Achieving High Performance and High Productivity in Next Generational Parallel Programming Languages

Abstract

Processor design has turned toward parallelism and heterogeneity cores to achieve performance and energy efficiency. Developers find high-level languages attractive because they use abstraction to offer productivity and portability over hardware complexities. To achieve performance, some modern implementations of high-level languages use work-stealing scheduling for load balancing of dynamically created tasks. Work-stealing is a promising approach for effectively exploiting software parallelism on parallel hardware. A programmer who uses work-stealing explicitly identifies potential parallelism and the runtime then schedules work, keeping otherwise idle hardware busy while relieving overloaded hardware of its burden.

However, work-stealing comes with substantial overheads. These overheads arise as a necessary side effect of the implementation and hamper parallel performance. In addition to runtime-imposed overheads, there is a substantial cognitive load associated with ensuring that parallel code is data-race free. This dissertation explores the overheads associated with achieving high performance parallelism in modern high-level languages.

My thesis is that, by exploiting existing underlying mechanisms of managed runtimes; and by extending existing language design, high-level languages will be able to deliver productivity and parallel performance at the levels necessary for widespread uptake.

The key contributions of my thesis are: 1) a detailed analysis of the key sources of overhead associated with a work-stealing runtime, namely sequential and dynamic overheads; 2) novel techniques to reduce these overheads that use rich features of managed runtimes such as the yieldpoint mechanism, on-stack replacement, dynamic code-patching, exception handling support, and return barriers; 3) comprehensive analysis of the resulting benefits, which demonstrate that work-stealing overheads can be significantly reduced, leading to substantial performance improvements; and 4) a small set of language extensions that achieve both high performance and high productivity with minimal programmer effort.

A managed runtime forms the backbone of any modern implementation of a high-level language. Managed runtimes enjoy the benefits of a long history of research and their implementations are highly optimized. My thesis demonstrates that converging these highly optimized features together with the expressiveness of high-level languages, gives further hope for achieving high performance and high productivity on modern parallel hardware