PDS: The Illinois Aggressive Cache-Only Memory Architecture Multiprocessor- (I-ACOMA)

这项工作的目标是伊利诺伊州侵略性缓存-仅内存架构（I-ACOMA）多处理器的点设计研究，这是一种新型的分布式共享内存多处理器（DSM），在不放弃可编程性的情况下提供千万亿次的性能。I-ACOMA旨在解决架构师在未来20年将面临的一些最重要的挑战，即多ghz处理器的速度与更慢的缓存层次结构、内存和网络子系统之间的差距日益扩大，可以嵌入处理器芯片上的大量晶体管的利用率日益不足，以及在许多关键应用程序（如稀疏计算和数据库工作负载）中令人失望的低机器利用率。为了解决这些问题，I-ACOMA包含了非常激进的内存层次结构、多线程超标量处理器以及对稀疏计算和数据库的体系结构支持。在软件方面，运行在I-ACOMA上的代码是由自动并行Fortran 90程序的编译器生成的。编译器公开循环级和指令级并行性。此外，它还生成数据转发指令。最后，在应用程序方面，虽然I-ACOMA是一台通用机器，但它将重点关注两组大挑战应用程序，即稀疏计算和数据库应用程序。

The objective of this work is the point design study of the Illinois Aggressive Cache-Only Memory Architecture (I-ACOMA) multiprocessor, a novel distributed shared-memory, multiprocessor (DSM) that delivers petaflop-scale performance without surrendering programmability. I-ACOMA is designed to address some of the most important challenges that architects will face in the next two decades, namely the widening gap between the speed of multi-Ghz processors and the much slower cache hierarchy, memory and network subsystems, the increasing under-utilization of the large number of transistors that can be embedded on a processor chip, and the disappointingly poor machine utilization in many key applications like sparse computations and database workloads. To address these issues, I-ACOMA includes a very aggressive memory hierarchy, multithreaded superscalar processors, and architectural support for sparse computations and databases. On the software side, the code running on I-ACOMA is generated by a compiler that automatically parallelizes Fortran 90 programs. The compiler exposes both loop- and instruction-level parallelism. In addition, it generates data forward directives. Finally, on the application side, while I-ACOMA is a general-purpose machine, it will focus on two sets of grand challenge applications, namely sparse computations and database applications.