I have a large list of strings stored in one huge memory block (usually there is 100k+ or even 1M+ of them). These are actually hashes, so the alphabet of the strings is limited to A-F0-9 and each string is exactly 32 bytes long (so its stored 'compressed'). I will call this list the main list from now on.
I want to be able to remove items from the main list. This will be usually done in bulks, so i get a large list (about 100 to 10k usually) of hashes which i need to find in this list and remove them. At the end of this operation there cannot be any empty blocks in the large memory block, so i need to take that into account. It is not guaranteed that all of the items will be in the main list, but none will be there multiple times. No rellocation can be done, the main block will always stay the same size.
The naive approach of iterating through the main list and checking if given hash shall be removed of course works, but is a bit slow. Also there is a bit too much moving of small memory blocks, because every time when a hash is flagged for removal i rewrite it with the last element of the main list, thus satisfying the condition of no empty blocks. This of course creates thousands of small memcpy's which in turn slow the thing down more because i get tons of cache misses.
Is there a better approach?
Some important notes:
the main list is not sorted and i cannot waste time sorting it, this
is a limitation imposed by the whole project and rewriting it so the
list is always sorted is not an option (it might not even be
possible)
memory is not really a problem, but the less is used the better
i can use STL, but not boost
Okay, here's what I'd do if I absolutely had to optimize the hell out of this.
I'm assuming order doesn't matter, which seems to be the case as you (IIUC) remove items by swapping them with the last item.
Store 128 bit integers (however you represent them, either your compiler supports them natively or you use a small array of 32/64 bit integers) instead of 32-char strings. See my comment on the question.
Roll my own hash set of 128 bit integers. Note that you can optimize a lot here if you're willing to think a bit, make some assumption, and get down 'n dirty. Some notes:
You only need to store the hashes themselves (for collision resolution), and a bit or two of metadata to identify deleted/unused slots. Have a look at what existing hash tables do if you're unsure how to guarantee correctness. I figure it's even simpler if you only ever delete (not add) after building the hash set. Though I think you could even do without that metadata if you had a value that's not a valid hash to denote empty slots, but this way removal is easier (just flip a bit, instead of overwriting 128 bit).
You don't need a hash function, as your inputs are already integers. You just need to do what every hash tables does anyway: Take the hashes modulo 2^n to derive an index that's not freaking huge. Choose n such that the load factor (the percentage of table entries used) is reasonable (< 2/3 seems standard). Choosing a power of makes the modulo operation cheaper (masking off bits via binary AND), and allows you to just do it on the lower 32 or 64 bit (ignoring the rest).
Choosing a collision resolution strategy is hard. I'd probably go with open addressing with linear probing, as first attempt. It may work badly, but if your input hashes are any good, this seems unlikely. There's also a probing scheme that factors in more and more of the bits you originally cut off, used by CPython's dict.
Now, this is a lot more work and maintenance burden than using off-the-shelf solutions. I wouldn't advise it unless this really is as performance-critical as it sounds in your description.
If C++11 is an option, and your compiler's unordered_set is any good, maybe you should just use it and save yourself most of the hassle (but be aware that this probably increases memory requirements). You still need to specialize std::hash and std::equal_to or operator==. Alternative supply your own Hash and KeyEqual for unordered_set, but that probably doesn't offer any benefit.
Two things might help. First, at least sort the list of items
to be removed; that way, you can use a binary search
(std::lower_bounds) on it. Second, keep two pointers:
a source and a destination. If the source points to something
not in the list to be removed, copy it to the destination, and
advance both. If the source points to something to be removed,
just advance the source pointer, without copying. There should
never be a reason to copy an entry more than once.
Related
I am looking for input on an associative data structure that might take advantage of the specific criteria of my use case.
Currently I am using a red/black tree to implement a dictionary that maps keys to values (in my case integers to addresses).
In my use case, the maximum number of elements is known up front (1024), and I will only ever be inserting and searching. Searching happens twenty times more often than inserting. At the end of the process I clear the structure and repeat again. There can be no allocations during use - only the initial up front one. Unfortunately, the STL and recent versions of C++ are not available.
Any insight?
I ended up implementing a simple linear-probe HashTable from an example here. I used the MurmurHash3 hash function since my data is randomized.
I modified the hash table in the following ways:
The size is a template parameter. Internally, the size is doubled. The implementation requires power of 2 sizes, and traditionally resizes at 75% occupation. Since I know I am going to be filling up the hash table, I pre-emptively double it's capacity to keep it sparse enough. This might be less efficient when adding small number of objects, but it is more efficient once the capacity starts to fill up. Since I cannot resize it I chose to start it doubled in size.
I do not allow keys with a value of zero to be stored. This is okay for my application and it keeps the code simple.
All resizing and deleting is removed, replaced by a single clear operation which performs a memset.
I chose to inline the insert and lookup functions since they are quite small.
It is faster than my red/black tree implementation before. The only change I might make is to revisit the hashing scheme to see if there is something in the source keys that would help make a cheaper hash.
Billy ONeal suggested, given a small number of elements (1024) that a simple linear search in a fixed array would be faster. I followed his advice and implemented one for side by side comparison. On my target hardware (roughly first generation iPhone) the hash table outperformed a linear search by a factor of two to one. At smaller sizes (256 elements) the hash table was still superior. Of course these values are hardware dependant. Cache line sizes and memory access speed are terrible in my environment. However, others looking for a solution to this problem would be smart to follow his advice and try and profile it first.
I have a map<size_t, set<size_t>>, which, for better performance, I'm actually representing as a lexicographically-sorted vector<pair<size_t, vector<size_t>>>.
What I need is a set<T> with fast insertion times (removal doesn't matter), where T is the data type above, so that I can check for duplicates (my program runs until there are no more unique T's being generated.).
So far, switching from set to unordered_set has turned out to be quite beneficial (it makes my program run > 25% faster), but even now, inserting T still seems to be one of the main bottlenecks.
The maximum number of integers in a given T is around ~1000, and each integer is also <= ~1000, so the numbers are quite small (but there are thousands of these T's being generated).
What I have already tried:
Using unsigned short. It actually decreases performance slightly.
Using Google's btree::btree_map.
It's actually much slower because I have to work around the iterator invalidation.
(I have to copy the keys, and I think that's why it turned out slow. It was at least twice as slow.)
Using a different hash function. I haven't found any measurable difference as long as I use something reasonable, so it seems like this can't be improved.
What I have not tried:
Storing "fingerprints"/hashes instead of the actual sets.
This sounds like the perfect solution, except that the fingerprinting function needs to be fast, and I need to be extremely confident that collisions won't happen, or they'll screw up my program.
(It's a deterministic program that needs exact results; collisions render it useless.)
Storing the data in some other compact, CPU-friendly way.
I'm not sure how beneficial this would be, because it might involve copying around data, and most of the performance I've gained so far is by (cleverly) avoiding copying data in many situations.
What else can I do to improve the speed, if anything?
I am under the impression that you have 3 different problems here:
you need the T itself to be relatively compact and easy to move around
you need to quickly check whether a T is a possible duplicate of an already existing one
you finally need to quickly insert the new T in whatever data structure you have to check for duplicates
Regarding T itself, it is not yet as compact as it could be. You could probably use a single std::vector<size_t> to represent it:
N pairs
N Indexes
N "Ids" of I elements each
all that can be linearized:
[N, I(0), ..., I(N-1),
R(0) = Size(Id(0)), Id(0, 0), ... , Id(0, R(0)-1),
R(1) = ... ]
and this way you have a single chunk of memory.
Note: depending on the access pattern you may have to tweak it, specifically if you need random access to any ID.
Regarding the possibility of duplicates, a hash-map seems indeed quite appropriate. You will need a good hash function, but with a single array of size_t (or unsigned short if you can, it is smaller), you can just pick MurmurHash or CityHash or SipHash. They all are blazing fast and do their damnest to produce good quality hash (not cryptographic ones, emphasis is on speed).
Now, the question is when is it slow when checking for duplicates.
If you spend too much time checking for non-existing duplicates because the hash-map is too big, you might want to invest in a Bloom Filter in front of it.
Otherwise, check your hash function to make sure that it is really fast and has a low collision rate and your hash-map implementation to make sure it only ever computes the hash once.
Regarding insertion speed. Normally a hash-map, specifically if well-balanced and pre-sized, should have one of the quickest insertion. Make sure you move data into it and do not copy it; if you cannot move, it might be worth using a shared_ptr to limit the cost of copying.
Don't be afraid of collisions, use cryptographic hash. But choose a fast one. 256 bit collision is MUCH less probable than hardware error. Sun used it to deduplicate blocks in ZFS. ZFS uses SHA256. Probably you can use less secure hash. If it takes $1000000 to find one collision hash isn't secure but one collision don't seem to drop your performance. Many collisions would cost many $1000000. You can use something like (unordered) multimap<SHA, T> to deal with collisions. By the way, ANY hash table suffer from collisions (or takes too many memory), so ordered map (rbtree in gcc) or btree_map has better guaranteed time. Also hash table can be DOSed via hash collisions. Probably a secret salt can solve this problem. It is due to table size is much less than number of possible hashes.
You can also:
1) use short ints
2) reinterpret your array as an array of something like uint64_t for fast comparison (+some trailing elements), or even reinterpret it as an array of 128-bit values (or 256-bit, depending on your CPU) and compare via SSE. This should push your performance to memory speed limit.
From my experience SSE works fast with aligned memory access only. uint64_t comparison probably needs alignment for speed too, so you have to allocate memory manually with proper alignment (allocate more and skip first bytes). tcmalloc is 16 byte-aligned, uint64_t-ready. It is strange that you have to copy keys in btree, you can avoid it using shared_ptr. With fast comparisons and slow hash btree or std::map may turn out to be faster than hash table. I guess any hash is slower than memory. You can also calculate hash via SSE and probably find a library that does it.
PS I strongly recommend you to use profiler if you don't yet. Please tell % of time your program spend to insert, compare in insert and calculate hash.
I am looking for some map which has fixed keys (fixed during initialization) and that does faster look-up. It may not support adding/updating elements later. Is there some algorithm which looks the list of keys and formulates a function so that it is faster to look-up later. In my case, keys are strings.
Update:
Keys are not known at compile time. But during initialization time of the application. There wont be any further insertions later but there will be lots of look-ups. So I want look-ups to be optimized.
CMPH may be what you're looking for. Basically this is gperf without requiring the set at compile-time.
Though of course std::unordered_map as by C++11 might just do too, though possibly with a few collisions.
Since you lookup strings, for strings, a trie (any of the different trie flavours, crit-bit or whatever funky names they have) may also be worthwhile to look into, especially if you have many of them. There are a lot of free trie implementations freely available.
The advantage of tries is that they can index-compress strings, so they use less memory, which has a higher likelihood of having data in cache. Also the access pattern is less random, which is also cache-friendly. A hash table must store the value plus the hash, and indexes more or less randomly (not randomly, but unpredictably) into memory. A trie/trie-like structure ideally only needs one extra bit that distinguishes a key from its common prefix in each node.
(Note by the way that O(log(N)) may quite possibly be faster than O(1) in such a case, because big-O does not consider things like that.)
Note that these are distinct things: do you need an upper limit, do you need a fast typical rate, or do you need the fastest lookup ever, no questions asked? The last one will cost you, the first two ones may be conflicting goals.
You could attempt to create a perfect hash function based on the input (i.e. one that does not have collisions of the input set). This is a somehow-solved problem (e.g. this, this). However, they usually generate source code and may spend significant time generating the hash function.
A modification of this would be using a generic hash function (e.g. shift-multiply-add) and do a brute force search over suitable parameters.
This has to be traded off with the cost of a few string comparisons (which aren't that terribly expensive if you don't have to collate).
Another option is to use two distinct hash functions - this increases the cost of a single lookup but makes degradation slightly less likely than aliens stealing your clock cylces. It is rather unlikely that this would be a problem with typical strings and a decent hash function.
Try google-sparsehash: http://code.google.com/p/google-sparsehash/
An extremely memory-efficient hash_map implementation. 2 bits/entry overhead!
The SparseHash library contains several hash-map implementations, including
implementations that optimize for space or speed.
In a similar topic ((number of) items known at compile time) , I produced this one: Lookups on known set of integer keys. Low overhead, no need for perfect hash. Fortunately, it is in C ;-)
In one of the applications I work on, it is necessary to have a function like this:
bool IsInList(int iTest)
{
//Return if iTest appears in a set of numbers.
}
The number list is known at app load up (But are not always the same between two instances of the same application) and will not change (or added to) throughout the whole of the program. The integers themselves maybe large and have a large range so it is not efficient to have a vector<bool>. Performance is a issue as the function sits in a hot spot. I have heard about Perfect hashing but could not find out any good advice. Any pointers would be helpful. Thanks.
p.s. I'd ideally like if the solution isn't a third party library because I can't use them here. Something simple enough to be understood and manually implemented would be great if it were possible.
I would suggest using Bloom Filters in conjunction with a simple std::map.
Unfortunately the bloom filter is not part of the standard library, so you'll have to implement it yourself. However it turns out to be quite a simple structure!
A Bloom Filter is a data structure that is specialized in the question: Is this element part of the set, but does so with an incredibly tight memory requirement, and quite fast too.
The slight catch is that the answer is... special: Is this element part of the set ?
No
Maybe (with a given probability depending on the properties of the Bloom Filter)
This looks strange until you look at the implementation, and it may require some tuning (there are several properties) to lower the probability but...
What is really interesting for you, is that for all the cases it answers No, you have the guarantee that it isn't part of the set.
As such a Bloom Filter is ideal as a doorman for a Binary Tree or a Hash Map. Carefully tuned it will only let very few false positive pass. For example, gcc uses one.
What comes to my mind is gperf. However, it is based in strings and not in numbers. However, part of the calculation can be tweaked to use numbers as input for the hash generator.
integers, strings, doesn't matter
http://videolectures.net/mit6046jf05_leiserson_lec08/
After the intro, at 49:38, you'll learn how to do this. The Dot Product hash function is demonstrated since it has an elegant proof. Most hash functions are like voodoo black magic. Don't waste time here, find something that is FAST for your datatype and that offers some adjustable SEED for hashing. A good combo there is better than the alternative of growing the hash table.
#54:30 The Prof. draws picture of a standard way of doing perfect hash. The perfect minimal hash is beyond this lecture. (good luck!)
It really all depends on what you mod by.
Keep in mind, the analysis he shows can be further optimized by knowing the hardware you are running on.
The std::map you get very good performance in 99.9% scenarios. If your hot spot has the same iTest(s) multiple times, combine the map result with a temporary hash cache.
Int is one of the datatypes where it is possible to just do:
bool hash[UINT_MAX]; // stackoverflow ;)
And fill it up. If you don't care about negative numbers, then it's twice as easy.
A perfect hash function maps a set of inputs onto the integers with no collisions. Given that your input is a set of integers, the values themselves are a perfect hash function. That really has nothing to do with the problem at hand.
The most obvious and easy to implement solution for testing existence would be a sorted list or balanced binary tree. Then you could decide existence in log(N) time. I doubt it'll get much better than that.
For this problem I would use a binary search, assuming it's possible to keep the list of numbers sorted.
Wikipedia has example implementations that should be simple enough to translate to C++.
It's not necessary or practical to aim for mapping N distinct randomly dispersed integers to N contiguous buckets - i.e. a perfect minimal hash - the important thing is to identify an acceptable ratio. To do this at run-time, you can start by configuring a worst-acceptible ratio (say 1 to 20) and a no-point-being-better-than-this-ratio (say 1 to 4), then randomly vary (e.g. changing prime numbers used) a fast-to-calculate hash algorithm to see how easily you can meet increasingly difficult ratios. For worst-acceptible you don't time out, or you fall back on something slower but reliable (container or displacement lists to resolve collisions). Then, allow a second or ten (configurable) for each X% better until you can't succeed at that ratio or reach the no-pint-being-better ratio....
Just so everyone's clear, this works for inputs only known at run time with no useful patterns known beforehand, which is why different hash functions have to be trialed or actively derived at run time. It is not acceptible to simple say "integer inputs form a hash", because there are collisions when %-ed into any sane array size. But, you don't need to aim for a perfectly packed array either. Remember too that you can have a sparse array of pointers to a packed array, so there's little memory wasted for large objects.
Original Question
After working with it for a while, I came up with a number of hash functions that seemed to work reasonably well on strings, resulting in a unique - perfect hashing.
Let's say the values ranged from L to H in the array. This yields a Range R = H - L + 1.
Generally it was pretty big.
I then applied the modulus operator from H down to L + 1, looking for a mapping that keeps them unique, but has a smaller range.
In you case you are using integers. Technically, they are already hashed, but the range is large.
It may be that you can get what you want, simply by applying the modulus operator.
It may be that you need to put a hash function in front of it first.
It also may be that you can't find a perfect hash for it, in which case your container class should have a fall back position.... binary search, or map or something like that, so that
you can guarantee that the container will work in all cases.
A trie or perhaps a van Emde Boas tree might be a better bet for creating a space efficient set of integers with lookup time bring constant against the number of objects in the data structure, assuming that even std::bitset would be too large.
I need a fast container with only two operations. Inserting keys on from a very sparse domain (all 32bit integers, and approx. 100 are set at a given time), and iterating over the inserted keys. It should deal with a lot of insertions which hit the same entries (like, 500k, but only 100 different ones).
Currently, I'm using a std::set (only insert and the iterating interface), which is decent, but still not fast enough. std::unordered_set was twice as slow, same for the Google Hash Maps. I wonder what data structure is optimized for this case?
Depending on the distribution of the input, you might be able to get some improvement without changing the structure.
If you tend to get a lot of runs of a single value, then you can probably speed up insertions by keeping a record of the last value you inserted, and don't bother doing the insertion if it matches. It costs an extra comparison per input, but saves a lookup for each element in a run beyond the first. So it could improve things no matter what data structure you're using, depending on the frequency of repeats and the relative cost of comparison vs insertion.
If you don't get runs, but you tend to find that values aren't evenly distributed, then a splay tree makes accessing the most commonly-used elements cheaper. It works by creating a deliberately-unbalanced tree with the frequent elements near the top, like a Huffman code.
I'm not sure I understand "a lot of insertions which hit the same entries". Do you mean that there are only 100 values which are ever members, but 500k mostly-duplicate operations which insert one of those 100 values?
If so, then I'd guess that the fastest container would be to generate a collision-free hash over those 100 values, then maintain an array (or vector) of flags (int or bit, according to what works out fastest on your architecture).
I leave generating the hash as an exercise for the reader, since it's something that I'm aware exists as a technique, but I've never looked into it myself. The point is to get a fast hash over as small a range as possible, such that for each n, m in your 100 values, hash(n) != hash(m).
So insertion looks like array[hash(value)] = 1;, deletion looks like array[hash(value)] = 0; (although you don't need that), and to enumerate you run over the array, and for each set value at index n, inverse_hash(n) is in your collection. For a small range you can easily maintain a lookup table to perform the inverse hash, or instead of scanning the whole array looking for set flags, you can run over the 100 potentially-in values checking each in turn.
Sorry if I've misunderstood the situation and this is useless to you. And to be honest, it's not very much faster than a regular hashtable, since realistically for 100 values you can easily size the table such that there will be few or no collisions, without using so much memory as to blow your caches.
For an in-use set expected to be this small, a non-bucketed hash table might be OK. If you can live with an occasional expansion operation, grow it in powers of 2 if it gets more than 70% full. Cuckoo hashing has been discussed on Stackoverflow before and might also be a good approach for a set this small. If you really need to optimise for speed, you can implement the hashing function and lookup in assembler - on linear data structures this will be very simple so the coding and maintenance effort for an assembler implementation shouldn't be unduly hard to maintain.
You might want to consider implementing a HashTree using a base 10 hash function at each level instead of a binary hash function. You could either make it non-bucketed, in which case your performance would be deterministic (log10) or adjust your bucket size based on your expected distribution so that you only have a couple of keys/bucket.
A randomized data structure might be perfect for your job. Take a look at the skip list – though I don't know any decend C++ implementation of it. I intended to submit one to Boost but never got around to do it.
Maybe a set with a b-tree (instead of binary tree) as internal data structure. I found this article on codeproject which implements this.
Note that while inserting into a hash table is fast, iterating over it isn't particularly fast, since you need to iterate over the entire array.
Which operation is slow for you? Do you do more insertions or more iteration?
How much memory do you have? 32-bits take "only" 4GB/8 bytes, which comes to 512MB, not much for a high-end server. That would make your insertions O(1). But that could make the iteration slow. Although skipping all words with only zeroes would optimize away most iterations. If your 100 numbers are in a relatively small range, you can optimize even further by keeping the minimum and maximum around.
I know this is just brute force, but sometimes brute force is good enough.
Since no one has explicitly mentioned it, have you thought about memory locality? A really great data structure with an algorithm for insertion that causes a page fault will do you no good. In fact a data structure with an insert that merely causes a cache miss would likely be really bad for perf.
Have you made sure a naive unordered set of elements packed in a fixed array with a simple swap to front when an insert collisides is too slow? Its a simple experiment that might show you have memory locality issues rather than algorithmic issues.