Using program behaviour to exploit heterogeneous multi-core processors

QA75 Electronic computers. Computer science;QA76 Computer software

McIlroy, Ross ; Sventek, Joseph

University:University of Glasgow

Department:School of Computing Science

关键词: Heterogeneous multi-core architectures, Virtual machines, Runtime Systems, Code annotations, Runtime monitoring, Non-uniform memory access, Java, multi-core programming;   

Others  :  http://theses.gla.ac.uk/1755/1/2010McIlroyPhD.pdf

来源: University of Glasgow

【 摘 要 】

Multi-core CPU architectures have become prevalent in recent years. A number of multi-core CPUs consist of not only multiple processing cores, but multiple different types of processing cores, each with different capabilities and specialisations. These heterogeneous multi-core architectures (HMAs) can deliver exceptional performance; however, they are notoriously difficult to program effectively.This dissertation investigates the feasibility of ameliorating many of the difficulties encountered in application development on HMA processors, by employing a behaviour aware runtime system. This runtime system provides applications with the illusion of executing on a homogeneous architecture, by presenting a homogeneous virtual machine interface. The runtime system uses knowledge of a program's execution behaviour, gained through explicit code annotations, static analysis or runtime monitoring, to inform its resource allocation and scheduling decisions, such that the application makes best use of the HMA's heterogeneous processing cores. The goal of this runtime system is to enable non-specialist application developers to write applications that can exploit an HMA, without the developer requiring in-depth knowledge of the HMA's design. This dissertation describes the development of a Java runtime system, called Hera-JVM, aimed at investigating this premise. Hera-JVM supports the execution of unmodified Java applications on both processing core types of the heterogeneous IBM Cell processor. An application's threads of execution can be transparently migrated between the Cell's different core types by Hera-JVM, without requiring the application's involvement. A number of real-world Java benchmarks are executed across both of the Cell's core types, to evaluate the efficacy of abstracting a heterogeneous architecture behind a homogeneous virtual machine. By characterising the performance of each of the Cell processor's core types under different program behaviours, a set of influential program behaviour characteristics is uncovered. A set of code annotations are presented, which enable program code to be tagged with these behaviour characteristics, enabling a runtime system to track a program's behaviour throughout its execution. This information is fed into a cost function, which Hera-JVM uses to automatically estimate whether the executing program's threads of execution would benefit from being migrated to a different core type, given their current behaviour characteristics. The use of history, hysteresis and trend tracking, by this cost function, is explored as a means of increasing its stability and limiting detrimental thread migrations. The effectiveness of a number of different migration strategies is also investigated under real-world Java benchmarks, with the most effective found to be a strategy that can target code, such that a thread is migrated whenever it executes this code. This dissertation also investigates the use of runtime monitoring to enable a runtime system to automatically infer a program's behaviour characteristics, without the need for explicit code annotations. A lightweight runtime behaviour monitoring system is developed, and its effectiveness at choosing the most appropriate core type on which to execute a set of real-world Java benchmarks is examined. Combining explicit behaviour characteristic annotations with those characteristics which are monitored at runtime is also explored.Finally, an initial investigation is performed into the use of behaviour characteristics to improve application performance under a different type of heterogeneous architecture, specifically, a non-uniform memory access (NUMA) architecture. Thread teams are proposed as a method of automatically clustering communicating threads onto the same NUMA node, thereby reducing data access overheads. Evaluation of this approach shows that it is effective at improving application performance, if the application's threads can be partitioned across the available NUMA nodes of a system.The findings of this work demonstrate that a runtime system with a homogeneous virtual machine interface can reduce the challenge of application development for HMA processors, whilst still being able to exploit such a processor by taking program behaviour into account.