Slashdot Mirror
Probing Hash Tables?
David Rusenko asks: "I've been taking a datastructures class at CMU as part of a summer CS program. One of these structures we have gone over is hash tables. After going through many different probing methods (linear, quadratic), multiple hash functions, and double hashing, I was all too curious to know if these are the best methods currently known. Some other interesting ideas came up, such as using the Fibonacci numbers for probing, but I haven't had time to test them yet. Any comments?"
Do yourself a favour and read the bible on it. The information might be a bit dated, but I doubt hashing has changed that much since the last edition.
What about a linked list with an array?
What about another hash table that holds the conflicts (kind of like double hash functions but cooler)?
As I dig out my Algorithms textbook, I see mention that Fibonacci heaps give fast times in Dijkstra and Prim's algorithms for Shortest dag path and Minimum Spanning Trees respectively: O(m+nlogn) in both cases. This is very fast for dense graphs.
Shrug.
--onyx--
What's that about a little bit of knowledge and it being dangerous. In school you'll beat every algorithm under the sun to death. In the real world you'll link to the STL and use a hash map.
Write well-designed, clean, maintainable code. Then profile it. If your table lookup blips on the profiler, then *THINK* about optimizing it. After you've *THOUGHT* about optimizing it, then decide if it makes sense to squeeze the time and effort into the schedule.
Random quotes:
Premature optimization is the root of all evil. -- Knuth
There is never a best solution, only tradeoffs to consider. --Eberly
The best optimizer is between your ears. --Abrash
The Ferrai's are only gravy, honest! --Carmack
Alright I threw the last one in for shits and giggles. Don't blame me, I can't get through a Sunday night without drinking a poop load of beer. Anyway, the point is that there is a difference between *FAST* and *FAST ENOUGH*. If it's fast enough, who cares.
[CYNICISM]
Unless you're an academic and you're looking for funding.
[/CYNICISM]
That said, the most thorough treatment I remember about probing was from David Gries' book Compiler Construction for Digital Computers, published in 1971 when memory wasn't so cheap. It's probably long out of print by now, because that stuff isn't really important any more.
Similarly, a lot of the stuff in Knuth vol. 3 is about sorting data on magtape, which was important 20 years ago but nobody cares about now. In the introduction to the second edition, Knuth says he left the material in just because it's mathematically beautiful and because tape-like media may make a comeback, but it's possible that he'll remove it from future editions.
In the worst case, double hashing is better than quadratic probing, and quadratic probing is better than linear probing. But the trick for finding the "right" method is to ensure that the worst case never happens. Then you can avoid expensive advanced stuff, and keep it simple and lean.
If you don't know much about your data, or your keys, or your insertion/removal pattern, than a separate chaining scheme might be best. If you know everything, then you can use a perfect hash function generator (but even that doesn't necessarily guarantee best performance). The reason they teach you several different methods is because there is no single "best" answer.
I better add before someone makes a tragic mistake, enlarge the table only if you're getting enough collisions to actually slow your program down enough to matter. If you have a lot of entries you'll get a few collisions even with an enormous table that's mostly empty, because of the birthday paradox. Don't worry about it til you're getting collisions on a significant fraction of your lookups.
Higher performance applications call for probing. It's a little harder to program though. One of my CS professors worked on speech recognition for IBM and they used hash tables for n-gram language modeling, and probed using double hasing I believe. The largest problem with probing is removal, but their program didn't actually remove anything from the table.
Ok, am I the only person who saw this article's title as being a bit sexually suggestive?
Ugh. I must have spent far too much time on the computer again this weekend... :)
Please consider making an automatic monthly recurring donation to the EFF
CAM tables if you're gluing together your own hardware. kinda like a hash table in hardware that never gets collisions until it is full.
the insertion/deletion maintance isn't free so you gotta be careful with your search/modify ratio.
Unfortunately, by reinventing the wheel, all you've succeeded in doing is confusing the hell out of anyone who has to maintain your code in the future.
I hope you commented real-world uses of this profusely, otherwise you may start receiving suspicious packages in the mail from those who take over for you some day.
- A.P.
"Remember when the U.S. had a drug problem, and then we declared a War On Drugs, and now you can't buy drugs anymore?"
Additionally, it would seem to me that obtaining the Nth element of this "hash" would be an O(N) operation, which provides, let me calcuate.... yes, zero gain over using a standard array.
The only real advantage I see is a short-term increase in job security due to the likelihood that people will look at that and say 'What the fuck is he doing this for?' and thinking you're far more clever than you are.
For example, if your data is composed primairly of upper case squences of A T G C (e.g., genetic code) you would tend to have long elements of highly repeatable letters.
If, on the other hand, you were memorizing, say, a binary image for uniqueness (such as in a virus scanner) you would have large files with binary data.
Thus... each data type possible can be *tuned* to be more efficient when hashed, based on what you know of the incoming data. To ask for a *generic* algorithm that works well on all data will automatically result in less efficiency.
The real secret to good hashing is to allow two things... first, allow the algorithm (usually hash length and table size) to be modifyable by the code. Second, allow simple statistics to be kept on collision rate and maximum child length - these two statistics can be calculated very very quickly at the key-add phase of the hashing, and only add a few instructions to the process.
Now... throw a good deal of data at your hash table... all sorts of data that represent the type you EXPECT to get. Tune the hash table algorithm (using the exposed algorithm hooks) until your statistics see a collision of between 2 and 5 for all positions in the table (in general, our hashers, and we use LOTS of hashers, rarely go beyond 3 to 5 children).
The tradeoff is obviously memory (table size) versus efficiency of the hash algorithm.
One of my favorite hash algorithms for trivial hashing (say, hashing of a label or variable name) is simply incremental add and xor. Very quick and usually generates a good spread (this is only for simple hashing).
MORE IMPORTANTLY (to me at least) is how you STORE your hashing and data info. We *always* store the data as a (void *) pointer. By doing so we can store ANY type of element, be it a structure, a function pointer, a pointer to text data, a LONG or a INT or a DOUBLE, doesn't matter, because we always store the POINTER to the data, not the data itself. This is an extremely powerful concept.
One last thought... one of the more interesting things you can do with hash tables is to use them as on-the-fly indexed arrays that can grow to any length. In this case, the hash code becomes the index you want to store in. This is an interesting concept because it means you can create an array that grows in real-time. For example, you can store in hash_array[1] and hash_array[5923] and it will not take 5924 positions, it will only take two positions. And reading the two items only takes 2 instances of the loop, not 5924 instances. The array grows at will, taking only as much memory as what you require. Obviously, for this to work, you are hashing a small structure that contains the real data AND the real index, thus during a collision you then do a compare for the real index down that childs tree. But this is extremely quick and low overhead and solves an infinite array with gusto.
Aloha
It's just a repeating pop-up.
In fact, you just took out IE on my machine in the middle of another post. Thanks a bunch, fuckwit.
If you disagree, post your argument. (-1, Overrated) isn't your personal censorship tool for views you don't like.
I hate to be churlish about this, but none of the early posts have addressed the core issues. Knuth's treatment is rather narrow. Buckets have very little to do with it.
The question to address is your key structure. Ideally you have the notion of your keys in cannonical representation. In object oriented contexts, the byte representation of your objects is not necessarily cannonical.
Next step is to analyze the size and distribution of your key space in cannonical representation, using as a function of some N which represents the scale of the problem instance.
At this point, if your the size of your key space is a weak function of N life is easy. Weak functions are where the average bit size of your cannonical keys is logarithmic in N or at worst sqrt(N). This represents the order of key entropy extraction.
A best case scenario is where all your keys are eight byte fully randomized GUIDs. Entropy extraction from your keys can be handled in just a few machine instructions. Worst case: you have to gather your key entropy by traversing deeply linked object hierarchies, in which case the efficiency of bucket access is swamped by the cost of constructing the bucket index.
When life is ugly, the games begin. There are no end of variations on how you can arrange to collect enough entropy (most of the time) for each key (most of the time) quickly enough (most of the time).
An extreme measure involves memoizing your key entropy with all the hassles that entails of making sure every modification to the key structure maintains key hash correspondence. If key modification is rare, you can really brute force this. If all your keys are always in the table, then keys can't change without rehashing (so one term starts to absorb another). On the other hand, if you have many keys to check, but few keys to check against, you might get sucked into finding clever and/or complicated methods for maintaining your key entropy memos.
If you have concocted a model with known degeneracies, who pays on the degenerate case? This is a game of hot potato, which again presents endless options, most of which are difficult to formally analyze (supposing you even have a sufficient model of your key space).
There are human hazards when you enter into this terrain not to be taken lightly. For some unknown reason a rather large slice of the population cannot comprehend any event more certain than 99% or less certain that 1%. pow(2,-6) is rounded up by these people to 1%, which in Murphy's calculas is as good as sunk. Nothing you can do about it. Sooner or later the pointy haired what-ifers will grind you down. Especially if they studied arithmetic for a couple of years as an undergraduate thinking it was mathematics.
Let's suppose now that you've done the dirty deed and concocted some method to extract at least log K bits of uniform key entropy (most of the time) where K somewhere between the total number of keys you might need to check and the total number of keys you might need to store. Only now does bucket management begin to matter.
Probably the best thing to do at this point is to slice the world into rough orders of magnitude: less than 10 CPU clock cycles to on-chip L1/L2 caches, 10-100 CPU clock cycles to on chip L3/off chip L3 cache, 100-1000 CPU cycles to external memory, 1000-10,000 cycles via interprocess communication/message passing, 10,000+ cycles for data structures not necessarily memory resident.
With a GHz class CPU, 1000 cycles still allows you to check one million keys per second. Do you really need to hone this down another 2.5 orders of magnitude? Sometimes you can if you really want to.
I could go on for a long time yet, but I've covered most of the ground that the rest of this thread has largely ignored.
On a more theoretical level, it is even possible to construct perfect hashing schemes with space efficiency near the Shannon limit for data sets where the average key entropy is less than one bit. Did someone mention Markov models for speech recognition? Forget Knuth volume 3. It has a lot more to do with Knuth volume 4.
pow(2,-6) was meant to be pow (2,-64) which rounds to 1% a with a great deal more drama.