Table of Contents
In its default configuration, UCS is suitable for environments with up to 5,000 users.
This document describes configuration modifications which can increase performance in larger environments.
As a core element in the operation and administration of a UCS domain, the performance of the LDAP server plays a central role in the overall performance.
Comparable with other database systems, OpenLDAP runs indexes about commonly requested attributes. For indexed attributes, a search is not performed via the full database contents, but over an optimized subsection.
With newer UCS versions, the indexes are occasionally expanded and automatically
activated. The automatic activation can be deactivated using the UCR variable
ldap/index/autorebuild. In this case, the indexes should be set
manually to ensure that there is no loss of performance as a result. The indexes
are controlled by the UCR variables ldap/index/eq,
ldap/index/pres, ldap/index/sub and
ldap/index/approx. Once the variables have been changed, the
OpenLDAP server must be stopped and the slapindex command run.
To determine whether not-indexed variables are used, you can activate OpenLDAP
debug level -1 and search for the string 'not indexed' in the log file
/var/log/syslog. For example:
ucr set ldap/debug/level=-1 invoke-rc.d slapd restart grep 'not indexed' /var/log/syslog
The memory mapped database (MDB) has been used for new installations since UCS
4.0. If BDB is still in use, a migration to MDB should be performed for amd64 systems (not i386).
The database backend can be controlled via the UCR variable
ldap/database/type. A migration can be performed as follows:
/etc/init.d/slapd stop slapcat -l ldif mkdir /var/lib/univention-ldap/ldap.BACKUP mv /var/lib/univention-ldap/ldap/* /var/lib/univention-ldap/ldap.BACKUP ucr set ldap/database/type=mdb slapadd -l ldif /etc/init.d/slapd start
In the default configuration the memory mapped database needs more I/O operations than
the BDB backend. With the Univention Configuration Registry variable ldap/database/mdb/envflags this
behavior can be configured. The following flags can be set (multiple values are
separated by spaces):
nosync specify that on-disk database contents should not be
immediately synchronized with in memory changes. Enabling
this option may improve performance at the expense of data security. In
particular, if the operating system crashes before changes are flushed, some
number of transactions may be lost. By default, a full data flush/sync is
performed when each transaction is committed.
nometasync Flush the data on a commit, but
skip the sync of the meta page. This mode is slightly faster than doing a full
sync, but can potentially lose the last committed transaction if the operating
system crashes. If both nometasync and nosync are set, the nosync flag takes
precedence.
writemap Use a writable memory map instead of
just read-only. This speeds up write operations but makes the database
vulnerable to corruption in case any bugs in slapd cause stray writes into the
memory mapped region.
mapasync When using a writable memory map and
performing flushes on each commit, use an asynchronous flush instead of a
synchronous flush (the default). This option has no effect if writemap has not
been set. It also has no effect if nosync is set.
nordahead Turn off file read-ahead.
Usually the OS performs read-ahead on every read request. This usually boosts
read performance but can be harmful to random access read performance if the
system's memory is full and the DB is larger than RAM.
Access to the information contained in the LDAP directory is controlled by access control lists (ACLs) on the server side. General information on the configuration of ACLs in UCS can be found in the LDAP chapter of the [ucs-manual].
Nested groups are also supported. The Univention Configuration Registry variable
ldap/acl/nestedgroups can be used to deactivate the nested
groups function for LDAP ACLs, which will result in a speed increase for
directory requests.
The Univention Directory Listener can perform safety
checks to prevent a user name being added into a group twice.
These checks add some overhead to replication and can
be deactivated by setting the Univention Configuration Registry variables
listener/memberuid/skip and
listener/uniquemember/skip to no.
Starting with UCS 3.1 the variables are not set and the checks are
not activated any longer by default.
Name resolutions can be cached by the Name Service Cache Daemon (NSCD) in order to speed up frequently recurring requests for unchanged data. Thus, if a repeated request occurs, instead of querying the LDAP server, the data are simply drawn directly from the cache.
The size of the cache held by the NSCD is preconfigured for an environment with 5,000 users. If more users or hosts are created, the cache should be enlarged as otherwise it will not be possibly to cache enough entries.
The following Univention Configuration Registry variables can be set:
nscd/hosts/size should be at least the same as the number
of all the computers entered in the DNS.
nscd/passwd/size should be at least the same as the number
of users.
To allow an efficient cache allocation, the value selected should always be a prime number, in case of doubt the next highest prime number should be selected.
A script can be downloaded from https://updates.software-univention.de/download/scripts/nscdCachesize.sh which suggests corresponding values based on the objects currently included in the system.
In the default setting, the group cache is regenerated every time changes are
made to a group. This avoids cache effects whereby group memberships only
become visible for a service after the next scheduled group
cache rewrite (in the default setting after 15 minutes and
after 15 seconds of inactivity in the Univention Directory Listener. In
larger environments with a lot of group changes, this function should be
deactivated by setting the Univention Configuration Registry variable
nss/group/invalidate_cache_on_changes to
false. This setting takes effect immediately and does not
require a restart of the Univention Directory Listener.
When the group cache file is being generated, the script verifies whether the group members
are still present in the LDAP directory. If only the Univention Management Console is used for the management of
the LDAP directory, this additional check is not necessary and can be disabled by setting the
Univention Configuration Registry variable nss/group/cachefile/check_member to false.
By default all objects are automatically searched for in the domain management
modules of the Univention Management Console. This behavior can be disabled by setting the
Univention Configuration Registry variable directory/manager/web/modules/autosearch to
0.
The Univention Configuration Registry variable directory/manager/web/sizelimit is used to impose an upper limit
for search results. If, e.g., this variable is set to 2000 (as is the default), searching for
more than 2000 users would not be performed and instead the user is asked to refine the
search.
If the Squid proxy service is used with NTLM authentication,
up to five running NTLM requests can be processed in parallel.
If many proxy requests are received in parallel, the
Squid user may occasionally receive an authentication error. The number
of parallel NTLM authentication processes can be configured with
the Univention Configuration Registry variable squid/ntlmauth/children.
BIND can use two different backends for its configuration: OpenLDAP or the internal LDB
database of Samba 4. The backend is configured via the Univention Configuration Registry variable
dns/backend.
When using the Samba backend, a search is performed in the LDAP for every DNS request. With the OpenLDAP backend, a search is only performed in the directory service if the DNS data has changed. For this reason, using the OpenLDAP backend can reduce the load on a Samba 4 domain controller.
In medium and larger environments the maximum number of open files allowed by the Linux kernel may be set too low by default. As each instance requires some unswappable memory in the Linux kernel, too many objects may lead to a resource depletion and denial-of-service problems in multi-user environments. Because of that the number of allowed file objects is limited by default.
The maximum number of open files can be configured on a per-user or per-group basis. The default for all users can be set through the following Univention Configuration Registry variables:
security/limits/user/default/hard/nofilesecurity/limits/user/default/soft/nofileA similar problem exists with the Inotify sub-system of the kernel, which can be used by all users and applications to monitor changes in file systems.
kernel/fs/inotify/max_user_instanceskernel/fs/inotify/max_user_watcheskernel/fs/inotify/max_queued_events
Samba uses its own mechanism to specify the maximum number of open files.
This can be configured through the Univention Configuration Registry variable samba/max_open_files.
The default is 32808.
If the log file /var/log/samba/log.smbd contains errors
like Failed to init inotify - Too many open
files, the kernel and Samba limits should be increased and the
services should be restarted.
The log file /var/log/univention/system-stats.log can be
checked for further performance analyses. The system status is logged every 30
minutes. If more regular logging is required, it can be controlled via the UCR
variable system/stats/cron.
Univention Corporate Server configures Dovecot to run in "High-security mode" by default. Each connection is served by a separate login process. This security has a price: for each connection at least two processes must run.
Thus installations with 10.000s of users hit operating system boundaries. For this case Dovecot offers the "High-performance mode". To activate it, login processes are allowed to serve more than one connection. To configure this run
ucr mail/dovecot/limits/imap-login/service_count=0
If client_limit=1000 and process_limit=100 are set, only 100 login processes are started, but each serves up to 1000 connections – a total of 100.000 connections.
The cost of this is that if a login process is compromised, an attacker might read the login credentials and emails of all users this login process is serving.
To distribute the load of the login processes evenly between CPU cores, mail/dovecot/limits/imap-login/process_min_avail should be set to the number of CPU cores in the system.
[ucs-manual] Univention GmbH. 2019. Univention Corporate Server - Manual for users and administrators. https://docs.software-univention.de/manual-4.4.html.