In this paper we describe Buffered Coscheduling, a new approach to
design the system software of large scale parallel computers. A buffered
coscheduled system can tolerate inefficient programs, programs that
have communication and load imbalance, and can substantially increase
the resource utilization, by overlapping computation, communication
and I/O over several jobs. In addition to that, Buffered Coscheduling
creates a framework to easily implement fault-tolerance, arguably
the most important problem to solve to build usuable, large scale
parallel computers.