Run-time Adaptation for Multi-Core Execution

Contention over shared resources is one of many variables affecting parallel performance on modern multi-core processors. The issue is of particular interest because it the primary factor limiting fully independent threads/processes which would otherwise showcase linear speedup with each additional core. Furthermore, addressing the problem is a non-trivial task. Shared-resource contention, while fundamentally a product of architectural design decisions, is significantly influenced by the running workload as well as code-generation decisions made at compile time. This paper investigates the effects of workload selection and code-generation on parallel performance and showcases that there is no single factor or characteristic that can be leveraged to provide a best static compilation or workload selection with respect to parallel performance. Addressing resource contention must be handled in an adaptive manner at runtime. To this end, this paper proposes a hybrid runtime system that uses hardware performance counters to influence runtime code selection. The system is evaluated using a subset of the SPEC CPU2006 benchmark suite and is demonstrated across variety of x86 multi-core architectures.