Empirical Studies of Process Distribution and Redistributionin Multicomputers

本项目审查非侵入性机制的好处， 在分布式存储器中分布负载和状态信息 多计算机 为了从并行的可能加速中受益， 在计算机中，程序必须被划分为工作在 并联 一旦分区，这些单元必须分配给特定的 处理器执行。 在分布式内存多计算机上， 与重新布置现有的 任务，称为负载重新分配。 在以前的工作中，程序跟踪 在参考架构上模拟小粒度程序， 使用现代网络架构。 互联网络确实 不需要很强的通信局部性，因为消息可以 快速直通中间节点，有利于简化流程 分销战略。 在这个项目中， 通过用于非侵入式加载信息的固件机制来增强 分布 这一信息被证明是至关重要的过程 分布;这些机制也适用于过程 同步和网络资源预留。 上一 示例程序，处理器没有什么优势， 重新分配;这项研究将确定这是否适用于平行 数字应用的仪器和测量 程序并与当前和预期的系统性能进行比较。

This project examines the benefit of non-intrusive mechanisms for distributing load and status information in a distributed memory multicomputer. To benefit from the possible speedup of parallel computers, programs must be partitioned into units that work in parallel. Once partitioned, these units, must be assigned to specific processors for execution. On distributed memory multicomputers, this is called load distribution, in comparison to rearranging extant tasks, termed load redistribution. In previous work, program traces from small-grain programs were simulated on a reference architecture, using a modern network architecture. The interconnection network did not require strong communication locality because messages could quickly cut-through intermediate nodes, favoring simple process distribution strategies. In this project, the reference architecture is augmented by firmware mechanisms for non-intrusive load information distribution. This information was shown to be critical for process distribution; these mechanisms are also useful for process synchronization and network resource reservation. For the previous sample programs, there was little advantage to processor redistribution; this study will determine if this is true for parallel numerical applications by instrumentation and measurement of those programs and comparison to current and expected system performance.