Breaking away from the OS shadow: A program execution model aware thread virtual machine for multicore architectures

The Cyclops-64 (C64) project questions fundamentally the suitability of conventional operating systems to achieve high performance. Domain-specific application experts who have participated in the conception of all aspects of the system software for the C64 supercomputer mandated that Linux was not adequate. Their previous experience with parallel applications that did not scale well for various reasons motivated us to develop a standalone Thread Virtual Machine (TVM) from scratch. Its implementation in the form of the TiNy Threads (TNT) library had the clear goal of allowing applications to achieve full resource utilization.;This dissertation is about the C64 system software in general, and the TNT library in particular. TNT replaces the conventional OS with a non-intrusive runtime system. Even though it is implemented as a user-level library, TNT manages the hardware resources directly. In addition, TNT provides a solid foundation for the development of advanced program execution models. However, for rapid prototyping of applications, TNT also provides a familiar Linux-like programming environment.;As evidence that the TNT model provides a good platform to experiment with innovative execution models, we developed MAGMA. Defined as a memory adaptive program execution model for multicore architectures, MAGMA uses percolation to migrate data that the user (programmer or compiler) identifies, to a level of the memory hierarchy local to the processing element before computation starts. MAGMA takes advantage of the large number of thread units in C64. MAGMA implements a multithreaded percolation engine that runs on a number of cores to maximize bandwidth utilization.