Semantic information based speculative parallel execution
Semantic information based speculative parallel execution
复制标题
基于语义信息的推测并行执行
DOI：
--
复制
发表时间：
2010
期刊：
影响因子：
--
通讯作者：
P. Stenström
中科院分区：
文献类型：
--
作者：
A. Vajda;P. Stenström
随着一个处理器内的核的数量将继续增加，而每核频率可能停滞，一般的推测性执行和具体的数据值预测将再次成为用于加速具有有限的固有并行性的应用的执行的主要候选者。然而，为了取得成功，我们需要改进的语义信息的应用程序将能够提供给硬件，在本文中，我们提出了初步结果的基础上，使用正式的合同，以指导投机执行的类型和范围，将这些信息传递到运行时环境。对传统上被认为难以并行化的应用程序（霍夫曼解码）的实验显示了这种方法的潜力，而无需重写实际应用程序。这种方法的优点之一是，它可以有效地利用大量的核心，打开了一个可能的路径扩展应用程序与大规模多核处理器上的并行度有限。
As the number of cores within one processor will continue to increase, while per-core frequency is likely to stagnate, speculative execution in general and data value prediction in particular, will once again be a prime candidate for speeding up execution of applications with limited inherent parallelism. In order to succeed however, we need to improve the semantic information the applications will be able to provide to the hardware; in this paper we present initial results based on using formal contracts to convey such information to the run-time environment in order to guide the type and scope of speculative execution. Experiments with an application that is traditionally considered hard to parallelize – Huffman decoding – show the potential of such an approach, without the need to re-write the actual application. One of the strengths of this method is that it can make efficient use of large number of cores, opening a possible path for scaling applications with limited amount of parallelism on massively multi-core processors.