashValue)%256 to get a hash table number from 0..255. Use that same algorithm when assigning a string to a bucket, and when checking/retrieving it.
divide and conquer - partition data by first 2 letters (say)
dictionary of xx=>dictionary of string=> count
A Dictionary<> is internally organized as a list of lists. You won't get close to the (2GB/8)^2 limit on a 64-bit machine.
I would use a database, any database would do.
Probably the fastest because modern databases are optimized for speed and memory usage.
You need only one column with index, and then you can count the number of records.
Have you tried a Hash-map (Dictionary in .Net)? Dictionary<String, byte> would only take up 5 bytes per entry on x86 (4 for the pointer to the string pool, 1 for the byte), which is about 400M elements. If there are many duplicates, they should be able to fit. Implementation-wise, it might be verrryy slow (or not work), since you also need to store all those strings in memory.
If the strings are very similar, you could also write your own Trie implementation.
Otherwise, you best bets would be to sort the data in-place on disk (after which counting unique elements is trivial), or use a lower-level, more memory-tight language like C++.
I agree with the other posters regarding a database solution, but further to that, a reasonably-intelligent use of triggers, and a potentially-cute indexing scheme (i.e. a numerical representation of the strings) would be the fastest approach, IMHO.
+1 for the SQL/Db solutions, keeps things simple --will allow you to focus on the real task at hand.
But just for academic purposes, I will like to add my 2 cents.
-1 for hashtables. (I cannot vote down yet). Because they are implemented using buckets, the storage cost can be huge in many practical implementation. Plus I agree with Eric J, the chances of collisions will undermine the time efficiency advantages.
Lee, the construction of a trie or DAWG will take up space as well as some extra time (initialization latency). If that is not an issue (that will be the case when you may need to perform search like operations on the set of strings in the future as well and you have ample memory available), tries can be a good choice.
Space will be the problem with Radix sort or similar implementations (as mentioned by KirarinSnow) because the dataset is huge.
The below is my solution for a one time duplicate counting with limits on how much space can be used.
If we have the storage available for holding 1 billion elements in my memory, we can go for sorting them in place by heap-sort in Θ(n log n) time and then by simply traversing the collection once in O(n) time and doing this:
if (a[i] == a[i+1]) dupCount++;
If we do not have that much memory available, we can divide the input file on disk into smaller files (till the size becomes small enough to hold the collection in memory); then sort each such small file by using the above technique; then merge them together. This requires many passes on the main input file.
I will like to keep away from quick-sort because the dataset is huge. If I could squeeze in some memory for the second case, I would better use it to reduce the number of passes rather than waste it in merge-sort/quick-sort (actually, it depends heavily on the type of input we have at hand).
Edit: SQl/DB solutions are good only when you need to store this data for a long duration.