|
Hi,
I just manged to setup a nice 5 system cluster.
And i have a couple of questions.
I have two machines setup as directors, with hearthbeat for failover.
Now i have 3 nodes, and the local node feature on the current active
director.
This works.
From the client i can connect to each of the RIP's and get their apache
reponse/ssh into them.
From the client i can connect with ssh to the VIP (gives me the active
director)
From the client i can browse to the VIP gives me semi random, (altough
usually the same) realserver.
When i do a AB (apachebench) from the client to each of the rip's i get
aprox 560 pages/sec.
When i do a AB from the client to the VIP, i get almost the same. (560
pages/sec)
When i ab the VIP, i see the load rising on all servers. (and see apache
processes in top)
When i use persistent connections i get 3 times more pages/sec.
I have a troughput of 460Kbytes/sec. So its not the lan.
I have 50% cpu useage on the client, so it could do more.
I changes the amount of concurrent connections from 5 to 200, no change
(altough the load gets higer on all of the machines)
Even a single RIP, with 200 connections and 50000 requests and o load of
5 gets the 560 pages/sec done.
Does anyone know where i can find my bottleneck. (all servers are 80%
idle when i AB the VIP)
The machines are Compq dl360's with 265MB ram (100MB ram free) and 10k
rpm scsi disks. each server has one 1ghz p3.
Now my seccond question.
Is it possible to have the failover director be a realserver as long as
it is not a director?
So it would have the VIP with noarp when it is just a failover, and thus
it is a realserver.
And then when the master director goes down, instead of adding the VIP,
it should just disable noarp for the VIP and thus become the director.
(it could offcourse then use local node feature do still be a realserver)
Thank you
|
