7 This document is out-of-date. It describes an intermediate file format used
8 during the development from 0.983 to 1.0000. It will be rewritten soon.
12 This is a document describing the internal workings of L<DBM::Deep>. It is
13 not necessary to read this document if you only intend to be a user. This
14 document is intended for people who either want a deeper understanding of
15 specifics of how L<DBM::Deep> works or who wish to help program
20 L<DBM::Deep> is broken up into five classes in three inheritance hierarchies.
26 L<DBM::Deep> is the parent of L<DBM::Deep::Array> and L<DBM::Deep::Hash>.
27 These classes form the immediate interface to the outside world. They are the
28 classes that provide the TIE mechanisms as well as the OO methods.
32 L<DBM::Deep::Engine> is the layer that deals with the mechanics of reading
33 and writing to the file. This is where the logic of the file layout is
38 L<DBM::Deep::File> is the layer that deals with the physical file. As a
39 singleton that every other object has a reference to, it also provides a place
40 to handle datastructure-wide items, such as transactions.
46 DBM::Deep uses a tagged file layout. Every section has a tag, a size, and then
53 =item * File Signature
55 The first four bytes are 'DPDB' in network byte order, signifying that this is
60 This is the tagging of the file header. The file used by versions prior to
61 1.00 had a different fifth byte, allowing the difference to the determined.
65 This is four bytes containing the file version. This lets the file format change over time.
69 These are the file-wide constants that determine how the file is laid out.
70 They can only be set upon file creation.
72 =item * Transaction information
74 The current running transactions are stored here, as is the next transaction
77 =item * Freespace information
79 Pointers into the next free sectors of the various sector sizes (Index,
80 Bucketlist, and Data) are stored here.
86 The Index parts can be tagged either as Hash, Array, or Index. The latter
87 is if there was a reindexing due to a bucketlist growing too large. The others
88 are the root index for their respective datatypes. The index consists of a
89 tag, a size, and then 256 sections containing file locations. Each section
90 corresponds to each value representable in a byte.
92 The index is used as follows - whenever a hashed key is being looked up, the
93 first byte is used to determine which location to go to from the root index.
94 Then, if that's also an index, the second byte is used, and so forth until a
99 This is the part that contains the link to the data section. A bucketlist
100 defaults to being 16 buckets long (modifiable by the I<max_buckets>
101 parameter used when creating a new file). Each bucket contains an MD5 and a
102 location of the appropriate key section.
106 This is the part that handles transactional awareness. There are
107 I<max_buckets> sections. Each section contains the location to the data
108 section, a transaction ID, and whether that transaction considers this key to
113 This is the part that actual stores the key, value, and class (if
114 appropriate). The layout is:
120 =item * length of the value
122 =item * the actual value
126 =item * the actual key
128 =item * a byte indicating if this value has a classname
130 =item * the classname (if one is there)
134 The key is stored after the value because the value is requested more often
139 L<DBM::Deep> is written completely in Perl. It also is a multi-process DBM
140 that uses the datafile as a method of synchronizing between multiple
141 processes. This is unlike most RDBMSes like MySQL and Oracle. Furthermore,
142 unlike all RDBMSes, L<DBM::Deep> stores both the data and the structure of
143 that data as it would appear in a Perl program.
147 DBM::Deep attempts to be CPU-light. As it stores all the data on disk,
148 DBM::Deep is I/O-bound, not CPU-bound.
152 DBM::Deep uses extremely little RAM relative to the amount of data you can
153 access. You can iterate through a million keys (using C<each()>) without
154 increasing your memeory usage at all.
158 DBM::Deep is I/O-bound, pure and simple. The faster your disk, the faster
159 DBM::Deep will be. Currently, when performing C<my $x = $db-E<gt>{foo}>, there
160 are a minimum of 4 seeks and 1332 + N bytes read (where N is the length of your
161 data). (All values assume a medium filesize.) The actions taken are:
165 =item 1 Lock the file
167 =item 1 Perform a stat() to determine if the inode has changed
169 =item 1 Go to the primary index for the $db (1 seek)
171 =item 1 Read the tag/size of the primary index (5 bytes)
173 =item 1 Read the body of the primary index (1024 bytes)
175 =item 1 Go to the bucketlist for this MD5 (1 seek)
177 =item 1 Read the tag/size of the bucketlist (5 bytes)
179 =item 1 Read the body of the bucketlist (144 bytes)
181 =item 1 Go to the keys location for this MD5 (1 seek)
183 =item 1 Read the tag/size of the keys section (5 bytes)
185 =item 1 Read the body of the keys location (144 bytes)
187 =item 1 Go to the data section that corresponds to this transaction ID. (1 seek)
189 =item 1 Read the tag/size of the data section (5 bytes)
191 =item 1 Read the value for this data (N bytes)
193 =item 1 Unlock the file
197 Every additional level of indexing (if there are enough keys) requires an
198 additional seek and the reading of 1029 additional bytes. If the value is
199 blessed, an additional 1 seek and 9 + M bytes are read (where M is the length
202 Arrays are (currently) even worse because they're considered "funny hashes"
203 with the length stored as just another key. This means that if you do any sort
204 of lookup with a negative index, this entire process is performed twice - once
205 for the length and once for the value.
211 Obviously, DBM::Deep isn't going to be as fast as some C-based DBMs, such as
212 the almighty I<BerkeleyDB>. But it makes up for it in features like true
213 multi-level hash/array support, and cross-platform FTPable files. Even so,
214 DBM::Deep is still pretty fast, and the speed stays fairly consistent, even
215 with huge databases. Here is some test data:
217 Adding 1,000,000 keys to new DB file...
219 At 100 keys, avg. speed is 2,703 keys/sec
220 At 200 keys, avg. speed is 2,642 keys/sec
221 At 300 keys, avg. speed is 2,598 keys/sec
222 At 400 keys, avg. speed is 2,578 keys/sec
223 At 500 keys, avg. speed is 2,722 keys/sec
224 At 600 keys, avg. speed is 2,628 keys/sec
225 At 700 keys, avg. speed is 2,700 keys/sec
226 At 800 keys, avg. speed is 2,607 keys/sec
227 At 900 keys, avg. speed is 2,190 keys/sec
228 At 1,000 keys, avg. speed is 2,570 keys/sec
229 At 2,000 keys, avg. speed is 2,417 keys/sec
230 At 3,000 keys, avg. speed is 1,982 keys/sec
231 At 4,000 keys, avg. speed is 1,568 keys/sec
232 At 5,000 keys, avg. speed is 1,533 keys/sec
233 At 6,000 keys, avg. speed is 1,787 keys/sec
234 At 7,000 keys, avg. speed is 1,977 keys/sec
235 At 8,000 keys, avg. speed is 2,028 keys/sec
236 At 9,000 keys, avg. speed is 2,077 keys/sec
237 At 10,000 keys, avg. speed is 2,031 keys/sec
238 At 20,000 keys, avg. speed is 1,970 keys/sec
239 At 30,000 keys, avg. speed is 2,050 keys/sec
240 At 40,000 keys, avg. speed is 2,073 keys/sec
241 At 50,000 keys, avg. speed is 1,973 keys/sec
242 At 60,000 keys, avg. speed is 1,914 keys/sec
243 At 70,000 keys, avg. speed is 2,091 keys/sec
244 At 80,000 keys, avg. speed is 2,103 keys/sec
245 At 90,000 keys, avg. speed is 1,886 keys/sec
246 At 100,000 keys, avg. speed is 1,970 keys/sec
247 At 200,000 keys, avg. speed is 2,053 keys/sec
248 At 300,000 keys, avg. speed is 1,697 keys/sec
249 At 400,000 keys, avg. speed is 1,838 keys/sec
250 At 500,000 keys, avg. speed is 1,941 keys/sec
251 At 600,000 keys, avg. speed is 1,930 keys/sec
252 At 700,000 keys, avg. speed is 1,735 keys/sec
253 At 800,000 keys, avg. speed is 1,795 keys/sec
254 At 900,000 keys, avg. speed is 1,221 keys/sec
255 At 1,000,000 keys, avg. speed is 1,077 keys/sec
257 This test was performed on a PowerMac G4 1gHz running Mac OS X 10.3.2 & Perl
258 5.8.1, with an 80GB Ultra ATA/100 HD spinning at 7200RPM. The hash keys and
259 values were between 6 - 12 chars in length. The DB file ended up at 210MB.
260 Run time was 12 min 3 sec.
264 One of the great things about L<DBM::Deep> is that it uses very little memory.
265 Even with huge databases (1,000,000+ keys) you will not see much increased
266 memory on your process. L<DBM::Deep> relies solely on the filesystem for storing
267 and fetching data. Here is output from I<top> before even opening a database
270 PID USER PRI NI SIZE RSS SHARE STAT %CPU %MEM TIME COMMAND
271 22831 root 11 0 2716 2716 1296 R 0.0 0.2 0:07 perl
273 Basically the process is taking 2,716K of memory. And here is the same
274 process after storing and fetching 1,000,000 keys:
276 PID USER PRI NI SIZE RSS SHARE STAT %CPU %MEM TIME COMMAND
277 22831 root 14 0 2772 2772 1328 R 0.0 0.2 13:32 perl
279 Notice the memory usage increased by only 56K. Test was performed on a 700mHz
280 x86 box running Linux RedHat 7.2 & Perl 5.6.1.