Once we had decided to go with OpenLDAP, membership on the openldap-software list was used to relay the issues we were having to see if anyone else was experiencing the same problems.
A consistent result of this was email from a developer at Symas, who suggested that we hire them to see if they could resolve the problems we were experiencing. We decided to go with this suggestion, and in November of 2002, we hired Symas to investigate the issues we were seeing.
This led to a flurry of work on the part of Symas which resulted in some major changes :
- A DB_CONFIG file for BDB (increased database stability)
- Work on setting idletimeout value in slapd.conf
- Evaluation of an appropriate threadlevel setting
- Evaluation of an appropriate concurrency setting
- Evaluation of different settings to IDL cache
- Evaluation of different settings of the thread stack size
- Using Heimdal Krb5 instead of MIT krb5
Symas also contributed some patches back into Heimdal and Cyrus-SASL that addressed threading issues that were the cause of several of our problems.
At the end of December 2002, we finally had a system that not only met our performance requirements, but greatly exceeded them (An average of 64 queries/second per server). On top of that, we were now able to run the entire directory load at Stanford on a single server, if necessary. This exceptional performance has only improved over time (The average is now 165 queries/second per server), and even with a massive growth in clients, we continue to be able to run the entire load on a single server when needed.
A detailed explanation of what our setup is at this time, and how it works, can be found at: OpenLDAP configuration information
