Re: [Wrapper-user] Wrapper restarts GC'ing JVM - why?

[Leif M. <le...@ta...>]

Christian,
    The wrapper.cpu.timeout is used to detect when the entire machine 
freezes up.  Not
really the problem that you are having.  Leave it at the default.   This 
timeout is so either
process can handle a loss of CPU for its own internal timeouts.  In this 
case the Wrapper
process is timing out waiting for the Java process.

    In your case, the Wrapper is waiting for the specified 2 minutes, 
but the Java side GC
appears to be taking even longer to complete.   After two minutes, the 
Java process is
still effectively hung until the GC completes, so the Wrapper is doing 
as it is told and
assuming that since it has not received a ping response, the JVM is hung.

    Can you try extending the timeouts to a longer value?  Try 5 or 10 
minutes.  Make sure
to give it more time than your longest possible GC.  That should fix the 
problem.

    Let me know how it works.  If it does not fix the problem, please 
turn on the debug
output and post that.  It will give me more to work with.

Cheers,
Leif

Christian Beedgen wrote:

>Hi. While overall i am very pleased with the functionality and stability
>of the Wrapper, i just cannot seem to figure out one problem. I went
>through the docs and the mailing list archives, yet no idea. So here
>goes:
>
>The basic problem is that the Wrapper will restart my application if it
>takes too long to garbage collect (this is as per observation). Consider
>the following Wrapper log:
>
>INFO   | jvm 1    | 2002/10/14 13:43:49 | Memory Status: 112 MB Used,
>127 MB Total
>INFO   | jvm 1    | 2002/10/14 13:43:56 | [Full GC
>ERROR  | wrapper  | 2002/10/14 13:46:02 | JVM appears hung: Timed out
>waiting for signal from JVM.
>ERROR  | wrapper  | 2002/10/14 13:46:02 | Java Virtual Machine did not
>exit on request, terminated
>STATUS | wrapper  | 2002/10/14 13:46:08 | Launching a JVM...
>
>As can be seen, i am hitting the heap limit, at which point a full GC
>kicks in (JVM running in verbose mode). The machine the JVM is running
>on is short on main memory and hits the page file at this point,
>swapping for about two minutes (see timestamps) - finally, the Wrapper
>decides it's dead and restarts it.
>
>While i don't want to argue the fact the 2 minute GCs are pretty bad ;)
>- i still think the Wrapper should not have restarted the JVM. I have
>gone through most of the docs and tuned some of the advanced parameters,
>namely:
>
>wrapper.ping.timeout=120
>
>>From my understanding, this will make the Wrapper wait for 120 seconds
>if no pings get answered from the JVM, then decide it's dead. From the
>above log, it looks like this timeout is working (death comes after some
>2 minutes). Now i did look into the 
>
>wrapper.cpu.timeout=10
>
>parameter and left it at the default. my understanding is, that once the
>wrapper detects that either itself or the JVM is stalled, then the ping
>timeout is extended. from the docs, it looks like this functionality is
>there to address the problem i am describing above. however, in my case,
>it looks like the decision to restart is made before the stalled process
>detection kicks in - is this a bug or the expected behaviour? 
>
>I'd appreciate any help in figuring this out. Oh, and i've seen this on
>(underpowerd) Windows and Solaris boxen.
>
>
>thanks,
>
>
>chr. 
>  
>