[Mono-list] Too many heap sections: IncreaseMAXHINCRorMAX_HEAP_SECTS

[Jonathan Gagnon]

I changed my code to serialize smaller chunks of data, but I still seem to
have some kind of leak that happens only in mono, but I haven't been able to
reproduce it in a small test case yet.  It seems like the memory I allocate
is not always returned to the OS and that after a while, I run out of
memory.
 
Is suspect this is a bug inside the GC like you said.  Should this bug be a
high priority since it seems to mean that mono doesn't handle heavy memory
allocations (unless there is something wrong with my code that happens to
work well with .NET)?
 
I noticed that there is a new compacting GC under development.  Would it be
easy for me to test my app with this GC to see if it fixes the problem?
 
Thanks,
 
Jonathan
 
  _____  

De : Alan McGovern [mailto:alan.mcgovern at gmail.com] 
Envoyé : Thursday, May 10, 2007 6:45 PM
À : Jonathan Gagnon
Cc : Mono-list at lists.ximian.com
Objet : Re: [Mono-list] Too many heap sections:
IncreaseMAXHINCRorMAX_HEAP_SECTS


Ok.

So the problem is this (as far as i can make out). You're fragmenting the
heap and running out of "free" memory or the GC is just choking on the
amount of data you're spitting out. This is caused by two things. 
1) The memorystream increasing in size
2) A hashtable internal to mono constantly increasing in size while
serialisation is taking place.

If you run your testcase with int num = 9000; instead of int num = 15000; it
works fine. So the best advice i can offer is to serialise your data in
smaller chunks for the moment. 

Also, i never managed to get an OOM exception when running on MacOS, but the
program did seem to crash/hang.

Alan.




On 5/10/07, Jonathan Gagnon <jonathan.gagnon at croesus.com> wrote: 

I simplified the test a little bit.  I also tried serializing to a
FileStream instead of a MemoryStream and I got the same result.
 
I ran the test on Mono Windows and it runs for a while before exiting with
an OutOfMemoryException.
 
Jonathan

  _____  

De : mono-list-bounces at lists.ximian.com
[mailto:mono-list-bounces at lists.ximian.com] De la part de Jonathan Gagnon
Envoyé : Thursday, May 10, 2007 2:53 PM
À : 'Alan McGovern'; Mono-list at lists.ximian.com
Objet : Re: [Mono-list] Too many heap sections:
IncreaseMAXHINCRorMAX_HEAP_SECTS



Here is a little test that reproduces the problem.  I thought that
initializing the memory stream to a size bigger than the entire list would
fix the problem but it only makes it happen less quickly in some cases.
 
If you play with the numbers in my little test to reduce the size of
allocated memory, you will notice that it takes longer to run out of memory
but it still happens after a while.  The way it behaves, it really looks
like a leak since I have a loop that does the same thing at every run and I
would expect the memory allocator to be able to reuse the same memory
instead of growing the heap.
 
Note that I compiled the test with VS 2005.  I don't know if I could
reproduce the bug using the mono compiler.
 
Jonathan


  _____  

De : mono-list-bounces at lists.ximian.com
[mailto:mono-list-bounces at lists.ximian.com] De la part de Alan McGovern
Envoyé : Thursday, May 10, 2007 11:45 AM
À : Mono-list at lists.ximian.com
Objet : [Mono-list] Too many heap sections: Increase
MAXHINCRorMAX_HEAP_SECTS



>Also, as a test, could you initialise the memory stream to roughly the size
required to store the entire list<T> and see if it works then.




 
That works fine if I do it that way.  But my problem is that I can't really
know in advance how much memory the serialization will use, so it's not
really a viable solution.  I'm thinking of trying to split up my list into
smaller chunks to see it this could fix the problem by avoiding the large
object heap, if there is such a heap in mono.


Sounds like your problem is due to heap fragmentation. The only solution is
to use a best-guess for the approximate size of the memory stream and
initialise the memorystream to that to start off with. For example if the
average size of your class is 68 bytes, then initialise the memorystream to
array.Length * 68. Or some such thing.

Still, a testcase may prove useful.

Alan.



_______________________________________________
Mono-list maillist  -  Mono-list at lists.ximian.com
http://lists.ximian.com/mailman/listinfo/mono-list





-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-list/attachments/20070511/d622988e/attachment-0001.html