Learning quantitative representation synthesis

Learning quantitative representation synthesis
复制标题

学习定量表示综合

DOI:
10.1145/3394450.3397467
复制
发表时间:
2020
期刊:
Proceedings of the 4th ACM SIGPLAN International Workshop on Machine Learning and Programming Languages
影响因子:
--
通讯作者:
Lesani, Mohsen
Lesani, Mohsen
中科院分区:
--
文献类型:
--
作者:
Patil, Mayur;Houshmand, Farzin;Lesani, Mohsen

文献摘要

参考文献

相似文献

软件系统通常使用数据结构的专门组合来存储和检索数据。设计和维护自定义数据结构特别是并发数据结构是一个耗时且易出错的问题,我们让用户将所需数据声明为关系和方法接口的高级规范,并自动合成正确高效的并发数据表示。我们提出了可证明的合理的语法推导来合成结构,有效地支持接口。然后,我们合成同步,以支持并发执行的结构。多个候选代表可能满足相同的规格,我们的目标是定量选择最有效的候选人。以前的作品要么使用动态自动调谐器来执行和测量的候选人的性能或使用静态成本函数来估计其性能。然而,对许多候选对象重复执行是耗时的,并且单个性能模型不能有效地预测所有平台上的所有工作负载。我们提出了一种新的方法来定量合成,学习的性能模型。我们开发了一个名为Leqsy的合成工具,它训练人工神经网络来静态预测候选表示的性能。实验评估表明,Leqsy可以合成接近最佳的表示。
Software systems often use specialized combinations of data structures to store and retrieve data. Designing and maintaining custom data structures particularly concurrent ones is time-consuming and error-prone.We let the user declare the required data as a high-level specification of a relation and method interface, and automatically synthesize correct and efficient concurrent data representations. We present provably sound syntactic derivations to synthesize structures that efficiently support the interface.We then synthesize synchronization to support concurrent execution on the structures. Multiple candidate representations may satisfy the same specification and we aim at quantitative selection of the most efficient candidate. Previous works have either used dynamic auto-tuners to execute and measure the performance of the candidates or used static cost functions to estimate their performance. However, repeating the execution for many candidates is time-consuming and a single performance model cannot be an effective predictor of all workloads across all platforms. We present a novel approach to quantitative synthesis that learns the performance model. We developed a synthesis tool called Leqsy that trains an artificial neural network to statically predict the performance of candidate representations. Experimental evaluations demonstrate that Leqsy can synthesize near-optimum representations.
并发数据结构客户端的自动原子性验证
DOI: 10.1007/978-3-319-08867-9_37
发表时间: 2014
期刊: --
影响因子: --
作者:
M. Lesani;T. Millstein;J. Palsberg
通讯作者: J. Palsberg
Elixir:一个用于综合并发图程序的系统
DOI: --
发表时间: 2012
期刊: Conference on Object-Oriented Programming Systems, Languages, and Applications
影响因子: --
作者:
Dimitrios Prountzos;R. Manevich;K. Pingali
通讯作者: K. Pingali
智能数据结构:多核数据结构的在线机器学习方法
DOI: --
发表时间: 2011
期刊: International Conference on Automation and Computing
影响因子: --
作者:
J. Eastep;David Wingate;A. Agarwal
通讯作者: A. Agarwal
DOI: --
发表时间: 2012
期刊:
影响因子: --
作者:
HawkinsPeter;AikenAlex;FisherKathleen;RinardMartin;SagivMooly
通讯作者: SagivMooly
实现自动锁移除以实现可扩展同步
DOI: --
发表时间: 2015
期刊: International Symposium on Distributed Computing
影响因子: --
作者:
M. Arbel;Guy Golan;Eshcar Hillel;I. Keidar
通讯作者: I. Keidar