基于SPARC架构面向确定性重演的多核访存竞争记录方法的研究

发布时间：2018-06-19 08:04

本文选题：多核不确定性 + 确定性重演　；参考：《哈尔滨工业大学》2013年硕士论文

【摘要】：在电子信息技术的不断发展进步的今天，人们对高性能计算机的需求越来越强。然而受材料、加工工艺和散热功耗等原因的限制，通过不断提高CPU主频而提升处理器性能的设计模式变得愈来愈困难，因此业界逐渐转向了在单个芯片上集成多个处理器的片上多核处理器结构。与此同时也带来了某些在单核心处理器架构中不会出现的新问题，由多核系统的访存竞争导致的多核不确定性问题就是其中之一，这个问题逐渐成为计算机多核系统结构与并行计算中研究的热点领域之一。现有的多核架构下的Cache一致性协议能有效的防止并行程序共享内存中数据的不一致，但是对多线程间访存的顺序没有控制。也就是说如果程序员在程序中对各处理器内部高速缓存的共享数据的访问未采取严格的同步措施，就可能会使程序出现违反期望执行顺序的Bug，即在相同的输入情况下多次执行的输出可能不同。为了解决这一难题，人们提出了确定性重演技术，该技术是指在存在并发错误的并行程序的某次执行过程中记录必要的线程执行信息，然后在程序的第二次执行过程中索引这些信息，并强制程序按照相同的方式或顺序执行，以确保此次执行的结果和前次执行的相同。本文在理解了了SPARC架构、Cache一致性协议的基础上，针对共享存储式CMP架构的片上多核SPARC系统在并发地访问共享数据时出现访存竞争从而导致的多核不确定性问题进行研究，提出了一种硬件辅助的面向确定性重演的访存竞争记录方法——ERTR方法。该方法法采用了基于滑动窗口的分块约减竞争序的思想，通过先对各个处理器线程的指令进行分块记录并约减掉冗余的竞争序，再引入RTR方法中的滑动窗口思想进行替换性约减，因此能够实现对访存竞争序更为高效的约减。最后，通过搭建SIMICS+GEMS的仿真测试平台，，创建UltraSPARC架构四核处理器虚拟机对ERTR方法进行验证测试，测试的项目是SPLASH-2并行程序集，在日志记录空间大小、时间开销、带宽开销等方面与其他方法进行对比测试，实验结果表明ERTR方法的是一种正确高效的访存冲突记录方法。
[Abstract]:With the continuous development and progress of electronic information technology, the demand for high-performance computers is becoming stronger and stronger. However, due to the limitation of material, processing technology and heat dissipation, the design mode of improving CPU performance by increasing CPU frequency becomes more and more difficult. As a result, the industry has gradually shifted to a multi-core processor architecture that integrates multiple processors on a single chip. At the same time, it also brings some new problems that will not appear in the single-core processor architecture, one of which is the multi-core uncertainty caused by the multi-core system memory access competition. This problem has gradually become one of the hot research fields in the architecture and parallel computing of computer multi-core systems. The existing cache conformance protocol under multi-core architecture can effectively prevent the inconsistency of data in shared memory of parallel programs, but there is no control over the order of memory access between multiple threads. That is, if the programmer does not take strict synchronization measures to access shared data cached within each processor in the program, A Bug that may cause the program to violate the expected order of execution, that is, the output that is executed multiple times with the same input may be different. In order to solve this problem, a deterministic replay technique is proposed to record the necessary thread execution information during the execution of a parallel program with concurrent errors. This information is then indexed during the second execution of the program and enforced in the same manner or order to ensure that the result of this execution is the same as that of the previous execution. On the basis of understanding the cache consistency protocol of SPARC architecture, this paper studies the multi-core uncertainty caused by the multi-core SPARC system accessing the shared data concurrently, which is caused by the multi-core SPARC system in the shared storage CMP architecture. In this paper, a hardware aided memory access and contention record (ERTR) method for deterministic replay is proposed. The method adopts the idea of reducing the competition order by block reduction based on sliding window. The instructions of each processor thread are recorded in blocks and the redundant competition order is reduced. Then the sliding window in RTR method is introduced to reduce the substitutability, so that the reduction of the competitive order of memory access can be realized more efficiently. Finally, by setting up the simulation test platform of Simics gems, creating a four-core processor virtual machine of UltraSPARC architecture to verify and test ERTR method, the test item is SPLASH-2 parallel assembly, the space size of log record, the time cost, the test item is SPLASH-2 parallel assembly. The bandwidth overhead is compared with other methods. The experimental results show that ERTR is a correct and efficient method for recording access to memory conflicts.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP332

【相似文献】