To duplicate the issue:
1) Create a four member cluster using the following configuration: (two member clusters exhibit the problem as well, just not as exaggerated)

TCP(start_port=7801):
TCPPING(initial_hosts=<ip addresses go here>;port_range=3;timeout=3500;num_initial_members=3;up_thread=true;down_thread=true):
MERGE2(min_interval=5000;max_interval=10000):
FD(shun=true;timeout=2500;max_tries=5;up_thread=true;down_thread=true):
VERIFY_SUSPECT(timeout=2000;down_thread=false;up_thread=false):
pbcast.NAKACK(down_thread=true;up_thread=true;gc_lag=100;retransmit_timeout=3000):
pbcast.STABLE(desired_avg_gossip=20000;down_thread=false;up_thread=false):
pbcast.GMS(join_timeout=5000;join_retry_timeout=3500;shun=true;print_local_addr=true;down_thread=true;up_thread=true)

2) I was running JGroups in a Tomcat servlet application. Start up the cluster. To determine the number of threads on Linux I executed the following commands:

ps -ef | grep tomcat
echo "" > catalina.out
kill -QUIT <pid from ps command above>
grep ".Sender [" catalina.out | wc -l

You get the process id of Tomcat using the ps command. Then clear the content of the catalina.out file. The kill command causes the threads to be printed into the catalina.out file. Then the grep searches for and counts all of the "ConnectionTable.Connnection.Sender" threads that are currently active on the system.

3) Pick one of the cluster member boxes and pull the network cable out of the box such that all communication with the other three members is terminated.

4) After one or two minutes, replace the network cable.

5) Repeat the steps to determine the number of threads currently active on the system.

6) Repeat steps 3 through 5, each time watching the number of threads. Each iteration will cause more and more threads to be orphaned on the system. It seems to grow exponentially, after about 4 iterations we have around 300-400 Sender threads. The Receiver threads will be orphaned also in similar numbers.

After investigating the issue, I came up with the following "fix" which cleared the problem up.

In the file ConnectionTable.java there is a method called retainAll(). It appears that this method is called by the TCP protocol when a view change occurs. This method removes Connnections from the "Connection Pool" (member variable conns) but does not destroy them. We initially thought the reaper thread may clean them up, but since the Connection objects are actually removed from the Connection Pool, the reaper does not help the situation. As we watched our connections we noticed that the Connections orphaned by this routine were the ones filling up the system's set of threads. So, we added code to call destroy() on all of the Connection objects that retainAll() removes from the Connection Pool. The "diff" is provided below. Note that we did our change in the JGroups 2.3 SP1 file ConnectionTable.java. Scott Marlow did this diff for me, the same change, but applied to the BasicConnectionTable from the 2.4 source set.

Index: BasicConnectionTable.java
===================================================================
RCS file: /cvsroot/javagroups/JGroups/src/org/jgroups/blocks/BasicConnectionTable.java,v
retrieving revision 1.8
diff -r1.8 BasicConnectionTable.java
22,26c22
< import java.util.Map;
< import java.util.Iterator;
< import java.util.HashMap;
< import java.util.Vector;
< import java.util.Collection;
—
> import java.util.*;
263c259,289
< conns.keySet().retainAll(c);
—
> // conns.keySet().retainAll(c);
> ArrayList alConnsToDestroy = new ArrayList();
> synchronized(conns)
> {
> HashMap copy=new HashMap(conns);
> conns.keySet().retainAll(c);
> Set ks = copy.keySet();
> Iterator iter = ks.iterator();
> while (iter.hasNext())
> {
> Object oKey = iter.next();
> if (null == conns.get(oKey))
> { // This connection NOT in the resultant connection set
> Connection conn = (Connection)copy.get(oKey);
> if (null != conn)
>

{ // Destroy this connection > alConnsToDestroy.add(conn); > }

> }
> }
> }
> // All of the connections that were not retained must be destroyed
> // so that their resources are cleaned up.
> for (int a=0; a<alConnsToDestroy.size(); a++)
>

{ > Connection conn = (Connection)alConnsToDestroy.get(a); > if(log.isTraceEnabled()) > log.trace("Destroy this orphaned connection: " + conn); > conn.destroy(); > }

>