root / tags / 1.0.8 / xapian-core / docs / scalability.html

Revision 9599, 6.7 kB (checked in by olly, 14 months ago)

docs/bm25.html,docs/scalability.html: Remove references to Muscat
3.6.

  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
Line 
1<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
2<HTML>
3<HEAD>
4<TITLE>Xapian: Scalability</TITLE>
5</HEAD>
6<BODY BGCOLOR="white" TEXT="black">
7
8<H1>Scalability</H1>
9
10<P>People often want to know how Xapian will scale.  The short answer is "very
11well" - a previous version of the software powered BrightStation's Webtop
12search engine, which offered a search over around 500 million web pages (around
131.5 terabytes of database files).  Searches took less than a second.</P>
14
15<P>The largest recent installation we're aware of is probably <A
16HREF="http://search.gmane.org/">gmane</A>, which currently indexes
17over 50 million mail messages.</P>
18
19<H2>Benchmarking</H2>
20
21<P>One effect to be aware of when designing benchmarks is that queries will be
22a lot slower when nothing is cached.  So the first few queries on a database
23which hasn't been searched recently will be unrepresentatively slow compared to
24the typical case.</P>
25
26<P>In real use, pretty much all the non-leaf blocks from the B-trees being used
27for the search will be cached pretty quickly, as well as many commonly used
28leaf blocks.</P>
29
30<H2>General Scalability Considerations</H2>
31
32<P>In a large search application, I/O will end up being
33the limiting factor.  So you want a RAID setup optimised for fast reading,
34lots of RAM in the box so the OS can cache lots of disk blocks (the access
35patterns typically mean that you only need to cache a few percent of the
36database to eliminate most disk cache misses).</P>
37
38<P>It also means that reducing the database size is usually a win.  The Flint
39and Quartz backends both compress the information in the tables in ways which
40work well given the nature of the data but aren't too expensive to unpack (e.g.
41lists of sorted docids are stored as differences with smaller values encoded in
42fewer bytes).  There is further potential for improving the encodings used.</P>
43
44<P>Another way to reduce disk I/O is to run databases through xapian-compact.
45The Btree manager usually leaves some spare space in each block so that updates
46are more efficient (though there are heuristics which will fill blocks fuller
47if they detect a long sequence of sequential insertions so adding documents
48to the end of an empty database will produce fairly compact tables, apart from
49the postlist table).  Compacting makes all blocks as full as possible, and so
50reduces the size of the database.  It also produces a database with revision 1
51which is inherently faster to search.  The penalty is that updates will be slow
52for a while, as they'll result in a lot of block splitting when all blocks are
53full.</P>
54
55<P>Splitting the data over several databases is generally a good strategy.
56Once each has finished being updated, compact it
57to make it small and faster to search.</P>
58
59<P>A multiple-database scheme works particularly well if you want a rolling web
60index where the contents of the oldest database can be rechecked and live links
61put back into a new database which, once built, replaces the oldest database.
62It's also good for a news-type application where older documents should expire
63from the index.</P>
64
65<P>You could take this idea further by implementing an ultra-compact read-only
66backend which would take a flint Btree and convert it to something like
67a cdb hashed file.  The same information would be stored, but with less
68overhead than the flint Btrees (which need to allow for update).  If
69you need serious performance, implementing such a backend is worth
70considering.</P>
71
72<H2>Size Limits in Xapian</H2>
73
74<P>The flint backend (which is now the recommended backend) stores
75the indexes in several files containing Btree tables.  If you're indexing with
76positional information (for phrase searching) the term positions table is
77usually the largest.</P>
78
79<P>
80The current limits are:
81
82<ul>
83<li>
84  Xapian uses unsigned 32-bit ints for document ids, so you're limited to
85  just over 4 billion documents in a database (as of 2003-09-10 that's
86  more than a third larger than Google).  The other limits will cut in first
87  for a single database, but searches over multiple databases are done by
88  interleaving the document ids, so this might start to matter (especially if
89  one database is much larger than the others).  This interleaving
90  technique could be changed fairly easily if it proves problematic.
91
92<li>
93  If you search many databases concurrently, you may hit the per-process
94  file-descriptor limit - each flint database uses 5 fds.  Some Unix-like OSes
95  allow this limit to be raised.  Another way to avoid it (and to spread the
96  search load) is to use the remote backend to search databases on a cluster of
97  machines.  Flint could be made to not open fds for tables which aren't being
98  used during search (values and positions may not be), or to juggle fds - the
99  record table is typically only used for results, while the posting table is
100  typically only used during matching.
101
102<li>
103  If the OS has a filesize limit, that obviously applies to Xapian (a 2GB limit
104  is common for older operating systems).  The xapian-core configure script
105  will attempt to detect and automatically enable support for "LARGE FILES"
106  where possible.
107  <P>
108  So what is the limit for a modern OS?  Taking Linux 2.4 with the ext2
109  filing system as an example, quoting from
110  linux/Documentation/filesystems/ext2.txt:
111  <blockquote><pre>
112  Filesystem block size:     1kB        2kB        4kB        8kB
113
114  File size limit:          16GB      256GB     2048GB     2048GB
115  Filesystem size limit:  2047GB     8192GB    16384GB    32768GB
116
117  There is a 2.4 kernel limit of 2048GB for a single block device, so no
118  filesystem larger than that can be created at this time.  There is also
119  an upper limit on the block size imposed by the page size of the kernel,
120  so 8kB blocks are only allowed on Alpha systems (and other architectures
121  which support larger pages).
122  </pre></blockquote>
123
124<li>
125  The B-trees use a 32-bit unsigned block count.  The default blocksize is 8K
126  which limits you to 32TB tables.  You can increase the blocksize if this
127  is a problem, but it's best to do it before you create the database as
128  otherwise you need to use xapian-compact to rebuild the
129  database with the new blocksize, and that will take a while for such a large
130  database.  The maximum blocksize allowed is 64K, which limits you to
131  256TB tables.
132
133<li>
134  Flint stores the total length (i.e. number of terms) of all the documents in
135  a database so it can calculate the average document length.  This is
136  currently stored as an unsigned 64-bit quantity so it's not likely to be a
137  limit you'll hit.  It's listed for completeness.
138</ul>
139
140If you've further questions about scalability, ask on the mailing lists -
141people using Xapian to search large databases may be able to make further
142suggestions.
143
144<!-- FOOTER $Author$ $Date$ $Id$ -->
145</BODY>
146</HTML>
Note: See TracBrowser for help on using the browser.