"GUMMALAM,MOHAN (HP-Cupertino,ex2)" wrote:
>
> -1 for beta
>
> Was able to build and run it on HP-UX on PA-RISC -- with problems.
>
> It still exhibits the same problem as was evident in 2.0.19 and later
> releases, there is only 1 process up and running, (apart from the cgid and
> the watchdog process). The configuration I used were as follows:
>
> StartServers 6
> MinSpareThreads 5
> MaxSpareThreads 50
> ThreadsPerChild 25
> MaxRequestsPerChild 0
>
> This behaviour on a running system would be as follows. At peak load, say a
> webserver is serving 130 simultaneous request => I have 6 worker processes
> running, a total of 150 threads. At some point later, the load (suddenly)
> falls down to say, 90 requests => 60 spare threads. The current logic would
> cause the p_i_s_m() to send a pod, to kill one process. Since all the
> processes are busy serving requests, the process which recd the pod will not
> die immediately. Hence, p_i_s_m() will continue sending pods. Eventually
> all the processes will die (and almost all of them simultaneously). And
> finally, there will be 0 processes running.
hmmm, a lot of this was broken by the removal of the
idle_thread_count_addition logic from p_i_s_m in revision 1.43 of
threaded.c on July 3, while I was on a sailboat. It is broken in the
worker MPM as well. If you can get mod_status to run, you will see "G"s
in the scoreboard when this hits.
This should fix threaded. Cliff & Paul, this may well take care of the
bad stuff you're seeing. I'll beat on it & commit fixes after supper.
Sorry I didn't figure it out sooner.
Index: server/mpm/threaded/threaded.c
===================================================================
RCS file: /home/cvs/httpd-2.0/server/mpm/threaded/threaded.c,v
retrieving revision 1.51
diff -u -d -b -r1.51 threaded.c
--- server/mpm/threaded/threaded.c 2001/07/28 23:39:07 1.51
+++ server/mpm/threaded/threaded.c 2001/07/30 22:58:31
@@ -1010,6 +1010,7 @@
*/
if (status <= SERVER_READY && status != SERVER_DEAD &&
ps->generation == ap_my_generation &&
+ !ps->quiescing &&
/* XXX the following shouldn't be necessary if we
clean up
* properly after seg faults, but we're not yet
GLA
*/
Greg
