面向图计算应用的JVM可扩展性研究

Graph data structure and algorithm are widely used in many scenarios like social network, searching service and machine learning. The mainstream open source platforms for big data processing, e.g., Hadoop and Spark, are developed using JVM-based language, thus those graph computing platforms also based on JVM, such as Giraph and GraphX, are popular and widely deployed. Currently, the mainstream in-memory graph computing frameworks are more suitable to be deployed on small cluster systems compose of high-performance computing nodes. However, the performance and scalability of these frameworks are limited by the underlying JVM platform, which cannot fully leverage the computing resource provided by many-core processors. Taking the characteristics of modern many-core architecture as well as graphic computing into consideration, this project focuses on the research of the performance and scalability of JVM. First, we will address the problems including poor scalability and large overhead of resource abstraction and management by diagnosing the performance bottleneck and designing analysis tools and benchmarks for many-core based JVM. Second, by using grey-box design methodology and considering the characteristics of application and architecture, we will propose optimizations of JVM for graphic computing applications. Last but not least, we will do further research on white-box design in order to propose and implement an efficient customized JVM for graph computing applications.

图数据结构及算法被广泛应用于社交网络、搜索服务和机器学习等场景。主流大数据开源 平台(如Hadoop和Spark)多使用基于JVM的语言,因此基于JVM的图计算框架Giraph和GraphX 得到了普遍关注和广泛应用。内存图计算是图计算的主流框架,更适合运行在基于高性能单机 的小规模集群系统上。然而,主流开源图计算系统的性能与可扩展性严重受限于现有JVM平台 ,无法充分利用众核处理器所提供的丰富计算资源。本项目拟结合当前众核体系结构及图计算 的特征,研究JVM的性能与可扩展性。首先,针对图计算应用在众核JVM平台上可扩展性差、资 源抽象与管理开销大等问题,分析其性能瓶颈,设计面向众核平台的JVM分析工具与基准测试 集。其次,采用灰盒设计的方法,结合应用与体系结构特征,提出面向图计算应用的JVM优化 方案。最后,进一步研究白盒设计的方法,针对图计算应用设计并实现一个高效的定制化JVM 。

本项目针对非易失性内存（NVM）这一新兴硬件设备，重点分析基于Java虚拟机运行的图计算等应用在NVM上的可扩展性及性能。本项目首次提出并实现了为NVM提供原生支持的定制化Java虚拟机——Espresso，允许Java应用以普通对象形式直接在NVM上读写数据，充分利用了NVM低访问时延的优势，并在后续工作GCPersist中提出新的持久化模型以提高NVM在图计算应用中的可用性。相关论文发表于程序语言及系统领域顶级会议ASPLOS上，是最早提出从语言虚拟机层面提供NVM支持的研究工作之一。.本项目还分析了图计算应用在运行中存在的可扩展性问题，并通过程序合成和垃圾回收优化的方法予以解决。.本项目共发表5篇论文，其中3篇为CCF A类会议论文。项目中关于垃圾回收优化的部分代码已经进入了OpenJDK主线。项目共培养5名博士生（3名毕业），7名硕士生（5名毕业）。