The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.
=head1 NAME

DBM::Deep::Internals - Out of date documentation on DBM::Deep internals

=head1 OUT OF DATE

This document is out-of-date. It describes an intermediate file format used
during the development from 0.983 to 1.0000. It will be rewritten soon.

So far, the description of the header format has been updated.

=head1 DESCRIPTION

This is a document describing the internal workings of L<DBM::Deep>. It is
not necessary to read this document if you only intend to be a user. This
document is intended for people who either want a deeper understanding of
specifics of how L<DBM::Deep> works or who wish to help program
L<DBM::Deep>.

=head1 CLASS LAYOUT

L<DBM::Deep> is broken up into five classes in three inheritance hierarchies.

=over 4

=item *

L<DBM::Deep> is the parent of L<DBM::Deep::Array> and L<DBM::Deep::Hash>.
These classes form the immediate interface to the outside world. They are the
classes that provide the TIE mechanisms as well as the OO methods.

=item *

L<DBM::Deep::Engine> is the layer that deals with the mechanics of reading
and writing to the file. This is where the logic of the file layout is
handled.

=item *

L<DBM::Deep::File> is the layer that deals with the physical file. As a
singleton that every other object has a reference to, it also provides a place
to handle datastructure-wide items, such as transactions.

=back

=head1 FILE LAYOUT

This describes the 1.0003 and 2.0000 formats, which internally are numbered
3 and 4, respectively. The internal numbers are used in this section. These
two formats are almost identical.

DBM::Deep uses a tagged file layout. Every section has a tag, a size, and then
the data.

=head2 File header

The file header consists of two parts. The first part is a fixed length of
13 bytes:

  DPDB h VVVV SSSS
  \  / |    \   \
   \/  '---. \   '--- size of the second part of the header
  file      \ '--- version
 signature  tag
  
=over 4

=item * File Signature

The first four bytes are 'DPDB' in network byte order, signifying that this is
a DBM::Deep file.

=item * File tag

A literal ASCII 'h', indicating that this is the header. The file used by
versions prior to 1.00 had a different fifth byte, allowing the difference
to be determined.

=item * Version

This is four bytes containing the file version. This lets the file format change over time.

It is packed in network order, so version 4 is stored as "\0\0\0\cD".

=item * Header size

The size of the second part of the header, in bytes. This number is also
packed in network order.

=back

The second part of the header is as follows:

  S B S T T(TTTTTTTTT...) (SS SS SS SS ...)  (continued...)
  | | | |              \       |
  | | | '----------.    \  staleness counters
  | | '--------.    \  txn bitfield
  | '------.    \  number of transactions
 byte size  \  data sector size
          max buckets

 (continuation...)
  BB(BBBBBB) DD(DDDDDD) II(IIIIII)
      |        |            |
      |    free data        |
  free blist           free index

=over

=item * Constants

These are the file-wide constants that determine how the file is laid out.
They can only be set upon file creation.

The byte size is the number of bytes used to point to an offset elsewhere
in the file. This corresponds to the C<pack_size> option. This and the
next three values are stored as packed 8-bit integers (chars), so 2 is
represented by "\cB".

C<max_buckets> and C<data_sector_size> are documented in the main
L<DBM::Deep> man page. The number stored is actually one less than what is
passed to the constructor, to allow for a range of 1-256.

The number of transactions corresponds to the C<num_txns> value passed to
the constructor.

=item * Transaction information

The transaction bitfield consists of one bit for every available
transaction ID. It is therefore anywhere from 1 byte to 32 bytes long.

The staleness counters each take two bytes (packed 32-bit integers), one
for each transaction, not including the so-called HEAD (the main
transaction that all processes share I<before> calling C<begin_work>). So
these take up 0 to 508 bytes.

Staleness is explained in L<DBM::Deep::Engine|DBM::Deep::Engine/STALENESS>.

=item * Freespace information

Pointers into the first free sectors of the various sector sizes (Index,
Bucketlist, and Data) are stored here. These are called chains internally,
as each free sector points to the next one.

The number of bytes is determined by the byte size, ranging from 2 to 8.

=back

=head2 Index

The Index parts can be tagged either as Hash, Array, or Index. The latter
is if there was a reindexing due to a bucketlist growing too large. The others
are the root index for their respective datatypes. The index consists of a
tag, a size, and then 256 sections containing file locations. Each section
corresponds to each value representable in a byte.

The index is used as follows - whenever a hashed key is being looked up, the
first byte is used to determine which location to go to from the root index.
Then, if that's also an index, the second byte is used, and so forth until a
bucketlist is found.

=head2 Bucketlist

This is the part that contains the link to the data section. A bucketlist
defaults to being 16 buckets long (modifiable by the I<max_buckets>
parameter used when creating a new file). Each bucket contains an MD5 and a
location of the appropriate key section.

=head2 Key area

This is the part that handles transactional awareness. There are
I<max_buckets> sections. Each section contains the location to the data
section, a transaction ID, and whether that transaction considers this key to
be deleted or not.

=head2 Data area

This is the part that actual stores the key, value, and class (if
appropriate). The layout is:

=over 4

=item * tag

=item * length of the value

=item * the actual value

=item * keylength

=item * the actual key

=item * a byte indicating if this value has a classname

=item * the classname (if one is there)

=back

The key is stored after the value because the value is requested more often
than the key.

=head1 PERFORMANCE

L<DBM::Deep> is written completely in Perl. It also is a multi-process DBM
that uses the datafile as a method of synchronizing between multiple
processes. This is unlike most RDBMSes like MySQL and Oracle. Furthermore,
unlike all RDBMSes, L<DBM::Deep> stores both the data and the structure of
that data as it would appear in a Perl program.

=head2 CPU

DBM::Deep attempts to be CPU-light. As it stores all the data on disk,
DBM::Deep is I/O-bound, not CPU-bound.

=head2 RAM

DBM::Deep uses extremely little RAM relative to the amount of data you can
access. You can iterate through a million keys (using C<each()>) without
increasing your memory usage at all.

=head2 DISK

DBM::Deep is I/O-bound, pure and simple. The faster your disk, the faster
DBM::Deep will be. Currently, when performing C<my $x = $db-E<gt>{foo}>, there
are a minimum of 4 seeks and 1332 + N bytes read (where N is the length of your
data). (All values assume a medium filesize.) The actions taken are:

=over 4

=item 1 Lock the file

=item 1 Perform a stat() to determine if the inode has changed

=item 1 Go to the primary index for the $db (1 seek)

=item 1 Read the tag/size of the primary index (5 bytes)

=item 1 Read the body of the primary index (1024 bytes)

=item 1 Go to the bucketlist for this MD5 (1 seek)

=item 1 Read the tag/size of the bucketlist (5 bytes)

=item 1 Read the body of the bucketlist (144 bytes)

=item 1 Go to the keys location for this MD5 (1 seek)

=item 1 Read the tag/size of the keys section (5 bytes)

=item 1 Read the body of the keys location (144 bytes)

=item 1 Go to the data section that corresponds to this transaction ID. (1 seek)

=item 1 Read the tag/size of the data section (5 bytes)

=item 1 Read the value for this data (N bytes)

=item 1 Unlock the file

=back

Every additional level of indexing (if there are enough keys) requires an
additional seek and the reading of 1029 additional bytes. If the value is
blessed, an additional 1 seek and 9 + M bytes are read (where M is the length
of the classname).

Arrays are (currently) even worse because they're considered "funny hashes"
with the length stored as just another key. This means that if you do any sort
of lookup with a negative index, this entire process is performed twice - once
for the length and once for the value.

=head1 ACTUAL TESTS

=head2 SPEED

Obviously, DBM::Deep isn't going to be as fast as some C-based DBMs, such as
the almighty I<BerkeleyDB>.  But it makes up for it in features like true
multi-level hash/array support, and cross-platform FTPable files.  Even so,
DBM::Deep is still pretty fast, and the speed stays fairly consistent, even
with huge databases.  Here is some test data:

    Adding 1,000,000 keys to new DB file...

    At 100 keys, avg. speed is 2,703 keys/sec
    At 200 keys, avg. speed is 2,642 keys/sec
    At 300 keys, avg. speed is 2,598 keys/sec
    At 400 keys, avg. speed is 2,578 keys/sec
    At 500 keys, avg. speed is 2,722 keys/sec
    At 600 keys, avg. speed is 2,628 keys/sec
    At 700 keys, avg. speed is 2,700 keys/sec
    At 800 keys, avg. speed is 2,607 keys/sec
    At 900 keys, avg. speed is 2,190 keys/sec
    At 1,000 keys, avg. speed is 2,570 keys/sec
    At 2,000 keys, avg. speed is 2,417 keys/sec
    At 3,000 keys, avg. speed is 1,982 keys/sec
    At 4,000 keys, avg. speed is 1,568 keys/sec
    At 5,000 keys, avg. speed is 1,533 keys/sec
    At 6,000 keys, avg. speed is 1,787 keys/sec
    At 7,000 keys, avg. speed is 1,977 keys/sec
    At 8,000 keys, avg. speed is 2,028 keys/sec
    At 9,000 keys, avg. speed is 2,077 keys/sec
    At 10,000 keys, avg. speed is 2,031 keys/sec
    At 20,000 keys, avg. speed is 1,970 keys/sec
    At 30,000 keys, avg. speed is 2,050 keys/sec
    At 40,000 keys, avg. speed is 2,073 keys/sec
    At 50,000 keys, avg. speed is 1,973 keys/sec
    At 60,000 keys, avg. speed is 1,914 keys/sec
    At 70,000 keys, avg. speed is 2,091 keys/sec
    At 80,000 keys, avg. speed is 2,103 keys/sec
    At 90,000 keys, avg. speed is 1,886 keys/sec
    At 100,000 keys, avg. speed is 1,970 keys/sec
    At 200,000 keys, avg. speed is 2,053 keys/sec
    At 300,000 keys, avg. speed is 1,697 keys/sec
    At 400,000 keys, avg. speed is 1,838 keys/sec
    At 500,000 keys, avg. speed is 1,941 keys/sec
    At 600,000 keys, avg. speed is 1,930 keys/sec
    At 700,000 keys, avg. speed is 1,735 keys/sec
    At 800,000 keys, avg. speed is 1,795 keys/sec
    At 900,000 keys, avg. speed is 1,221 keys/sec
    At 1,000,000 keys, avg. speed is 1,077 keys/sec

This test was performed on a PowerMac G4 1gHz running Mac OS X 10.3.2 & Perl
5.8.1, with an 80GB Ultra ATA/100 HD spinning at 7200RPM.  The hash keys and
values were between 6 - 12 chars in length.  The DB file ended up at 210MB.
Run time was 12 min 3 sec.

=head2 MEMORY USAGE

One of the great things about L<DBM::Deep> is that it uses very little memory.
Even with huge databases (1,000,000+ keys) you will not see much increased
memory on your process.  L<DBM::Deep> relies solely on the filesystem for storing
and fetching data.  Here is output from I<top> before even opening a database
handle:

    PID USER     PRI  NI  SIZE  RSS SHARE STAT %CPU %MEM   TIME COMMAND
  22831 root      11   0  2716 2716  1296 R     0.0  0.2   0:07 perl

Basically the process is taking 2,716K of memory.  And here is the same
process after storing and fetching 1,000,000 keys:

    PID USER     PRI  NI  SIZE  RSS SHARE STAT %CPU %MEM   TIME COMMAND
  22831 root      14   0  2772 2772  1328 R     0.0  0.2  13:32 perl

Notice the memory usage increased by only 56K.  Test was performed on a 700mHz
x86 box running Linux RedHat 7.2 & Perl 5.6.1.

=cut