# $OpenLDAP: pkg/openldap-guide/admin/tuning.sdf,v 1.8.2.1 2006/01/03 22:16:03 kurt Exp $
# Copyright 1999-2006 The OpenLDAP Foundation, All Rights Reserved.
# COPYING RESTRICTIONS APPLY, see COPYRIGHT.
H1: Performance Tuning
Note: this chapter needs to be updated to discuss BDB tuning.
There are several things you can do to tune the performance of
slapd for your system. Most of them have to do with the LDBM
backend. LDBM uses an index mechanism to store and retrieve
information in slapd. Each entry is assigned a unique ID, used to
refer to the entry in the indexes. A search for entries with a
surname of "Jensen", for example, would look up the index entry
"=JENSEN" in the surname index. The data returned is a list of
IDs of entries having that value for the surname attribute. We
have found several things to be useful in improving the
performance of this indexing scheme, especially on modify
operations.
H2: The allIDs threshold
Some index entries become so large as to be useless. For
example, if every entry in your database is a person entry, the
"=PERSON" index entry in the objectclass index contains every
entry. This returns very little useful information, and can cause
significant delays, especially on updates. To alleviate this
problem, we have introduced the idea of an allIDs index entry.
The allIDs entry stands for a real index entry containing the IDs
of every entry in the database, but it takes up very little space,
never needs updating, and can be manipulated quickly and
efficiently. The trade-off is that it does not prune the set of
candidate entries at all during a search. This must be done
using other, more "high-powered" index entries.
You can set the minimum number of IDs that an index entry may
contain before it turns into an allIDs block by changing the
{{EX: SLAPD_LDBM_MIN_MAXIDS}} variable in the
{{EX: include/ldapconfig.h}} file. The actual number is determined at
runtime by the LDBM backend, depending on the block size of
the underlying device (i.e., the number you provide is rounded up
to the nearest multiple of a block size).
H2: The entry cache
The LDBM backend can be configured to keep a cache of
entries in memory. Since the LDBM database spends much of its
time reading entries from the id2entry file into memory, this cache
can greatly speed performance. The trade-off is that the cache
uses some extra memory. The default cache size is 1000
entries. See the discussion of the cachesize option in Section
5.2.3 on LDBM configuration.
H2: The DB cache
The LDBM backend uses a number of disk-based index files. If
the underlying hash or B-tree package supports in-memory
caching of these files, performance can be greatly improved,
especially on modifies. The size of this in-memory file cache is
given by the dbcachesize option, discussed in more detail in
section 5.2.3 on LDBM configuration. The default {{EX: dbcachesize}} is
100K.
H2: Maintain the right indices
Finally, one of the best performance tune-ups you can do is to
make sure you are maintaining the right indices. Too few indices
can lead to poor search performance. Too many indices can
lead to poor update performance. For example, the LDBM
backend would be perfectly happy to maintain substring and
approximate indices for the {{EX: objectclass attribute}}, but this would
not be useful and would just slow down update operations. If
your database has many entries and is handling queries for
substring equality on the surname attribute, you should make
sure to maintain a surname substring index so these queries are
answered quickly.
So, take a look at the index lines in your slapd configuration file to
ensure that only those indices that make sense and are needed
are being maintained.