ScELA: Scalable and Extensible Launching Architecture for Clusters

As cluster sizes head into tens of thousands, current job launch mechanisms do not scale as they are limited by resource constraints as well as performance bottlenecks. The job launch process includes two phases-spawning of processes on processors and information exchange between processes for job intialization. Implementations of various programming models follow distinct protocols for the information exchange phase. We present the design of a sclable, extensible and high-performance job launch architecture for very large scale parallel computing. We present implementations of this architecture which achieve a speedup of more than 7--% in launching a simple Hello World MPI application on 10,240 processor cores and also scale to more than 3 times the number of processor cores compared to prior solutions.