[Date Prev][Date Next] [Chronological] [Thread] [Top]

RE: make test got stuck at Testing for slave slapd...





--On Thursday, April 03, 2003 5:07 PM -0800 Howard Chu <hyc@highlandsun.com> wrote:

-----Original Message-----
From: owner-openldap-software@OpenLDAP.org
[mailto:owner-openldap-software@OpenLDAP.org]On Behalf Of Quanah
Gibson-Mount

--On Thursday, April 03, 2003 6:41 PM -0500 Igor Brezac
<igor@ipass.net>
wrote:

> Hmm..  I do not find this to be correct.  I just ran a DirectoryMark
> benchmark (100,000 entry dataset, 20 simultaneous clients, 100,000
> different search samples) on Solaris 9 - fairly recent patches, (1 CPU
> 440Mhz Ultrasparc IIi, 512MB RAM)  against Openldap 2.1.16 (tpool
> patch, latest CVS SASL, Sleepycat 4.1.25).  I am able to get over 800
> searches per second and with more tweaking I can probably do better.
>
> This number is probably not indicative measure of a real world
> application, but it does not indicate performance problems with Solaris
9.
> I tested Solaris 8 awhile back and it performed roughly the same.

Igor,

I unfortunately have never seen the DirectoryMark's tool results have any
resemblance to reality.  You gave me the same sort of answer when I was
seeing massive problems with Solaris 8, with a rate  of 4-6
queries/second. After hiring Howard Chu to work on the system, in which
he found the exact same results, an extensive set of patches were put in
to OpenLDAP and related pieces of software (cyrus-sasl, heimdal, BDB,
etc) before we got our current performance of 66/queries a second.  I
have absolutely no faith in your results.

One principal difference between these two scenarios is that Stanford uses GSSAPI for all of their LDAP sessions. One major issue that we faced there was with thread-safety issues (and memory leaks) in the Kerberos libraries underlying the SASL/GSSAPI layer. It might be interesting to see the numbers DirectoryMark yields when all of its clients are also using mutual authentication and encryption using either GSSAPI or TLS. The two scenarios also differ in the volume of actual directory data, as well as ACLs controlling access to that data. So, as with any benchmark, the results of one scenario cannot be meaningfully compared to any other. DirectoryMark has some value, for comparing its results to itself given different server configurations, but that's all one should expect from it.

The fact that DirectoryMark yields comparable numbers for Solaris 8 and
Solaris 9 is mildly interesting, but as we learned in great detail at
Stanford, the addition of  other major components (SASL, Kerberos) to the
picture drastically alters the runtime environment and its behavior. Your
numbers won't tell the whole story until you take this into account.

Also of note, we found only slightly better performance numbers when allowing anonymous binds to our servers. The last anonymous run I did got a performance of 85 queries/second or so, which is slightly better, but not greatly better, than the performance we see when using SASL/GSSAPI. And still nowhere near 800-1000 queries/second. And again, the underlying issue of difference I see between Solaris 8 and 9 is its thread implementation, which I still believe affects its ability to perform. I can examine using Net::LDAPapi to create a persistent connection and gauge queries/second in an authenticated environment.


--Quanah

--
Quanah Gibson-Mount
Senior Systems Administrator
ITSS/TSS/Computing Systems
Stanford University
GnuPG Public Key: http://www.stanford.edu/~quanah/pgp.html