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

Re: 8 hours tests ends with inconsistent DB.





--On Friday, June 11, 2004 11:56 PM -0700 Trevor Warren <trevorwarren@yahoo.com> wrote:


 Have you bombarded your system with Zero think time
tests and 200-500+ users. Do you understand the amount
of stree the system is required to handle when
200-500+ users hit a system housing 1million records
with no think time ????.

 A 15-20 user test i am sure i will get openldap to
excel at. I want my feedback to help us all scale this
setup to Enterprise needs.

 BTW: A Zero think time test never can be imitated in
any real life situation since every user infuses a
large amount of think time into the action of using an
application.

 Also...a 200+ user test without Think Time is real
high load for any system. If it can withstand this
test as an 8 hour ordeal.....we have a REAL Stable
System here.

Trevor,

I think one thing to consider here is it is rather a bad design flaw to have a single server answering queries. We have 9 OpenLDAP servers in a pool that are all replica's of a 10th server that is the master. This way all incoming writes go to the master, which then spreads them out to the slaves via slurpd, and the slaves take the queries. 3 of the slaves are dedicated to mail delivery, and they each process around 550k connections a day, with about 3 times that many searches. Also, I find, at least in our case, it is generally applications and not users making queries. If you properly set up your environment, your servers are not being bombarded with writes and reads. The load tests I've performed with that type of setup still show a single server able to handle 30 hosts constantly pounding them, with a query rate of about 177 queries/second when the queries are bind+search+unbind. If the queries are bind+<multiple searches>+never unbind, the query rate can jump up to 1k/queries a second.

--Quanah

--
Quanah Gibson-Mount
Principal Software Developer
ITSS/Shared Services
Stanford University
GnuPG Public Key: http://www.stanford.edu/~quanah/pgp.html