SBIR Phase II: Technology for Integrated Computation and Communication

这个小型企业创新研究计划第二阶段的研究项目建议开发一种创新的并行程序开发和执行技术的原型产品，该技术可以在不使用消息传递接口（MPI）的情况下，在多处理器和超级计算机中异步运行并行程序，速度比目前可能的快100倍。三十多年来，人们一直认为有效编译和执行并行程序的唯一途径是通过MPI。尽管人们已经认识到，如果在数据可用性的基础上异步执行并行程序会运行得更快，但在集成计算和通信技术（TICC）出现之前，无法获得有效执行并行程序所需的技术。这种调谐技术消除了对时间协调动态检查的需要，并使控制信号交换协议的执行与计算并行成为可能。每秒可以交换超过4000万条消息。这消除了通信瓶颈，并允许在不使用MPI的情况下基于数据可用性异步执行并行程序。TICC定义了因果语句的语义，并为它们提供了非常有效的实现。TICC带来了以下附加功能：(1)基于组件的并行程序开发环境；(2)并行程序的动态调试；(3)消息和消息流量的动态监控和更改；(4)动态修复和故障恢复；(5)动态重构；(5)并行软件系统的动态演化。这样做的好处是减少并行程序的开发和维护成本，使它们更容易和广泛地使用。再加上多处理器成本的下降，到2007年有可能迎来桌面超级计算的新时代，对科学、技术、工业、教育、计算和通信理论以及整个社会产生深远的影响。

This Small Business Innovation Research Program Phase II research project proposes to develop a prototype product for an innovative parallel program development and execution technology, which can run parallel programs asynchronously in multiprocessors and supercomputers up to 100 times faster than what is currently possible, without using Message Passing Interfaces (MPI). For more than thirty years it had been assumed that the only way to efficiently compile and execute parallel programs was through MPI. Even though it had been recognized that parallel programs would run faster if executed asynchronously on the basis of data availability, technology needed to do that efficiently was not available, until Technology for Integrated Computation and Communication (TICC) came along. This tuning technology eliminates the need for dynamic checking of temporal coordination, and makes it possible to execute control signal exchange protocols in parallel with computations. More than 40 million messages may be exchanged per second. This eliminates communication bottleneck and allows asynchronous execution of parallel programs based on data availability without using MPI. TICC defines the semantics of causal statements and provides a very efficient implementation for them.TICC brings the following additional facilities: (1) Component based parallel program development environment, (2) Dynamic debugging of parallel programs (3) Dynamic monitoring and changing of messages and message traffic, (4) Dynamic repair and failure recovery, (5) Dynamic reconfiguration, and (5) Dynamic evolution parallel software systems. These have the consequent benefit of reducing parallel program development and maintenance costs, making them more easily and widely available. This, together with decreasing costs of multiprocessors, has the potential to usher in a new era of desktop supercomputing by 2007, with profound impact on science, technology, industry, education, theories of computation and communication, and society in general.