GEVO-ML: a proposal for optimizing ML code with evolutionary computation
GEVO-ML: a proposal for optimizing ML code with evolutionary computation
复制标题
DOI：
10.1145/3377929.3398139
复制
发表时间：
2020-07
期刊：
影响因子：
--
通讯作者：
Jhe-Yu Liou;Xiaodong Wang;S. Forrest;Carole-Jean Wu
中科院分区：
文献类型：
--
作者：
Jhe-Yu Liou;Xiaodong Wang;S. Forrest;Carole-Jean Wu
并行加速器（例如GPU）是大规模机器学习（ML）应用程序的关键推动剂。但是，程序员通常缺乏对基本体系结构的详细知识，并且无法完全利用其计算能力。本文提出了Gevo-ML，该工具是一种自动发现优化机会并调整ML内核性能的工具。 GEVO-ML通过直接关注ML框架，中级语言和目标体系结构来扩展对GEVO（使用进化计算的GPU优化）的早期工作。它保留为GEVO开发的多目标进化搜索，该搜索搜索编辑为汇编为LLVM-IR的GPU代码，并根据所需的标准提高性能，同时保留所需的功能。在较早的工作中，我们在GPU设置中研究了一些ML工作负载，发现Gevo可以通过从1.7倍到2.9倍的因素来提高内核速度，即使仅访问了整个ML框架的一小部分。该研讨会论文研究了ML工作负载的GEVO的局限性和约束，并讨论了我们目前正在实施的Gevo-ML设计。
Parallel accelerators, such as GPUs, are a key enabler of large-scale Machine Learning (ML) applications. However, programmers often lack detailed knowledge of the underlying architecture and fail to fully leverage their computational power. This paper proposes GEVO-ML, a tool for automatically discovering optimization opportunities and tuning the performance of ML kernels. GEVO-ML extends earlier work on GEVO (Gpu optimization using EVOlutionary computation) by focusing directly on ML frameworks, intermediate languages, and target architectures. It retains the multi-objective evolutionary search developed for GEVO, which searches for edits to GPU code compiled to LLVM-IR and improves performance on desired criteria while retaining required functionality. In earlier work, we studied some ML workloads in GPU settings and found that GEVO could improve kernel speeds by factors ranging from 1.7X to 2.9X, even with access to only a small portion of the overall ML framework. This workshop paper examines the limitations and constraints of GEVO for ML workloads and discusses our GEVO-ML design, which we are currently implementing.