Which timeout affects the master-slave connection?

We share our PROOF cluster with a number of other groups. In particular, the astro-physicists (who use iPython, which is very cool, for parallel work). The problem is every now and then their jobs get so memory hungry that they cause the Linux machine to effectively crash. The problem is that when we then connect to the PROOF cluster, the slaves on that machine never show up - and so the TProof::Open takes forever (minutes under some circumstances, it feels like).

The xrd config file has lots of timeouts. Which is the one we can use to shorten this?