Zero load time file formats
Posted 13 years ago
Sometimes you cannot afford to load data files from disk. Maybe you need
results immediately, or the data is simply too large to fit into memory. A
technique that I like to use is an on-disk data structure. Here is a toy
example for instantly accessing lists of related words.
In this article, I address the problem of the time needed to load data into memory from disk. However, I do not make any optimization for disk caches or blocks. I am not going to talk about B-Trees or cache-oblivious structures.
No waiting
Using an on-disk data structure, there is no need to load the whole file into memory or parse it. Instead of opening a file and reading its contents, we will use a memory mapped file. We tell the operating system the file name, and it will lazily load the parts of the file only when we access them. These parts remain in the disk cache even after our program exits. So if you later start the program again, it will execute similar queries more quickly. We let the operating system do the caching for us. In python, this is done using the mmap module. Mmap makes the file appear as a very long string.Toy example
Here is an on-disk structure for looking up related words that I prepared. (Download 11 MB of it). It has three sections: A header, an index, and a word section. The header contains the number of words. The index contains a list of pointers to word records. The word section contains the word records. It is constructed so that we can instantly query for words related to a given word by jumping around to different parts of the file.--- header 4 bytes: number of words --- index section. The words are listed in alphabetical order, so you can --- look one up using binary search. for each word: 4 byte ptr to word record --- word section: for each word: null terminated text 4 bytes: number of related words for each link, ptr to linked word recordHere is a short python program for accessing the data file.
#!/usr/bin/python # An on-disk data structure for finding related words # By Steve Hanov. This code and data file are released to the public domain. import sys, mmap, struct class FrozenThesaurus: def __init__(self, filename): self.f = file(filename, "rb") self.mmap = mmap.mmap( self.f.fileno(), 0, access=mmap.ACCESS_READ ) def getDword( self, ptr ): # return the 32 bit number beginning at the given byte offset in the # file. return struct.unpack("<I", self.mmap[ptr:ptr+4])[0] def getString( self, ptr ): # return the null terminated string beginning at the given byte offset. result = [] while self.mmap[ptr] != "\x00": result.append(self.mmap[ptr]) ptr += 1 return "".join(result) def getWordCount(self): # Retrive the number of words in the file. return self.getDword(0) def getWord(self, index): # Retrive a word, given its index. The index must be less then the word # count. return self.getString( self.getDword(4 + index * 4) ) def getIndexOf( self, word ): # perform a binary search through the index for the given word. high = self.getWordCount() low = -1 while (high - low > 1): probe = (high + low) / 2 candidate = self.getWord(probe) if candidate == word: return probe elif candidate < word: low = probe else: high = probe return None def getRelatedWords( self, word ): # Returns the list of related words to the given word. results = [] index = self.getIndexOf( word ) if index == None: return results ptr = self.getDword( 4 + index * 4 ) # skip past the word text while self.mmap[ptr] != '\x00': ptr += 1 ptr += 1 numRelated = self.getDword( ptr ) for i in range(numRelated): ptr += 4 results.append( self.getString( self.getDword( ptr ) ) ) return results; data = FrozenThesaurus("thesaurus.dat") print data.getRelatedWords(sys.argv[1])
When shouldn't you use this?
SQLITE uses memory mapped files internally. If you create the proper indicies, SQLITE will match the performance of any file format that you can come up with yourself, though it may be larger. If you can store your data in a relational database, you should not go through the trouble of creating your own on-disk data structure. In particular, a thesaurus could easily be stored in an SQLITE database.
Steve Hanov makes a living working on
Rhymebrain.com,
rapt.ink,
www.websequencediagrams.com,
and Zwibbler.com. He lives in
Waterloo, Canada.
Post comment
edit
Gregory
twelve years ago
Yeah well... that totally doesn't address endianness and struct member alignment
edit
Joshua Schachter
twelve years ago
This didn't actually work on python 2.7.1 on OS X. You have to change '' to 'x00' for it to match the nulls.
Joshua
edit
John
13 years ago
Hi Steve,
I translated your example to Factor, if you're curious. It's on my Re: Factor blog at re-factor.blogspot.com.
Thanks,
John.
edit
rjp
13 years ago
I did something similar to store a huge hash of URLs for a URL filtering proxy. Written in Ruby but was an order of magnitude faster than the linear-search C version.