3 ms·

I've done a bit of work with tries and hash functions and found that in-memory tries are almost universally faster than in-memory hashes, no matter how good the

• doty 17y ago
  I've done a bit of work with tries and hash functions and found that in-memory tries are almost universally faster than in-memory hashes, no matter how good the hash function is. Do you have any insight on why this would be? It's not very intuitive to me. A hash-table with low load has relatively small hash buckets and so touches fairly little memory, and string keys are fairly local, so I would have expected memory costs to dominate. Insertion and deletion from a hash table should require fewer allocations than insertion and deletion from a trie.
  • elblanco 17y ago
    see http://news.ycombinator.com/item?id=1240914 http://news.ycombinator.com/item?id=1240914
• AlisdairO 17y ago
  That's really interesting - but it's hard for me to understand. What makes them faster for lookup than a hash table with a really good, fast hash function (say, murmurhash)? Does it depend on factors like string length? Basically, what's the environment? My assumption is that it's faster because you might end up performing fewer string comparisons with the trie - if you have strings with common prefixes/sizes, this could cost you time on a hash with collisions. Although it seems like unless you have a lot of strings with common prefixes/sizes, this still ought to be pretty fast. And if you do have a lot of similar strings, you could store the hash value as well to speed this up. My work on cache-sensitive data structures is in the integer space, so perhaps my ignorance when it comes to working with strings is showing :-).
  • elblanco 17y ago
    Bingo, string comparisons will eat you alive. Even if you have a very small percentage of collisions in your hash (even close to 0%), you still have to check if you have a collision to determine if you need to iterate to the next slot in the hash no matter what. If you are hashing strings, you have to compare the entire string, each character comparison is guaranteed to be at least one integer sub operation for 1 byte wide character sets and 2 for unicode. But basically, tries compute almost as fast as just reading the string, while hashes have to read the string a few times AND compute the function AND deal with collisions (AND resize the hash which is stupid expensive). So to either insert or lookup in a hash you have to: 1) Compute the hash from the string, which naively involves reading the entire string into memory and computing the hash code. So for each character you have to move at least one or two bytes into memory, do something with it (likely some kind of integer op), store the result, then move the next byte or two in, do an operation with that and so on. Let's not forget that I also have to check for end of string, which uses at least one sub operation for each character as well for C style strings. Otherwise you keep a strlen counter in a register somewhere and do sub operations against that, either way. 1b)If you are smart you simply restrict the length you compute off of to some smaller length that will still do a pretty good job of assuring a reasonable level of uniqueness in the hash, but you can't ever guarantee it (uniqueness is the principle problem of hashing functions). 2) But still, let's assume I'm dealing with a nice small 16 character string (n=16). I have to compute many more operations than just the string length, at least 1 mov, 1 sub (to check for EOS), and then a few operations for your hash and probably some copies back into memory, so let's be generous and say 8 ops for each character. So if n=16 at least 128 ops to simply compute the hash. 3) Then we have to do some pointer arithmetic or some table lookups or whatever to get to the memory location and store the string. So in all we're talking about 500-1000 ops to insert a 16 character string without even checking for collisions. 4) Factor that in, assume no collision, we might be on the order of 1200 executed ops in the end. If we have a collision, a couple thousand or more. So for n=16, ops=~1200 minimum. 5) Even if there isn't a collision, I have to do a linear string compare to the hash->string value in the hash to assure me that there wasn't a collision. So we have to start moving integers into registers and doing a bunch of subs again (and checking for EOS on both strings now). This could easily introduce another 50-100 ops to ensure no collision. Lookups on the hash are also of a similar scale. Heaven help you if you have to expand the hash and move all that crap from one place to another. For a trie, there is no such thing as collisions. In the naive version, you simply allocate an array in memory the length of the character set + 1 in memory, which takes a couple of ops since you know the character set + 1 length ahead of time, use the character (I'm assuming 1 byte character lengths, but the principle is the same for 2 byte) as an operand in some simple pointer arithmetic. And store a memory location pointer in each array index to the next array (of character set + 1) until you finish off the string. If you are inserting, point to one more array and set that extra + 1 byte to something meaningful. If you are doing a lookup, you check that location to see if the byte is set to something meaningful (or you can just set that on the last character instead of one more down the tree, whatever). The point is, for a 16 character string you count say two ops for memory allocation, an op to find the slot in the array, and another op for allocation or value setting for each character. So in all, an insert or lookup happens almost as fast as a direct string compare, maybe on the order of n=16, ops=<100. Lookups can be even faster since you have have to short circuit the lookup (read: stop reading the string) as soon as you can't find just one character). I'm wagging a bit, I've never built either structure in asm, but if you analyze the number of operations each takes, the trie is insanely faster. Even with memory access lags. You can do some obvious speedups, like loading several characters into a register at a time for integer arithmetic, or loading the hash locations value chains or something... One of the confusing points on this is that the big-O estimates I've seen never bother to account for the string reads and comparisons. With a hash you basically have to do them a minimum of two times, with a trie, only once (and sometimes not even that much). Usually you see hashes as O(1) best and O(n) worst, but really they are O(k) and O(k*n) where k is the string length while tries are really O(k). All that being said, as soon as you try to work with either structure on disk, hashes easily overtake tries in terms of speed since hard drive random access times are much worse and tries require lots and lots and lots of random access (for each character). http://www.statemaster.com/encyclopedia/Trie http://www.statemaster.com/encyclopedia/Trie