Linux Large Datagram Problem

[kzablonski at ll.mit.edu (Kevin Zablonski)]

I am having a rather strange problem when it comes to UDP datagrams in 
linux.  I have found that when running software on linux that listens 
for multicast traffic on a given address, linux is having trouble 
receiving large UDP datagrams.  I know what everyone is thinking, I must 
be loosing packet fragments along the way due to maybe a busy network, 
but this is not the case.  When I run ethereal on the machine, I can see 
that all of the fragments are getting to the machine, but to the 
application.  It is like they are't even there.  To make things worse, 
when running the same software on windows(EEEK), everything works fine.  
To test out the problem, I have written two very basic java classes.  
One class sends out generic packets of a given size and the other class 
simply listens over the network and notifies you once a packet has been 
received.  I then ran the receiving class on several versions of linux 
to determine the size boundary of each system.  By slowly adjusting the 
size of the packets sent I have determined that Redhat 9 has a boundary 
of 41,433 bytes.  If you were to send data in the size of 41,432 bytes, 
everything is fine, but not 41,433.  The boundary information for each 
other version of linux can be found below.

Now to me, I would assume that each OS would be able to handle packet 
sizes up until the industry standard 65k.  I am puzzled as to why this 
would be happening.  It looks as if Fedora Core 2 is no longer having 
this issue and Slackware 10.1 also does not experience this problem.  I 
will eventually need to be able to receive packets of up to 57000 kb, 
but I would like to avoid running fedora core 2 if I can avoid it.

Has anyone seen this before or possibly know what I could do to fix it?  
Do I need to adjust an OS setting?  Any help or ideas would be greatly 
appreciated.  Thank You.