Hybrid Runtime Systems

As parallelism continues to increase in ubiquity---from mobile devices and GPUs to datacenters and supercomputers---parallel runtime systems occupy an increasingly important role in the system software stack. The needs of parallel runtimes and the increasingly sophisticated languages and compilers they support do not line up with the services provided by general-purpose OSes. Furthermore, the semantics available to the runtime are lost at the system-call boundary in such OSes. Finally, because a runtime executes at user-level in such an environment, it cannot leverage hardware features that require kernel-mode privileges---a large portion of the functionality of the machine is lost to it. These limitations warp the design, implementation, functionality, and performance of parallel runtimes. I make the case for eliminating these compromises by transforming parallel runtimes into hybrid runtimes (HRTs), runtimes that run as kernels, and that enjoy full hardware access and control over abstractions to the machine. The primary claim of this dissertation is that the hybrid runtime model can provide significant benefits to parallel runtimes and the applications that run on top of them.;I demonstrate that it is feasible to create instantiations of the hybrid runtime model by doing so for four different parallel runtimes, including Legion, NESL, NDPC (a home-grown language), and Racket. These HRTs are enabled by a kernel framework called Nautilus, which is a primary software contribution of this dissertation. A runtime ported to Nautilus that acts as an HRT enjoys significant performance gains relative to its ROS counterpart. Nautilus enables these gains by providing fast, light-weight mechanisms for runtimes. For example, with Legion running a mini-app of importance to the HPC community, we saw speedups of up to forty percent. We saw further improvements by leveraging hardware (interrupt control) that is not available to a user-space runtime.;In order to bridge Nautilus with a "regular OS" (ROS) environment, I introduce a concept we developed called the hybrid virtual machine (HVM). Such bridged operation allows an HRT to leverage existing functionality within a ROS with low overheads. This simplifies the construction of HRTs.;In addition to Nautilus and the HVM, I introduce an event system called Nemo, which allows runtimes to leverage events both with a familiar interface and with mechanisms that are much closer to the hardware. Nemo enables event notification latencies that outperform Linux user-space by several orders of magnitude.;Finally, I introduce Multiverse, a system that implements a technique called automatic hybridization. This technique allows runtime developers to more quickly adopt the HRT model by starting with a working HRT system and incrementally moving functionality from a ROS to the HRT.