Recently, my twitter pal @ifesdjeen wrote a line that resonated with me: "Looks like it's easier for people to read 40 blog posts than a single whitepaper." And although he used it in a negative context, I recognized it as a very precise (and, actually, positive) description of what a research engineer does: read a whitepaper (or a dozen, for what it's worth) and transform it into working code and - as a possible byproduct - into a blog post that other engineers will understand and be able to reproduce. I'm in the business of reproducing papers for about 7 years now, and, I believe, everyone with the same experience will confirm that it's not a skill every engineer should be easily capable of. Not all papers even can be reproduced (because the experiment was just not set up correctly), and of those, which, in principle, can be, only some are presented in the form that I can grasp well enough to be able to program them. And, after all, working code is an ultimate judge of your understanding.
But I digressed... :) This post is meant to explain (in simple "engineering" terms) the concept of minimal perfect hash-tables and how I've recently implemented one of their varieties to improve the memory efficiency of my language identification library WILD. Uncovering, in the process, a more abstract concept behind it.
The Justification of Perfect Hash-Tables
Minimal perfect hash-tables are persistent data structures that solve 2 basic deficiencies of normal hash-tables: they guarantee both constant (not only amortized) O(1) time for collision-free key-based access, while no reserved space is required (which for hash-tables may be in the range of 20%-50% of its size). It comes at a cost of two restrictions: the table should be filled with a keyset known ahead-of-time, and the process of building one takes longer than for a usual hash-table (although, for many methods, it can still be bounded by amortized O(1) time).
From my point of view, the main advantage of perfect HTs is the possibility to substantially reduce memory usage, which is important in such use cases as storing big dictionaries (relevant to many NLP tasks). Moreover, the space requirements can be lowered even more if the whole keyset is stored in the table, i.e. there can be no misses. Under such constraints, unlike with normal hash-tables, which still require storing the keys alongside the values due to the need to compare them in cases of hash collisions, in a perfect HT they can be just omitted. Unfortunately, such situation is rarely the case, but if some level of false positives is permitted, with the help of additional simple programmatic tricks, the memory usage for keys can be also reduced by orders of magnitude.
Hash-tables on their own are the trickiest among the basic data structures. But, in terms of complexity, they pale in comparison to minimal perfect hash-tables, which belong to the advanced algorithms family. One reason for that is that perfect hash-tables require more "moving parts", but the main one is, probably, that there's no common well-known algorithm to distribute keys in a perfect manner. And reading many perfect hash-table papers will scare away most programmers, at least it did me. However, after some research and trial-and-error, I've managed to find the one that presents a simple and scalable approach. So, I'm going to explain it in this post.
Hash-tables in SBCL
If you take a look at some particular hash-table implementation, your mental model of what a hash-table is may be quite seriously shaken. A straightforward open addressing table will assume a vector of key-value pairs as an underlying concrete structure, filled as the table is modified. On a 64-bit system, it will require: (8 × 1.5 + 8 × 16) × entries-count + total size of all keys + total size of all values + constant size of metadata. The 8 × 1.5 bytes in the formula is storage needed to hold a pointer to a hash-table entry including an average 33% extra overhead, the additional 8 × 16 bytes per entry will be spent on a cons cell (if we use this efficient although antiquated way to represent pairs in Lisp). It should be noted that depending on the size of keys and values the first term may be either a significant (up to half of the total size) or a negligible part of the table's memory profile.
However, this is not how SBCL implements hash-tables. See for yourself. Let's create a random hash-table with approximately 1000 keys:
> (defparameter *ht*
(pairs->ht (loop :repeat 1000
:collect (pair (fmt "~{~C~}"
(loop :repeat (random 10)
:collect (code-char (+ 65 (random 50)))))
(random 1000)))
:test 'equal))
And inspect its contents:
> (inspect *ht*)
The object is a STRUCTURE-OBJECT of type HASH-TABLE.
0. TEST: EQUAL
1. TEST-FUN: #
2. HASH-FUN: #
3. REHASH-SIZE: 1.5
4. REHASH-THRESHOLD: 1.0
5. REHASH-TRIGGER: 1024
6. NUMBER-ENTRIES: 844
7. TABLE: #(#{EQUAL
"plCjU" 985
"VVYZqKm[" 861
"N\\" 870
"fqfBdZP\\" 364
"cHNjIM]Y" 804
"p_oHUWc^" 630
"CqGRaMH" 408
"NecO" 636
"QDBq" 380
"M" 838
...
}
0 "plCjU" 985 "VVYZqKm[" 861 "N\\" 870 "fqfBdZP\\" 364 ...)
8. NEXT-WEAK-HASH-TABLE: NIL
9. %WEAKNESS: 0
10. NEXT-FREE-KV: 845
11. CACHE: 1688
12. INDEX-VECTOR: #(0 617 332 393 35 0 194 512 0 420 ...)
13. NEXT-VECTOR: #(0 72 0 0 253 499 211 0 349 488 ...)
14. HASH-VECTOR: #(9223372036854775808 1830284672361826498 3086478891113066655
24243962159539 2602570438820086662 2431530612434713043
4568274338569089845 3085527599069570347 2374133389538826432
3322613783177911862 ...)
15. LOCK: #
16. SYNCHRONIZED-P: NIL
As the fog of initial surprise clears, we can see that instead of a single vector it uses 4! The keys and values are placed in the one called TABLE (that starts with a reference to the hash-table object itself). Note that storing the entries as pairs is redundant both in terms of memory and access (additional dereferencing). That's why an obvious optimization is to put them directly in the backing array: so the keys and values here are interleaved. One more indirection that may be removed, at least in Lisp, is "unboxing" the numbers, i.e. storing them immediately in the array avoiding the extra pointer. This may be an additional specialized optimization for number-keyed hash-tables (to which we'll return later), but it is hardly possible with the interleaved scheme.
Perhaps, the biggest surprise is that the entries of the TABLE array are stored sequentially, i.e. there are no gaps. If they were to be added randomly, we'd expect the uniform distribution of 845×2 unique keys and corresponding values over the 2048-element array. Instead, this randomness is transferred to the 1024-element integer INDEX-VECTOR: another level of indirection. Why? For the same reason yet another 1024-element array is used — a HASH-VECTOR, which stores the values of hashes for all the table keys: efficient resizing. Although, it may seem that the biggest overhead of a hash-table is incurred due to reserved space for new keys, in fact, as we have now learned, resizing is a much heavier factor. Especially this relates to the HASH-VECTOR: if the hashes of all keys would not have been stored, each time the table got resized they would have had to be recalculated anew: an intolerable slowdown for larger tables. Having a separate INDEX-VECTOR is not so critical, but it provides two additional nice capabilities. Resizing the table is performed by adjusting the vectors' lengths (an optimized operation) without the need to redistribute the entries. Besides, unlike in theory, we can iterate this table in a deterministic order: the one of keys addition into the table. Which comes quite handy, although SBCL developers don't want to advertise this as a public API as it may restrict potential future modifications to the data structure. There's also a NEXT-VECTOR that is used to optimize insertion.
Overall, we can see that a production-optimized hash-table turns out to be much heavier than a textbook variant. And, indeed, these optimizations are relevant, as SBCL's hash-table are quite efficient. In my experiments several years ago, they turned out to be 2-3 times faster than, for instance, the CCL ones. Yet, it's clear that the speed optimizations, as usual, come at a cost of storage deoptimization. To restore storage sanity, we could fall back to the basic hash-table variant, and it's a possible solution for some cases, although a mediocre one, in general. Neither will it be the fastest, nor fully optimized.
Building an MPHT
Most of space in the SBCL's hash-table implementation is spent on metadata and keys, not on values. Yet, if we introduce a restriction that the table cannot be altered — no new values added after initial construction and no resizing possible — most of those optimizations and connected storage become redundant. An ideally space-efficient data structure is an array, but it doesn't allow key-based access. In theory, minimal perfect hash-tables are arrays with a minimal amount of metadata overhead. Yet, the precise amount is dependent on the algorithm, and there's still new research improving them going on. Overviewing all the existing approaches (most of which I don't fully understand) is beyond the scope of this post.
Besides, MPHTs require additional calculations at access time compared to normal hash-tables. So, if a hash-table is well-optimized it will usually be faster than and MPH.
And still, MPHs require a non-negligible amount of additional space for metadata: for the algorithm that will be discussed it's around 2 bytes per key. The best algorithms in the literature claim to reduce this amount more than 4 times to less than 4 bits per key. However, for now, I have picked the simpler approach, since this difference is not critical considering that every value in the table, on a 64-bit machine, occupies at least 16 bytes (a pointer plus the data), so an overhead of 2 bytes versus 0.5 bytes (that will probably be rounded to 1 byte) is already negligible.
Now, let's think of how we can distribute the keys in a hash-table so that there are no collisions and the backing array has the same number of elements as the number of hash-table keys. Theoretically, as the set of keys is known before the creation of the table, we can find a hash function that produces such distribution. Unfortunately, due to the Birthday paradox, it may be a long search. The algorithms for MPHTs suggest ways of structuring it. A good algorithm should have at most O(n) complexity as, otherwise, it will be infeasible for large keysets, which are the main use case for perfect hash-tables.
The algorithm I will briefly describe now was suggested by Botelho and Ziviani. It is a 2-step process:
- at first stage, using a normal hash-function (in particular, Jenkins hash), all keys are nearly uniformly distributed into buckets, so that the number of keys in each bucket doesn't exceed 256. This can be done by setting the hash divisor to
(ceiling (length keyset) 200 #| or slightly less |#); - next, for each bucket, a perfect hash function is constructed via a simple algorithm: for each key, two hash codes are calculated by one call to the Jenkins hash (this function outputs 3 hashes at once), which are treated as vertices of a graph. If the graph happens to be non-circular (which can be ensured with high probability with a suitable hash divisor), it is possible to construct the desired function as a sum of 2 hashes. Otherwise, we change the Jenkins hash seed and try constructing a new graph until a non-circular one is obtained. In practice, this requires just a couple of tries;
- the construction of the final hash function is described very clearly by Czech, Havas and Majewski who have proposed this method: it lies in performing a depth-first search on the graph, labelling the edges with unique numbers and deducing the corresponding number for each possible Jenkins hash value.
Here you can see one of the possible labelings (each edge corresponds to a key and its unique index, each vertex — to the value for each of the possible Jenkins hashes):
Now, the per-bucket hash-function can be reconstructed from an array of numbers (in the range 0-255) associated with each possible hash. The divisor of the hash is twice the number of keys in the bucket, though it can be any number above the number of keys: the greater the divisor, the more space is used for storing the numbers (the divisor is the length of an array) and the less time it takes to find an acyclic graph.
The algorithm works quite fast: on my laptop, it takes 8 seconds for the table of 725 359 character trigrams from a mid-size language identification model and 13 seconds for 1 236 452 words from the same model.
To sum up, this is how to find the index of an element (bytes argument) in our perfect hash-table:
(defun csindex (bytes cstab)
(with ((mod (/ (1- (length @cstab.meta)) 2)) ; divisor of the top-level hash
(hash (* (mod (jenkins-hash bytes) mod) 2)) ; determine the bucket
(off (? @cstab.meta hash))
(seed (? @cstab.meta (1+ hash))) ; each bucket has its own Jenkins seed
(mod2 (- (? @cstab.meta (+ 2 hash)) off))
(b c (jenkins-hash2 bytes seed (* mod2 2)))
(goff (* 2 off)))
;; bucket offset + in-bucket hash
(+ off (mod (+ (? gs (+ goff b))
(? gs (+ goff c)))
mod2))))
Note, that, in the code for this article, I refer to perfect hash tables as cstabs for the reasons described in the end.
Efficiency In Practice
So, let's now examine the memory efficiency of this method. Thankfully, just recently the SBCL developers started working on a critical for every algorithmic developer missing piece in the SBCL API: a function to determine the size in memory occupied by an arbitrary object. As we know from the famous koan, "LISP programmers know the value of everything and the cost of nothing". Indeed, from a certain point of view, this applies to SBCL. Although we, now, have a rough tool at our disposal that patches this hole... ;)
Using this unofficial function, we can roughly calculate the space occupied by the character trigrams hash-table mentioned above:
> (let ((ht (? wild:*lang-detector* '3gs)))
(+ (object-size ht)
(object-size @ht.table)
(object-size @ht.index-vector)
(object-size @ht.hash-vector)
(reduce '+ (map 'list
(lambda (obj)
;; the keys of the table are strings
;; and values -- alists of a symbol and a float
(reduce '+ (map 'list ^(if (listp %)
(sum 'object-size %)
(object-size %))
@ht.table)))))))
102856432
100 MB!
> (let ((ct (build-const-table (? wild:*lang-detector* '3gs))))
(+ (object-size ct)
(object-size @ct.gs)
(object-size @ct.meta)
(reduce '+ (map 'list ^(sum 'object-size %)
@ct.data))))
53372880
I.e., we have reduced the memory usage almost twice. Because all the metadata now occupies just 1479808 bytes or 1,5 MB.
One critical decision that allows for such drastic memory-use improvement is omitting the keys from the table. It should be noted that adding them back is trivial: (defstruct (keyed-pht (:include const-table)) (keys nil :type simple-vector) (test 'eql)), for which getting the value will work like this:
(defun csget (key cstab)
(let ((idx (csindex (bytes key) cstab)))
(when (call @cstab.test key (svref @cstab.keys idx))
(values (svref @cstab.data idx)
t)))
However, this will, basically, return us to the same ballpark as a textbook hash-table implementation in terms of memory usage while loosing in terms of speed. Yet, if we allow for some controlled level of false positives, there are a few algorithmic tricks that can be used to once again make keys almost negligible.
The first one is really simple and straightforward: replace the vector of keys with the vector of their hashes. In particular, if we take a single byte of the hash, such array will be 10-100 times smaller than the generic keys array and produce an FP-rate of 1/255.
Another trick will be to use a Bloom filter. For instance, a filter with 0.1 FP-rate for all the trigrams from our language identification model will require just around 0.4 MB of storage compared to 0.7 MB needed to store the 1-byte hashes and 30 MB needed to store just the keys. The disadvantage of a Bloom filter, however, is slower processing time: for the mentioned 10% FP-rate we'll need to perform 4 hash calculations, and if we'd like to reach the same 0.3% rate as the 1-byte hash array we'd have to increase the memory usage to 1MB and perform 8 hash calculations.
Finally, it should be noted that the main motivator for this work was reducing the memory footprint of my language identification library, for, to be widely adopted, such kind of project should be very frugal in its memory usage: 10-30 MB is its maximum polite footprint. By switching to a perfect hash-table, we haven't reached that goal (at least, for this particular model), but there's also plenty of room for optimizing values memory usage that I will return to later.
Const-tables
The other motivator for this work, in general, was my interest in the topic of efficient "static" data structures. In this context, I feel that the notion of a perfect hash table doesn't fully capture the essential features of the data structures class we're dealing with. First of all, the main distinguishing factor is static vs dynamic usage. A hash-table is thought of as a dynamic entity while our data structure is primarily a static one. Next, hashing is, for sure, involved in constructing all these structures, but it's only part of a bigger picture. The way of structuring the metadata, in particular, the index arrays may differ greatly not only between perfect and usual hash-table, but also within the ordinary hash-table class — within the different implementations.
So, I decided to come up with a different name for this data structure — a const-table (or cstab). It defines a broad class of persistent data structures that allow constant-time key-based access and, ideally, efficiently store the keys. The implementation, described here, is released as the library with the same name. It is still in its infancy, so major changes are possible — and suggestions on its improvement are also welcome.
No comments:
Post a Comment