java slave client dies on some nodes

On some of my build slaves, the java slave runs just fine, on others it dies very shortly (within a second or two) after starting. Hudson says about it:

Connection was broken

java.io.IOException: Unexpected termination of the channel
at hudson.remoting.Channel$ReaderThread.run(Channel.java:881)
Caused by: java.io.EOFException
at java.io.ObjectInputStream$BlockDataInputStream.peekByte(ObjectInputStream.java:2570)
at java.io.ObjectInputStream.readObject0(ObjectInputStream.java:1314)
at java.io.ObjectInputStream.readObject(ObjectInputStream.java:368)
at hudson.remoting.Channel$ReaderThread.run(Channel.java:875)

The log for it says:

Connecting to 10.8.3.81
Verifying that java exists
java full version "JRE 1.6.0 IBM Linux build pxa6460sr7-20091215_02 (SR7)"
Copying slave.jar
Launching slave agent
<===[HUDSON REMOTING CAPACITY]===>Slave.jar version: 1.372
This is a Unix slave
Copied maven-agent.jar
Copied maven-interceptor.jar
Copied maven2.1-interceptor.jar
I can start the slave manually and it will stay running:
[busy icon spinning]

It will stay running if I start it manually:

1. ssh -i ~hudson/.euca/hudson.priv root@10.8.3.81 java -jar /tmp/slave.jar
   <===[HUDSON REMOTING CAPACITY]===>[encyrpted/encoded data removed for privacy]
   [stays running, so in another terminal:]
2. ssh -i ~hudson/.euca/hudson.priv root@10.8.3.81 ps -ef | grep java
   root 2766 2761 3 12:53 ? 00:00:01 java -jar /tmp/slave.jar

Any ideas how to go about figuring out why it's dying so quickly on only some of my build slaves?