Changing KeepAlive to Off does not help, the issue occurs at this
point where obviously due to MaxConnectionsPerChild a Process
should become killed.

As observed, the problem is that (in the example stated) both
PIDs 11454 and 16555 do not become killed because of these
remaining connections, which is ok, since these could be
remaining downloads or something to a client.
But then, Apache does neither fork new processes to accept
new connections nor does it kill the both PIDs (which is as just
statet, ok) - the scoreboard is full then in that case and new
connections can not be accepted.

As I see a solution to this problem is that the main httpd process
should be 1. able to fork new processes 2. I see no reason why
the scoreboard is full then, 1+2 connections is 3, not 2x64=128,
in that given case 125 free slots are wasted since they are blocked
by the 2 remaining processes which show connections which fill up
the scoreboard, those are shown as "G"s in the scoreboard for both
PIDs - that should not be the case.

Since 2.4.2 (the first I've used) I did not understand why a
finishing/killing process shows all "G"s - however, that issue did
not occur in 2.2 when I used MPM event in experimental stage, but
in 2.2 I did not have that problem, the connections were counted as
usual. Probably because lots of the code was taken from MPM worker,
where the calculations were still correct, until fully adopted to 2.4.

By this theory, also MPM worker might experience this, however, due to
the different design MPM worker also may never get to that point to
show this behavior and error messages in the log.

So my conclusion is also a question: why are finishing processes showing
all "G"s in the scoreboard, and/or why are these finishing threads all
counted in the scoreboard (here: 2x64 instead of 1+2, causing 125 threads
to be wasted).

I would expect this issue to become much more important when most people
upgrade to 2.4, also I clearly would say this is a bug.

On Thu, Sep 6, 2012 at 11:21 PM, Sam <lennsen@chello.at> wrote:
Never, the server is not even under high load. There
are several thousand vhosts, it is about shared webhosting.

This is what server-status says:

PID    Connections     Threads    Async connections
total    accepting    busy    idle    writing    keep-alive    closing
21388    257    yes  44   20    0  171   44
11454    2      no    0    0    0    0    0
16555    1      no    0    0    0    0    0
Sum      260         44   20    0  171   44



How about those 171 in keep-alive? What are your KeepAlive and KeepAliveTimeout settings? Maybe try switching KeepAlive off? Whit only 192 threads and KeepAlive On you are imposing DoS on your self in case of long timeout.
 
That one PID 11454 has "2" connections, 16555 has "1", all other
values are 0, I guess it should become killed?  - but it does not,
it stays like this for a while (I don't know on what this "while" depends on).

At this point all 3 rows are filled with stuff, such as:

WWW_W_WWWWWW_WW_W_W_W_WW__WWW__WW__WWWWWW__WWW__W__W_RWWWWRWWWWW
GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG
GGGGGGGGGGGGGGGLGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG

But as you see, due to several _ this means that there are
not all 192 busy, in fact:

105 requests/sec - 2.5 MB/second - 24.8 kB/request
44 requests currently being processed, 20 idle workers

WHY can no others be used? With the statement above it means that
128 (which is 192-64) workers are unused because they should become
killed already, to spawn new ones, these are the ones in the 2. and
3. row with "G"s.

--

So the rest is occupied with "G". In Apache 2.2 I used MPM_event
as well, this obviously never occured because the both PIDs above
which are showing G only showed just dots in that case, now after
one PID should be finished, all slots become occupied with G, 
which obviously causes a full scoreboard.

I would consider this a bug, a serious one for my case, since it
often occurs (~1-3 times per hour) that the server stops accepting
new requests, just because of that full scoreboard, where the server
is neither busy, nor do I see what configuration directive should
be configured badly. Again:

KeepAlive                   On
ListenBacklog             4095
Timeout                     20
KeepAliveTimeout             8
MaxKeepAliveRequests     16192
MaxRanges                  200
MaxRangeOverlaps            20
MaxRangeReversals           20
GracefulShutdownTimeout      6
LimitInternalRecursion      10
LimitRequestFieldSize     4094
LimitRequestFields          40
LimitRequestLine          4094
LimitXMLRequestBody     786432

MaxRequestWorkers 192
ServerLimit         3
StartServers        2

MaxMemFree                2048
MinSpareThreads             64
MaxSpareThreads             64
ThreadLimit                128
ThreadsPerChild             64
MaxConnectionsPerChild   10240
AsyncRequestWorkerFactor    10

I assumed AsyncRequestWorkerFactor to have something to do 
with it, I raised it from 2 (default) to 10, no change.

By the way: after some time (in this case above both) PIDs
become killed finally, releasing 2x64 slots in the scoreboard,
the waring messages vanish and the server can respond again
well to requests. In the interest of finding a workaround until
this is fixed: what configuration directives control that timeout
or whatever, which cause these PIDs to become killed?


On Thu, Sep 6, 2012 at 7:06 PM, Sam <lennsen@chello.at> wrote:

Yes, I still see such messages even after upgrading to 2.4.3


On Sun, Aug 19, 2012 at 8:25 AM, Sam <lennsen@chello.at> wrote:
I upgraded to Apache 2.4 and use mod_event on a server
having average high load.

My issue is that I am getting pretty much often the error message
within the error log

AH00485: scoreboard is full, not at MaxRequestWorkers

What causes this? The server almost stops to handle new requests
when this occurs.

MaxRequestWorkers 192
ServerLimit         3
StartServers        2

MaxMemFree                2048
MinSpareThreads             64
MaxSpareThreads             64
ThreadLimit                 64
ThreadsPerChild             64
MaxConnectionsPerChild   10240
AsyncRequestWorkerFactor     2

So you never have more than 192 simultaneous requests at any given time ???