Live Webcast 15th Annual Charm++ Workshop

Complex systems, both natural and engineered, function well when they employ feedback control strategies. Complexity of upcoming parallel machines, combined with sophistication of next generation HPC applications, is creating a landscape where such feedback control systems are essential to attain the scientific and engineering breakthroughs we expect. The runtime systems of today need to evolve into smart, introspective and adaptive control systems. We will examine this idea in the context of two related systems: a single parallel job, and an entire parallel machine. These can be informally analyzed using vocabulary from control systems and discrete optimization, leading to a conclusion that it is necessary to create a rich set of control variables to facilitate effective control. Programming models will play a large role in enabling and empowering such adaptive control. To this end, this talk will also present an abstract programming paradigm called XMAPP, which captures essential properties a programming language must posses to create a specific, rich set of adaptive control variables.