Hi all,
I have this situation, an hardware loadbalancer configured with sticky
session based
on souce IP, two httpd servers behind it doing HTTPS termination with
mod_jk doing
load balancing vs four tomcats, i'll try to do my best in ascii art,
hopping my mailer
does the right thing:
----------------- {internet}---------------------
|
v
[HW LB with source IP stickyness]
| |
| |
[httpd1] [httpd2]
[mod_jk] [mod_jk]
| |
======================================
| | | |
[TC1] [TC2] [TC3] [TC4]
Each apache httpd talks with every tomcat.
mod_jk is configured with load balancing by business, sticky sessions, only
one try for ajp worker (so if it's busy i dont add more stuff to the
poor worker)
and only 1 retry for all load balanced workers, in short, here's the
interesting
part of my conf:
worker.worker_lb.balance_workers = w1, w2, w3, w4
sticky session on, sticky session force = off ecc
[...]
Now, even with this settings, on peak times i have an amplification
effect, all
tomcats are busy, mod_jk reaches atimeout and retries again and again
... in a
death spiral for the poor cats.
This behaviour brings me two problems:
1) Even if apache httpd / mod_jk timeouts (read_timeout) from his side
this does
not stop the tomcat thread proccessing that request. It goes sometimes
on and on
but after all finishes the request.
This seems a resource waste, is there a way to let tomcat know that
mod_jk droped
the connection and stop proccessingthe dropped request? I tried to find
a way to
get this behaviour without success, so any suggestions are welcome.
2) I noticed that the default behaviour of load balanced workers and AJP
workers is
as follows:
a) ajp worker reaches retry_timeout, at this point retries again to
the same worker
once, thus amplifying load. This onewas easy to solve, just add a
"worker.w1.retries = 1"
and this does not happen anymore
b) Load balancer worker goes to the next ajp worker.. till the last
one, than does
another round of requests:
lb -> w1 -> timeout
[...]
lb-> w4 -> timeout
LB try 2
lb -> w1 -> timeout
[...]
request fails.
The second round to the workers IMHO is wasteful, but easy solved
with:
"worker.worker_lb.retries = 1"
So at this point for every request all the busy tomcats get only one
request. I tryed finding a way to tell
the LB worker to try just once, only another worker for request but i
didnt find an option for doing this.
Well, as i had an each to scratch i made a patch that adds an option to
a LB worker telling how much times
to retry before giving up, i'm glad to post it if somebody is
interested. Not sure it's perfect, but works
for me.
Now, my questions:
1) Am i doing something wrong?
2) Is there a way to make a tomcat thread stop proccessing when AJP
timeouts?
Thanks in advance,
Frederik
---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org
For additional commands, e-mail: users-h...@tomcat.apache.org
