20. Tuning
This is perhaps one of the most important chapters in the guide, because if you have not tuned slapd(8) correctly or grasped how to design your directory and environment, you can expect very poor performance.
Reading, understanding and experimenting using the instructions and information in the following sections, will enable you to fully understand how to tailor your directory server to your specific requirements.
It should be noted that the following information has been collected over time from our community based FAQ. So obviously the benefit of this real world experience and advice should be of great value to the reader.
20.1. Performance Factors
Various factors can play a part in how your directory performs on your chosen hardware and environment. We will attempt to discuss these here.
20.1.1. Memory
Scale your cache to use available memory and increase system memory if you can.
See Caching
20.1.2. Disks
Use fast subsystems. Put each database and logs on separate disks configurable via DB_CONFIG:
# Data Directory
set_data_dir /data/db
# Transaction Log settings
set_lg_dir /logs
20.1.3. Network Topology
http://www.openldap.org/faq/data/cache/363.html
Drawing here.
20.1.4. Directory Layout Design
Reference to other sections and good/bad drawing here.
20.1.5. Expected Usage
Discussion.
20.2. Indexes
20.2.1. Understanding how a search works
If you're searching on a filter that has been indexed, then the search reads the index and pulls exactly the entries that are referenced by the index. If the filter term has not been indexed, then the search must read every single entry in the target scope and test to see if each entry matches the filter. Obviously indexing can save a lot of work when it's used correctly.
20.2.2. What to index
You should create indices to match the actual filter terms used in search queries.
index cn,sn,givenname,mail eq
Each attribute index can be tuned further by selecting the set of index types to generate. For example, substring and approximate search for organizations (o) may make little sense (and isn't like done very often). And searching for userPassword likely makes no sense what so ever.
General rule: don't go overboard with indexes. Unused indexes must be maintained and hence can only slow things down.
See slapd.conf(8) and slapdindex(8) for more information
20.2.3. Presence indexing
If your client application uses presence filters and if the target attribute exists on the majority of entries in your target scope, then all of those entries are going to be read anyway, because they are valid members of the result set. In a subtree where 100% of the entries are going to contain the same attributes, the presence index does absolutely NOTHING to benefit the search, because 100% of the entries match that presence filter.
So the resource cost of generating the index is a complete waste of CPU time, disk, and memory. Don't do it unless you know that it will be used, and that the attribute in question occurs very infrequently in the target data.
Almost no applications use presence filters in their search queries. Presence indexing is pointless when the target attribute exists on the majority of entries in the database. In most LDAP deployments, presence indexing should not be done, it's just wasted overhead.
See the Logging section below on what to watch our for if you have a frequently searched for attribute that is unindexed.
20.3. Logging
20.3.1. What log level to use
The default of loglevel stats (256) is really the best bet. There's a corollary to this when problems *do* arise, don't try to trace them using syslog. Use the debug flag instead, and capture slapd's stderr output. syslog is too slow for debug tracing, and it's inherently lossy - it will throw away messages when it can't keep up.
Contrary to popular belief, loglevel 0 is not ideal for production as you won't be able to track when problems first arise.
20.3.2. What to watch out for
The most common message you'll see that you should pay attention to is:
"<= bdb_equality_candidates: (foo) index_param failed (18)"
That means that some application tried to use an equality filter (foo=<somevalue>) and attribute foo does not have an equality index. If you see a lot of these messages, you should add the index. If you see one every month or so, it may be acceptable to ignore it.
The default syslog level is stats (256) which logs the basic parameters of each request; it usually produces 1-3 lines of output. On Solaris and systems that only provide synchronous syslog, you may want to turn it off completely, but usually you want to leave it enabled so that you'll be able to see index messages whenever they arise. On Linux you can configure syslogd to run asynchronously, in which case the performance hit for moderate syslog traffic pretty much disappears.
20.3.3. Improving throughput
You can improve logging performance on some systems by configuring syslog not to sync the file system with every write (man syslogd/syslog.conf). In Linux, you can prepend the log file name with a "-" in syslog.conf. For example, if you are using the default LOCAL4 logging you could try:
# LDAP logs
LOCAL4.* -/var/log/ldap
For syslog-ng, add or modify the following line in syslog-ng.conf:
options { sync(n); };
where n is the number of lines which will be buffered before a write.
20.4. Caching
We all know what caching is, don't we?
In brief, "A cache is a block of memory for temporary storage of data likely to be used again" - http://en.wikipedia.org/wiki/Cache
There are 3 types of caches, BerkeleyDB's own cache, slapd(8) entry cache and
20.4.1. Berkeley DB Cache
There are two ways to tune for the BDB cachesize:
(a) BDB cache size necessary to load the database via slapadd in optimal time
(b) BDB cache size necessary to have a high performing running slapd once the data is loaded
For (a), the optimal cachesize is the size of the entire database. If you already have the database loaded, this is simply a
du -c -h *.bdb
in the directory containing the OpenLDAP (/usr/local/var/openldap-data) data.
For (b), the optimal cachesize is just the size of the id2entry.bdb file, plus about 10% for growth.
The tuning of DB_CONFIG should be done for each BDB type database instantiated (back-bdb, back-hdb).
Note that while the
There is also an IDL cache which is used for Index Data Lookups. If you can fit all of your database into slapd's entry cache, and all of your index lookups fit in the IDL cache, that will provide the maximum throughput.
If not, but you can fit the entire database into the BDB cache, then you should do that and shrink the slapd entry cache as appropriate.
Failing that, you should balance the BDB cache against the entry cache.
It is worth noting that it is not absolutely necessary to configure a BerkeleyDB cache equal in size to your entire database. All that you need is a cache that's large enough for your "working set."
That means, large enough to hold all of the most frequently accessed data, plus a few less-frequently accessed items.
For more information, please see: http://www.oracle.com/technology/documentation/berkeley-db/db/ref/am_conf/cachesize.html
20.4.1.1. Calculating Cachesize
The back-bdb database lives in two main files, dn2id.bdb and id2entry.bdb. These are B-tree databases. We have never documented the back-bdb internal layout before, because it didn't seem like something anyone should have to worry about, nor was it necessarily cast in stone. But here's how it works today, in OpenLDAP 2.4.
A B-tree is a balanced tree; it stores data in its leaf nodes and bookkeeping data in its interior nodes (If you don't know what tree data structures look like in general, Google for some references, because that's getting far too elementary for the purposes of this discussion).
For decent performance, you need enough cache memory to contain all the nodes along the path from the root of the tree down to the particular data item you're accessing. That's enough cache for a single search. For the general case, you want enough cache to contain all the internal nodes in the database.
db_stat -d
will tell you how many internal pages are present in a database. You should check this number for both dn2id and id2entry.
Also note that id2entry always uses 16KB per "page", while dn2id uses whatever the underlying filesystem uses, typically 4 or 8KB. To avoid thrashing the, your cache must be at least as large as the number of internal pages in both the dn2id and id2entry databases, plus some extra space to accommodate the actual leaf data pages.
For example, in my OpenLDAP 2.4 test database, I have an input LDIF file that's about 360MB. With the back-hdb backend this creates a dn2id.bdb that's 68MB, and an id2entry that's 800MB. db_stat tells me that dn2id uses 4KB pages, has 433 internal pages, and 6378 leaf pages. The id2entry uses 16KB pages, has 52 internal pages, and 45912 leaf pages. In order to efficiently retrieve any single entry in this database, the cache should be at least
(433+1) * 4KB + (52+1) * 16KB in size: 1736KB + 848KB =~ 2.5MB.
This doesn't take into account other library overhead, so this is even lower than the barest minimum. The default cache size, when nothing is configured, is only 256KB.
This 2.5MB number also doesn't take indexing into account. Each indexed attribute uses another database file of its own, using a Hash structure.
Unlike the B-trees, where you only need to touch one data page to find an entry of interest, doing an index lookup generally touches multiple keys, and the point of a hash structure is that the keys are evenly distributed across the data space. That means there's no convenient compact subset of the database that you can keep in the cache to insure quick operation, you can pretty much expect references to be scattered across the whole thing. My strategy here would be to provide enough cache for at least 50% of all of the hash data.
(Number of hash buckets + number of overflow pages + number of duplicate pages) * page size / 2.
The objectClass index for my example database is 5.9MB and uses 3 hash buckets and 656 duplicate pages. So:
( 3 + 656 ) * 4KB / 2 =~ 1.3MB.
With only this index enabled, I'd figure at least a 4MB cache for this backend. (Of course you're using a single cache shared among all of the database files, so the cache pages will most likely get used for something other than what you accounted for, but this gives you a fighting chance.)
With this 4MB cache I can slapcat this entire database on my 1.3GHz PIII in 1 minute, 40 seconds. With the cache doubled to 8MB, it still takes the same 1:40s. Once you've got enough cache to fit the B-tree internal pages, increasing it further won't have any effect until the cache really is large enough to hold 100% of the data pages. I don't have enough free RAM to hold all the 800MB id2entry data, so 4MB is good enough.
With back-bdb and back-hdb you can use "db_stat -m" to check how well the database cache is performing.
For more information on db_stat: http://www.oracle.com/technology/documentation/berkeley-db/db/utility/db_stat.html
20.4.2. slapd(8) Entry Cache (cachesize)
The slapd(8) entry cache operates on decoded entries. The rationale - entries in the entry cache can be used directly, giving the fastest response. If an entry isn't in the entry cache but can be extracted from the BDB page cache, that will avoid an I/O but it will still require parsing, so this will be slower.
If the entry is in neither cache then BDB will have to flush some of its current cached pages and bring in the needed pages, resulting in a couple of expensive I/Os as well as parsing.
The most optimal value is of course, the entire number of entries in the database. However, most directory servers don't consistently serve out their entire database, so setting this to a lesser number that more closely matches the believed working set of data is sufficient. This is the second most important parameter for the DB.
As far as balancing the entry cache vs the BDB cache - parsed entries in memory are generally about twice as large as they are on disk.
As we have already mentioned, not having a proper database cache size will cause performance issues. These issues are not an indication of corruption occurring in the database. It is merely the fact that the cache is thrashing itself that causes performance/response time to slowdown.
20.4.3. IDL Cache (idlcachesize)
Each IDL holds the search results from a given query, so the IDL cache will end up holding the most frequently requested search results. For back-bdb, it is generally recommended to match the "cachesize" setting. For back-hdb, it is generally recommended to be 3x"cachesize".
{NOTE: The idlcachesize setting directly affects search performance}
20.4.4. slapd(8) Threads
slapd(8) can process requests via a configurable number of thread, which in turn affects the in/out rate of connections.
This value should generally be a function of the number of "real" cores on the system, for example on a server with 2 CPUs with one core each, set this to 8, or 4 threads per real core. This is a "read" maximized value. The more threads that are configured per core, the slower slapd(8) responds for "read" operations. On the flip side, it appears to handle write operations faster in a heavy write/low read scenario.
The upper bound for good read performance appears to be 16 threads (which also happens to be the default setting).