Managed Runtimes for Energy Efficient Multicore Architecture

点击翻译按钮获取中文摘要

Developing, controlling and optimizing distributed applications on modern heterogeneousarchitectures in the context of energy efficiency is a complex task involving trade-offs between rawperformance (operations per second), power consumption (joules per second) and, for manyapplications, the accuracy of the computation (suitably defined), all of which may vary in terms ofsupply and demand - from either the algorithm and/or the hardware - over the lifetime of anapplication.Looking at the Apache projects dedicated to Big Data processing(https://projects.apache.org/projects.html?category), there are 36 different software frameworks.Although there are some projects that are more popular than others, a quick analysis reveals two keytrends. First, most of the software frameworks have inside a data-driven or event-driven model ofcomputation. Secondly, most of the software frameworks run on Java Virtual Machines.The University of Manchester has unique knowledge on data-driven computations, JIT compilers,runtime systems and virtualization. For example, we have been porting the Maxine Virtual Machine (aresearch JVM, https://en.wikipedia.org/wiki/Maxine_Virtual_Machine) to the ARMv7. In addition wehave included and improved the support and performance of the GRAAL JIT compiler in Maxine VM(https://arxiv.org/abs/1509.04085). The current status of Maxine VM on ARMv7 is capable of runningof complex applications, e.g. SLAMBench, on platforms such as the Xilinx Zynq (dual-core ARMCortex A9 with 1GB of memory).For more information on GRAAL and how it is being integrated into the OpenJDK visithttps://github.com/graalvm/graalvm and http://hg.openjdk.java.net/graal/graal-jvmci-9/The University of Manchester has also started working on the port of Maxine VM to ARMv8. Thestudent Hartley started this process as part of his MSc thesis. There is a significant engineering effortthat needs to be dedicated so as to be able to optimize server side Big Data frameworks for ARMv8-based datacenters.The main research question is how and where we can optimize data-driven computations for Big Dataprocessing on ARMv8 servers? The approach taken by this PhD project is to investigate the runtimesystem. That is, look for improvements at the JIT compiler, memory management, JVM as well as theinteraction with the OS and or directly with a hypervisor.This PhD is systems-oriented and we will follow a methodology of investigation driven by buildingprototypes which demonstrate our novel ideas. In order to make progress, the PhD student needs tocommand the literature in, and become familiar with: Managed Runtime Systems optimizations forperformance and energy plus ; compiler and runtime technologies (specifically for ARM-basedsystems); ARM architecture (many-cores, particularly ARMv8 systems.