1 | A CRASH COURSE GUIDE TO MGMERGE
|
---|
2 | ===============================
|
---|
3 |
|
---|
4 | Shane Hudson
|
---|
5 | 15 November 1994
|
---|
6 |
|
---|
7 |
|
---|
8 | This document is intended as a note to the maintainers of mgmerge
|
---|
9 | outlining the mgmerge utility, the changes
|
---|
10 | that had to be made to the existing mg code to make mgmerge
|
---|
11 | work, and the new source code for mgmerge.
|
---|
12 |
|
---|
13 |
|
---|
14 | **** NOTE: ****
|
---|
15 | All the NEW/CHANGED files are in the file "new_mg.tar" which
|
---|
16 | you should have received with this; files that were
|
---|
17 | not changed are NOT in the tar file so it will have to be
|
---|
18 | extracted "on top of" the existing code. You'll figure it out :)
|
---|
19 |
|
---|
20 |
|
---|
21 | **** DESCRIPTION: ****
|
---|
22 | mgmerge adds documents to an existing mg database
|
---|
23 | without the need to rebuild the database from scratch.
|
---|
24 | It works by building a temporary new database from the
|
---|
25 | new documents, then merging the "old" and "new" databases
|
---|
26 | to give one merged database.
|
---|
27 |
|
---|
28 |
|
---|
29 | The source consists of:
|
---|
30 | mgmerge.sh The main script, a lot like mgbuild.sh
|
---|
31 | mg_get_merge.sh Script to retrieve new documents, a lot like
|
---|
32 | mg_get.sh
|
---|
33 | mg_merge.h header file used by mg_invf_merge.c
|
---|
34 | and mg_text_merge.c
|
---|
35 | mg_text_merge.c Program to append one compressed text file
|
---|
36 | to another
|
---|
37 | mg_invf_merge.c Program to merge two inverted files and
|
---|
38 | stemmed dictionaries
|
---|
39 |
|
---|
40 | See the man pages for mgmerge, mg_get_merge, mg_text_merge and
|
---|
41 | mg_invf_merge for a brief overview.
|
---|
42 |
|
---|
43 | mg_get_merge.sh should really be put into mg_get.sh, with a "-merge"
|
---|
44 | option if it is being called by the mgmerge utility, but I was
|
---|
45 | too lazy to change it to that.
|
---|
46 |
|
---|
47 |
|
---|
48 | The files updated by mgmerge are:
|
---|
49 |
|
---|
50 | *.text with mg_text_merge
|
---|
51 | *.text.idx with mg_text_merge
|
---|
52 | *.text.idx.wgt with mg_weights_build (after merging)
|
---|
53 | *.invf with mg_invf_merge
|
---|
54 | *.invf.idx with "
|
---|
55 | *.invf.dict with "
|
---|
56 | *.weight with mg_invf_merge or mg_weights_build
|
---|
57 | *.invf.dict.blocked with mg_invf_dict (after merging)
|
---|
58 |
|
---|
59 | The *.text.stats file and *.text.dict file remain the same.
|
---|
60 | Any other database files (ie, *.invf.dict.hash)
|
---|
61 | will be out of date after a merge and will need to be recomputed.
|
---|
62 |
|
---|
63 | --------------------------------------------------------------------------
|
---|
64 |
|
---|
65 | MG_TEXT_MERGE
|
---|
66 | =============
|
---|
67 |
|
---|
68 | The two "merging" utilities are mg_text_merge and mg_invf_merge
|
---|
69 |
|
---|
70 | mg_text_merge simply appends the compressed text for the new
|
---|
71 | documents to the old compressed text file.
|
---|
72 | For this to succeed, the two files should have been created
|
---|
73 | with the same parameters to mg_passes and the same compression
|
---|
74 | model must be used. So *.text.dict from the old (static) database
|
---|
75 | is used as the model for compressing the new documents.
|
---|
76 | The default "-C" option to mg_compression_dict in mgbuild was
|
---|
77 | changed to "-S" so novel words in the new documents can be coded.
|
---|
78 | The option must be the same in mgmerge -- it is "-S" there too.
|
---|
79 |
|
---|
80 | ---------------
|
---|
81 |
|
---|
82 | EFFECT on FILE COMPRESSION PERFORMANCE
|
---|
83 |
|
---|
84 | Since the compression model is not being updated,compression
|
---|
85 | performance for the text will slowly degrade. The only solution
|
---|
86 | is a periodic rebuild from scratch.
|
---|
87 |
|
---|
88 | ---------------------------------------------------------------------------
|
---|
89 |
|
---|
90 | MG_INVF_MERGE
|
---|
91 | =============
|
---|
92 |
|
---|
93 | mg_invf_merge updates the stemmed dictionary and
|
---|
94 | inverted file.
|
---|
95 | For reasons of simplicity, the stemmed dictionaries merged
|
---|
96 | are in the ".invf.dict" format, NOT the ".invf.dict.blocked"
|
---|
97 | format.
|
---|
98 | So if a database is to be merged the .invf.dict file should
|
---|
99 | not be deleted after the .invf.dict.blocked file used by mgquery
|
---|
100 | has been created.
|
---|
101 | Also for simplicity, level 3 inverted files are not supported, nor
|
---|
102 | are skipped format files.
|
---|
103 |
|
---|
104 |
|
---|
105 |
|
---|
106 | Merging the inverted files requires every entry in each file
|
---|
107 | to be decoded and merged if an entry for the same term appears
|
---|
108 | in the other inverted file.
|
---|
109 | This can be a slow process if the old inverted file is very large.
|
---|
110 | Especially since each entry is decoded bit-by-bit with
|
---|
111 | the bit-level I/O rouutines.
|
---|
112 | As was noted on page 94 of the book "Managing Gigabytes" (hereafter
|
---|
113 | called "MG"), switching from bernoulli coding (called "Bblock" in
|
---|
114 | the mg source code) to gamma or delta coding for the
|
---|
115 | inter-document gaps would be an advantage since they require no
|
---|
116 | parameters.
|
---|
117 | With Bblock coding, N (the number of documents) is a parameter and
|
---|
118 | as N changes whenever documents are added, every entry must
|
---|
119 | be decoded and recoded.
|
---|
120 | My approach was to keep Bblock coding, but keep the N parameter
|
---|
121 | constant by recording the artificial N used to code document
|
---|
122 | gaps (called "Nstatic") as well as the real number of documents.
|
---|
123 |
|
---|
124 | Then, any entry in the old inverted file (called "IFold") that
|
---|
125 | is not merged with an entry in the new inverted file (called "IFnew")
|
---|
126 | can be copied directly to the merged inverted file (called "IFmerge").
|
---|
127 | This can give a significant increase in speed, since usually the
|
---|
128 | size of IFnew is very small if only a few documents are being added.
|
---|
129 |
|
---|
130 | To store Nstatic, a field "static_num_of_docs" was added to
|
---|
131 | the invf_dict_header and stem_dict_header structs defined
|
---|
132 | in invf.h
|
---|
133 | Any program that decodes inverted file entries also had to be changed.
|
---|
134 | (The existing mg source code that I changed has been returned with
|
---|
135 | this document.)
|
---|
136 | Most changes were simply altering a line that
|
---|
137 | called BIO_Bblock_Init() so it used the static_num_of_docs
|
---|
138 | value rather than num_of_docs.
|
---|
139 |
|
---|
140 | So since the file format has changed, any existing collection will
|
---|
141 | have to be rebuilt first even using mgmerge with it is not
|
---|
142 | intended.
|
---|
143 |
|
---|
144 | -------------
|
---|
145 |
|
---|
146 | EFFECT on INVERTED FILE COMPRESSION PERFORMANCE
|
---|
147 |
|
---|
148 | As more merges are accumulatively performed on a database, the
|
---|
149 | compression performance will decline.
|
---|
150 | The solution is an option in mg_invf_merge to do a "slow"
|
---|
151 | merge where every inverted file entry is decoded and recoded using
|
---|
152 | the real value of N.
|
---|
153 | This is not a lot slower than a fast merge when compared to the
|
---|
154 | cost of completely rebuilding the collection from scratch, and
|
---|
155 | the option can be used periodically (when, say, 40% of the
|
---|
156 | inverted file has been added since the previous slow merge).
|
---|
157 |
|
---|
158 | -----------------
|
---|
159 |
|
---|
160 | THE EXACT WEIGHTS FILE
|
---|
161 |
|
---|
162 | Recomputing exact document weights requires a complete scan over the
|
---|
163 | merged inverted file. This is a waste since the weights for old
|
---|
164 | documents will (hopefully) not change much anyway, and the weights
|
---|
165 | for new documents can be computed exactly when merging the inverted files,
|
---|
166 | at no extra time cost.
|
---|
167 | So mg_invf_merge computes the weights for new documents and
|
---|
168 | updates the .weight file.
|
---|
169 | The weights for old documents are left untouched.
|
---|
170 | mgmerge has an option to recompute the weights file.
|
---|
171 | This is not much more expensive in terms of time, and is only
|
---|
172 | periodically needed.
|
---|
173 |
|
---|
174 | The effect of leaving the old weights unchanged is difficult to
|
---|
175 | asess since the "correct" ranking of documents for a query is
|
---|
176 | subjective.
|
---|
177 | The frequency with which a rebuild of the weights file with the
|
---|
178 | "-w" option should be done is hard to guess.
|
---|
179 | But if most merges involves only a few documents, it makes sense
|
---|
180 | to not bother recomputing it since the change in old weights values
|
---|
181 | is very small and the weights only stray from their "true"
|
---|
182 | values slowly over accumulated merging.
|
---|
183 |
|
---|
184 | -------------------------------------------------------------------------
|
---|
185 |
|
---|
186 | EXPERIMENTAL RESULTS
|
---|
187 | ====================
|
---|
188 |
|
---|
189 | With some small test collections (only a few Mb of text)
|
---|
190 | and a slow merge, and a weights file rebuild,
|
---|
191 | mgmerge typically took under 20% of the time to completely
|
---|
192 | rebuild. Using the default options (fast merge, dont rebuild weights)
|
---|
193 | took around 10% or more.
|
---|
194 | For larger collections, the savings were greater, especially if the
|
---|
195 | added text is very small.
|
---|
196 | For example, when the last 6Kb of text of the GNUBib collection was
|
---|
197 | added to the rest of it (14Mb) using mgmerge, the total mgmerge time
|
---|
198 | was around 20 seconds whereas a complete mgbuild took 280 seconds.
|
---|
199 | Only one larger collection was tested (resource contraints prevented
|
---|
200 | testing on any really large collections): two short (one-line)
|
---|
201 | documents were added to the Gutenberg collection, which comprised
|
---|
202 | of nearly 74Mb of source text and had an inverted file size of 9Mb.
|
---|
203 | mgmerge took 42 seconds (or 142 seconds with a slow merge and rebuilding the
|
---|
204 | weights file), and I'd estimate that an mgbuild on the same machine would
|
---|
205 | have taken at least 20 minutes.
|
---|
206 |
|
---|
207 | To summarise, mgmerge is not as fast as some sophisticated method
|
---|
208 | such as the fixed-length blocks described in section 5.7 of "MG",
|
---|
209 | but its huge advantage is that it required almost no changes to
|
---|
210 | the existing mg code and is still far better than using mgbuild
|
---|
211 | every time a document needs to be added to a collection.
|
---|
212 |
|
---|
213 | ----------------------------------------------------------------------------
|
---|
214 |
|
---|
215 | OTHER NOTES
|
---|
216 | ===========
|
---|
217 |
|
---|
218 | Since it works like mgbuild, mgmerge needs mg_get_merge
|
---|
219 | to return the SAME text each time "mg_get_merge -text .." is called.
|
---|
220 |
|
---|
221 | ----
|
---|
222 |
|
---|
223 | One other thing discovered (but not fixed) was a bug in "invf.pass2.c":
|
---|
224 | It crashes on parsed words longer than 128 characters.
|
---|
225 | So if any text with words >128 chars is piped to mg_passes,
|
---|
226 | the "-N1" and "-N2" options (the memory-efficient inversion method)
|
---|
227 | had better be used and not the "-I1"/"-I2" option (which would only
|
---|
228 | work on small collections anyway, being the memory-inefficient
|
---|
229 | method).
|
---|
230 |
|
---|