RE: Slapd frontend performance issues

[Sam Drake <drake@timesten.com>]

FYI, here is a short description of the changes I made.  I'll package up the
changes asap, but it may take a couple of days.

The performance numbers quoted in this report were seen at my location with
a 100,000 object database ... the slower numbers I mentioned earlier were
reported by a customer with a 1,000,000 object database.

I also can't explain the very poor performance I saw with OpenLDAP and LDBM
with a 100,000 object database.

...Sam Drake / TimesTen Performance Software

----------

Work Performed

OpenLDAP 2.0.9, including back-sql, was built successfully on Solaris
8 using gcc.  The LDAP server itself, slapd, passed all tests bundled
with OpenLDAP.  OpenLDAP was built using Sleepycat LDBM release 3.1.17
as the "native" storage manager.

The experimental back-sql facility in slapd was also built
successfully.  It was built using Oracle release 8.1.7 and the Oracle
ODBC driver and ODBC Driver Manager from Merant.  Rudimentary testing
was performed with the data and examples provided with back-sql, and
back-sql was found to be functional.

Slapd and back-sql were then tested with TimesTen, using TimesTen
4.1.1.  Back-sql was not immediately functional with TimesTen due to a
number of SQL limitations in the TimesTen product.

Functional issues encountered were:

1. Back-sql issued SELECT statements including the construct,
   "UPPER(?)".  While TimesTen supports UPPER, it does not support the
   use of parameters as input to builtin functions.  Back-sql was
   modified to convert the parameter to upper case prior to giving it
   to the underlying database ... a change that is appropriate for all
   databases.

2. Back-sql issued SELECT statements using the SQL CONCAT function.
   TimesTen does not support this function.  Back-sql was modified to
   concatentate the necessary strings itself (in "C" code) prior to
   passing the parameters to SQL.  This change is also appropriate for
   all databases, not just TimesTen.

Once these two issues were resolved, back-sql could successfully
process LDAP searches using the sample data and examples provided with
back-sql.

While performance was not measured at this point, numerous serious
performance problems were observed with the back-sql code and the
generated SQL.  In particular:

1. In the process of implementing an LDAP search, back-sql will
   generate and execute a SQL query for all object classes stored in
   back-sql.  During the source of generating each SQL query, it is
   common for back-sql to determine that a particular object class can
   not possibly have any members satisfying the search.  For example,
   this can occur if the query searches an attribute of the LDAP
   object that does not exist in the SQL schema.  In this case,
   back-sql would generate and issue the SQL query anyway, including a
   clause such as "WHERE 1=0" in the generated SELECT.  The overhead
   of parsing, optimizing and executing the query is non-trivial, and
   the answer (the empty set) is known in advance. Solution: Back-sql
   was modified to stop executing a SQL query when it can be
   predetermined that the query will return no rows.

2. Searches in LDAP are fundamentally case-insensitive ("abc" is equal
   to "aBc").  However, in SQL this is not normally the case.
   Back-sql thus generated SQL SELECT statements including clauses of
   the form, "WHERE UPPER(attribute) = 'JOE'".  Even if an index is
   defined on the attribute in the relational database, the index can
   not be used to satisfy the query, as the index is case sensitive.
   The relational database then is forced to scan all rows in the
   table in order to satisfy the query ... an expensive and
   non-scalable proposition.  Solution: Back-sql was modified to allow
   the schema designer to add additional "upper cased" columns to the
   SQL schema.  These columns, if present, contain an upper cased
   version of the "standard" field, and will be used preferentially
   for searching.  Such columns can be provided for all searchable
   columns, some columns, or no columns.  An application using
   database "triggers" or similar mechanisms can automatically
   maintain these upper cased columns when the standard column is
   changed.

3. In order to implement the hierarchical nature of LDAP object
   hierarchies, OpenLDAP uses suffix searches in SQL.  For example, to
   find all objects in the subtree "o=TimesTen,c=us", a SQL SELECT
   statement of the form, "WHERE UPPER(dn) LIKE '%O=TIMESTEN,C=US'"
   would be employed.  Aside from the UPPER issue discussed above, a
   second performance problem in this query is the use of suffix
   search.  In TimesTen (and most relational databases), indexes can
   be used to optimize exact-match searches and prefix searches.
   However, suffix searches must be performed by scanning every row in
   the table ... an expensive and non-scalable proposition.  Solution:
   Back-sql was modified to optionally add a new "dn_ru" column to the
   ldap_entries table.  This additional column, if present, contains a
   byte-reversed and upper cased version of the DN.  This allows
   back-sql to generate indexable prefix searches.  This column is
   also easily maintained automatically through the use of triggers.

Results

A simple database schema was generated holding the LDAP objects and
attributes specified by our customer.  An application was written to
generate test databases.  Both TimesTen and Oracle 8.1.7 were
populated with 100,000 entry databases.

Load Times 

Using "slapadd" followed by "slapindex", loading and indexing 100,000
entries in an LDBM database ran for 19 minutes 10 seconds.

Using a C++ application that used ODBC, loading 100,000 entries into
a disk based RDBMS took 17 minutes 53 seconds.

Using a C++ application that used ODBC, loading 100,000 entries into
TimesTen took 1 minute 40 seconds.

Search Times

The command, "timex timesearch.sh '(cn=fname210100*)'" was used to
test search times.  This command issues the same LDAP search 4000
times over a single LDAP connection.  Both the client and server
(slapd) were run on the same machine.

With TimesTen as the database, 4000 queries took 14.93 seconds, for a
rate of 267.9 per second.

With a disk based RDBMS as the database, 4000 queries took 77.79 seconds,
for a
rate of 51.42 per second.

With LDBM as the database, 1 query takes 76 seconds, or 0.076 per
second.  Something is clearly broken.