The new topology of a cluster system involves a "Cluster Manager" process on each node that controls communications with other "Cluster Managers" on other nodes, and with it's own collection of "slave" nodes.

There is no longer any need to manually start slaves on each node - that is handled by the Cluster Manager and is configured in the initial Pilot script. The Cluster Manager on the node the "master" node is executing on will contact each of the required Cluster Manager nodes and start the required number of nodes on each.

Error detection has been improved - if any nodes fail, they trigger a failure for all nodes in the run and the entire run will cleanly exit.

Slaves are now simply allocated an ID, and no specification of "process name" and number is required - any node from the pool that can perform the task is used. Intrastep tasks simply request the number of nodes desired. This ensures that multiple requests for the same node are now no longer possible.

A separate wait function has been implemented for Pilot for multi-step operations - "BARRIER". WAIT4FILES still exists, but as it only operates on files, it no longer functions as a wait for cluster operations to complete.