A multi-level scalable startup for parallel applications

High performance parallel machines with hundreds of thousands of processors and petascale performance are already in use, and even larger Exaflops scale computing systems which may have hundreds of millions of cores are planned. To run parallel applications on machines of such massive scale, one of the biggest challenges is the parallel startup process. This task involves two components: (1) parallel launching of appropriate processes on the given set of processors and (2) setting up communication channels to enable the processes to communicate with each other after process launching has completed. Most current startup mechanisms focus on either using special purpose daemons which waste system resources or using a startup manager which becomes a scalability bottleneck. In this paper, we investigate the design and scalability of a SMP-aware, multi-level startup scheme with batching of remote shell sessions, which provides a complete solution to startup of a parallel application and facilitates its management during execution. It monitors process health and can be used to support recovery from failures and provide scalable interaction with the application. We demonstrate the performance and scalability of this scheme by applying it to startup Charm++ applications. In particular, starting up a Charm++ program on 16,384 cores of Ranger (at TACC) with Ethernet as the underlying communication layer takes only 25 seconds and attains a speedup of over 400% compared to MPICH2 startup (using hydra as process manager) and over 800% compared to Open MPI startup on Ranger.