Question? Leave a message!




Index Construction

Index Construction
Introduction to Information Retrieval Introduction to Information Retrieval Index Construction www.ThesisScientist.comIntroduction to Information Retrieval Plan  Last lecture: nz ahu  Dictionary data structures hym  Tolerant retrieval  Wildcards  Spell correction m mace madden  Soundex mo among amortize on abandon among  This time:  Index construction www.ThesisScientist.comIntroduction to Information Retrieval Ch. 4 Index construction  How do we construct an index  What strategies can we use with limited main memory www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.1 Hardware basics  Many design decisions in information retrieval are based on the characteristics of hardware  We begin by reviewing hardware basics www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.1 Hardware basics  Access to data in memory is much faster than access to data on disk.  Disk seeks: No data is transferred from disk while the disk head is being positioned.  Therefore: Transferring one large chunk of data from disk to memory is faster than transferring many small chunks.  Disk I/O is blockbased: Reading and writing of entire blocks (as opposed to smaller chunks).  Block sizes: 8KB to 256 KB. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.1 Hardware basics  Servers used in IR systems now typically have several GB of main memory, sometimes tens of GB.  Available disk space is several (2–3) orders of magnitude larger.  Fault tolerance is very expensive: It’s much cheaper to use many regular machines rather than one fault tolerant machine. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.1 Hardware assumptions for this lecture  symbol statistic value −3  s average seek time 5 ms = 5 x 10 s −8  b transfer time per byte 0.02 μs = 2 x 10 s 9 −1  processor’s clock rate 10 s −8  p lowlevel operation 0.01 μs = 10 s (e.g., compare swap a word)  size of main memory several GB  size of disk space 1 TB or more www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 RCV1: Our collection for this lecture  Shakespeare’s collected works definitely aren’t large enough for demonstrating many of the points in this course.  The collection we’ll use isn’t really large enough either, but it’s publicly available and is at least a more plausible example.  As an example for applying scalable index construction algorithms, we will use the Reuters RCV1 collection.  This is one year of Reuters newswire (part of 1995 and 1996) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 A Reuters RCV1 document www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Reuters RCV1 statistics  symbol statistic value  N documents 800,000  L avg. tokens per doc 200  M terms (= word types) 400,000  avg. bytes per token 6 (incl. spaces/punct.)  avg. bytes per token 4.5 (without spaces/punct.)  avg. bytes per term 7.5  nonpositional postings 100,000,000 4.5 bytes per word token vs. 7.5 bytes per word type: why www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Term Doc I 1 Recall IIR 1 index construction did 1 enact 1 julius 1 caesar 1  Documents are parsed to extract words and these I 1 are saved with the Document ID. was 1 killed 1 i' 1 the 1 capitol 1 brutus 1 killed 1 me 1 Doc 1 Doc 2 so 2 let 2 it 2 be 2 I did enact Julius with 2 So let it be with caesar 2 Caesar I was killed Caesar. The noble the 2 noble 2 i' the Capitol; Brutus hath told you brutus 2 hath 2 Brutus killed me. Caesar was ambitious told 2 you 2 caesar 2 was 2 www.ThesisScientist.com ambitious 2Introduction to Information Retrieval Sec. 4.2 Term Doc Term Doc Key step I 1 ambitious 2 did 1 be 2 enact 1 brutus 1 brutus 2 julius 1 caesar 1 capitol 1  After all documents have been I 1 caesar 1 caesar 2 was 1 parsed, the inverted file is killed 1 caesar 2 i' 1 did 1 sorted by terms. the 1 enact 1 capitol 1 hath 1 I 1 brutus 1 killed 1 I 1 me 1 i' 1 We focus on this sort step. it 2 so 2 let 2 julius 1 We have 100M items to sort. it 2 killed 1 killed 1 be 2 with 2 let 2 me 1 caesar 2 noble 2 the 2 noble 2 so 2 the 1 brutus 2 hath 2 the 2 told 2 told 2 you 2 you 2 caesar 2 was 1 was 2 was 2 with 2 ambitious 2 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Scaling index construction  Inmemory index construction does not scale  Can’t stuff entire collection into memory, sort, then write back  How can we construct an index for very large collections  Taking into account the hardware constraints we just learned about . . .  Memory, disk, speed, etc. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Sortbased index construction  As we build the index, we parse docs one at a time.  While building the index, we cannot easily exploit compression tricks (you can, but much more complex)  The final postings for any term are incomplete until the end.  At 12 bytes per nonpositional postings entry (term, doc, freq), demands a lot of space for large collections.  T = 100,000,000 in the case of RCV1  So … we can do this in memory in 2009, but typical collections are much larger. E.g., the New York Times provides an index of 150 years of newswire  Thus: We need to store intermediate results on disk. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Sort using disk as “memory”  Can we use the same index construction algorithm for larger collections, but by using disk instead of memory  No: Sorting T = 100,000,000 records on disk is too slow – too many disk seeks.  We need an external sorting algorithm. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Bottleneck  Parse and build postings entries one doc at a time  Now sort postings entries by term (then by doc within each term)  Doing this with random disk seeks would be too slow – must sort T=100M records If every comparison took 2 disk seeks, and N items could be sorted with N log N comparisons, how long would this take 2 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 BSBI: Blocked sortbased Indexing (Sorting with fewer disk seeks)  12byte (4+4+4) records (term, doc, freq).  These are generated as we parse docs.  Must now sort 100M such 12byte records by term.  Define a Block 10M such records  Can easily fit a couple into memory.  Will have 10 such blocks to start with.  Basic idea of algorithm:  Accumulate postings for each block, sort, write to disk.  Then merge the blocks into one long sorted order. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 Sorting 10 blocks of 10M records  First, read each block and sort within:  Quicksort takes 2N ln N expected steps  In our case 2 x (10M ln 10M) steps  Exercise: estimate total time to read each block from disk and and quicksort it.  10 times this estimate – gives us 10 sorted runs of 10M records each.  Done straightforwardly, need 2 copies of data on disk  But can optimize this www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 How to merge the sorted runs  Can do binary merges, with a merge tree of log 10 = 4 layers. 2  During each layer, read into memory runs in blocks of 10M, merge, write back. 1 2 1 Merged run. 2 3 4 3 4 Runs being merged. Disk www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.2 How to merge the sorted runs  But it is more efficient to do a multiway merge, where you are reading from all blocks simultaneously  Providing you read decentsized chunks of each block into memory and then write out a decentsized output chunk, then you’re not killed by disk seeks www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.3 Remaining problem with sortbased algorithm  Our assumption was: we can keep the dictionary in memory.  We need the dictionary (which grows dynamically) in order to implement a term to termID mapping.  Actually, we could work with term,docID postings instead of termID,docID postings . . .  . . . but then intermediate files become very large. (We would end up with a scalable, but very slow index construction method.) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.3 SPIMI: Singlepass inmemory indexing  Key idea 1: Generate separate dictionaries for each block – no need to maintain termtermID mapping across blocks.  Key idea 2: Don’t sort. Accumulate postings in postings lists as they occur.  With these two ideas we can generate a complete inverted index for each block.  These separate indexes can then be merged into one big index. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.3 SPIMIInvert  Merging of blocks is analogous to BSBI. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.3 SPIMI: Compression  Compression makes SPIMI even more efficient.  Compression of terms  Compression of postings  See next lecture www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Distributed indexing  For webscale indexing (don’t try this at home): must use a distributed computing cluster  Individual machines are faultprone  Can unpredictably slow down or fail  How do we exploit such a pool of machines www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Web search engine data centers  Web search data centers (Google, Bing, Baidu) mainly contain commodity machines.  Data centers are distributed around the world.  Estimate: Google 1 million servers, 3 million processors/cores (Gartner 2007) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Massive data centers  If in a nonfaulttolerant system with 1000 nodes, each node has 99.9 uptime, what is the uptime of the system  Answer: 63  Exercise: Calculate the number of servers failing per minute for an installation of 1 million servers. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Distributed indexing  Maintain a master machine directing the indexing job – considered “safe”.  Break up indexing into sets of (parallel) tasks.  Master machine assigns each task to an idle machine from a pool. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Parallel tasks  We will use two sets of parallel tasks  Parsers  Inverters  Break the input document collection into splits  Each split is a subset of documents (corresponding to blocks in BSBI/SPIMI) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Parsers  Master assigns a split to an idle parser machine  Parser reads a document at a time and emits (term, doc) pairs  Parser writes pairs into j partitions  Each partition is for a range of terms’ first letters  (e.g., af, gp, qz) – here j = 3.  Now to complete the index inversion www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Inverters  An inverter collects all (term,doc) pairs (= postings) for one termpartition.  Sorts and writes to postings lists www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Data flow Master assign assign Postings af gp qz Parser Inverter af Parser af gp qz Inverter gp splits Inverter qz Parser af gp qz Map Reduce Segment files phase phase www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 MapReduce  The index construction algorithm we just described is an instance of MapReduce.  MapReduce (Dean and Ghemawat 2004) is a robust and conceptually simple framework for distributed computing …  … without having to write code for the distribution part.  They describe the Google indexing system (ca. 2002) as consisting of a number of phases, each implemented in MapReduce. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 MapReduce  Index construction was just one phase.  Another phase: transforming a termpartitioned index into a documentpartitioned index.  Termpartitioned: one machine handles a subrange of terms  Documentpartitioned: one machine handles a subrange of documents  As we’ll discuss in the web part of the course, most search engines use a documentpartitioned index … better load balancing, etc. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.4 Schema for index construction in MapReduce  Schema of map and reduce functions  map: input → list(k, v) reduce: (k,list(v)) → output  Instantiation of the schema for index construction  map: collection → list(termID, docID)  reduce: (termID1, list(docID), termID2, list(docID), …) → (postings list1, postings list2, …) www.ThesisScientist.comIntroduction to Information Retrieval Example for index construction  Map:  d1 : C came, C c’ed.  d2 : C died. →  C,d1, came,d1, C,d1, c’ed, d1, C, d2, died,d2  Reduce:  (C,(d1,d2,d1), died,(d2), came,(d1), c’ed,(d1)) → (C,(d1:2,d2:1), died,(d2:1), came,(d1:1), c’ed,(d1:1)) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Dynamic indexing  Up to now, we have assumed that collections are static.  They rarely are:  Documents come in over time and need to be inserted.  Documents are deleted and modified.  This means that the dictionary and postings lists have to be modified:  Postings updates for terms already in dictionary  New terms added to dictionary www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Simplest approach  Maintain “big” main index  New docs go into “small” auxiliary index  Search across both, merge results  Deletions  Invalidation bitvector for deleted docs  Filter docs output on a search result by this invalidation bitvector  Periodically, reindex into one main index www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Issues with main and auxiliary indexes  Problem of frequent merges – you touch stuff a lot  Poor performance during merge  Actually:  Merging of the auxiliary index into the main index is efficient if we keep a separate file for each postings list.  Merge is the same as a simple append.  But then we would need a lot of files – inefficient for OS.  Assumption for the rest of the lecture: The index is one big file.  In reality: Use a scheme somewhere in between (e.g., split very large postings lists, collect postings lists of length 1 in one file etc.) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Logarithmic merge  Maintain a series of indexes, each twice as large as the previous one  At any time, some of these powers of 2 are instantiated  Keep smallest (Z ) in memory 0  Larger ones (I , I , …) on disk 0 1  If Z gets too big ( n), write to disk as I 0 0  or merge with I (if I already exists) as Z 0 0 1  Either write merge Z to disk as I (if no I ) 1 1 1  Or merge with I to form Z 1 2 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Logarithmic merge  Auxiliary and main index: index construction time is 2 O(T ) as each posting is touched in each merge.  Logarithmic merge: Each posting is merged O(log T) times, so complexity is O(T log T)  So logarithmic merge is much more efficient for index construction  But query processing now requires the merging of O(log T) indexes  Whereas it is O(1) if you just have a main and auxiliary index www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Further issues with multiple indexes  Collectionwide statistics are hard to maintain  E.g., when we spoke of spellcorrection: which of several corrected alternatives do we present to the user  We said, pick the one with the most hits  How do we maintain the top ones with multiple indexes and invalidation bit vectors  One possibility: ignore everything but the main index for such ordering  Will see more such statistics used in results ranking www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Dynamic indexing at search engines  All the large search engines now do dynamic indexing  Their indices have frequent incremental changes  News items, blogs, new topical web pages  Sarah Palin, …  But (sometimes/typically) they also periodically reconstruct the index from scratch  Query processing is then switched to the new index, and the old index is deleted www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 4.5 Other sorts of indexes  Positional indexes  Same sort of sorting problem … just larger Why  Building character ngram indexes:  As text is parsed, enumerate ngrams.  For each ngram, need pointers to all dictionary terms containing it – the “postings”.  Note that the same “postings entry” will arise repeatedly in parsing the docs – need efficient hashing to keep track of this.  E.g., that the trigram uou occurs in the term deciduous will be discovered on each text occurrence of deciduous  Only need to process each term once www.ThesisScientist.comIntroduction to Information Retrieval Ch. 4 Resources for today’s lecture  Chapter 4 of IIR  MG Chapter 5  Original publication on MapReduce: Dean and Ghemawat (2004)  Original publication on SPIMI: Heinz and Zobel (2003) www.ThesisScientist.com
Website URL
Comment