Climate Change Data Portal
| Automatic Dataflow Modeling for HPC Applications | |
| 项目编号 | CRSK-2_190359 |
| Calotoiu Alexandru | |
| 项目主持机构 | ETH Zurich - ETHZ |
| 开始日期 | 2020-02-01 |
| 结束日期 | 2021-01-31 |
| 中文摘要 | Whether it is climate change, artificial intelligence or quantum physics, scientific progress today requires ever increasing computational power. New, powerful supercomputers are being continuously designed to fulfill the practically boundless demand for high performance computing. Modern supercomputers are complex, heterogeneous systems containing a staggering number of processing units working in parallel, ranging from commodity processors like those in personal computers to powerful accelerators specifically engineered towards scientific use. Using such systems is challenging, as the computational work must be divided among many processing units. On these systems, the task usually requiring the most time and energy is transferring data to, from, and between the distributed memory of these processing units, as well as accessing the local memory of each individual processing unit. Therefore, ensuring that the data moves as little as possible is paramount to achieving good performance and low energy consumption, and even the decision of which processing unit configuration to select is difficult without a detailed understanding of both application and hardware. Recent work shows how programs expressing dataflow explicitly can be optimized to reduce the runtime up to eight times compared to solving the same problem without a data-centric approach. By creating an intermediate representation called a Stateful Dataflow Multigraph (SDFG) and applying transformations to this SDFG locality is significantly improved. Having dataflow be explicit in the code requires a complete rewrite for most applications.The difficulty and productivity loss in rewriting entire scientific programs, result in optimizations needing a data-centric representation remaining out of reach for most scientific endeavors.In this project, we wish to generate data-centric representations of arbitrary parallel programs - similar to SDFGs - to allow many locality-improving optimizations to be applied even to applications without explicit dataflow information. We will expand polyhedral loop analysis and taint analysis techniques to create dependency hypotheses and determine which code regions benefit most from optimizing locality. We will measure instrumented versions of these applications in various configurations to test these dependency hypotheses and focus on critical performance hotspots. After iterating this process to refine our hypotheses, we should ideally create a complete SDFG based on dependence models for the application and allow the optimization process to start. Even an incomplete collection of SDFGs for parts of an application should allow significant optimizations to locality.This kind of analysis has never been attempted, and would not be possible without leveraging cutting edge research in dataflow programming and performance modeling. It requires expanding state-of-the-art performance measurement and instrumentation techniques. The main challenge is creating a code analysis method to incorporate static and dynamic information to create SDFGs for parts of the code and then iteratively linking these together if possible. Developers could take advantage of optimizations previously restricted to dataflow programming and make informed decisions about which hardware configuration is best for their application without needing either a performance expert or a complete rewrite of their application. |
| 英文关键词 | Economic History ; Early Modern Period; Social History; Agrarian History; Climate Impacts; Late Middle Ages; Historical Climatology; Comparative History |
| 学科分类 | 1201 - 电子学与信息系统;12 - 信息科学 |
| 资助机构 | CH-SNSF |
| 项目经费 | 99992 |
| 项目类型 | Project funding;Project funding (special) |
| 国家 | CH |
| 语种 | 英语 |
| 文献类型 | 项目 |
| 条目标识符 | http://gcip.llas.ac.cn/handle/2XKMVOVA/125582 |
| 推荐引用方式 GB/T 7714 | Calotoiu Alexandru.Automatic Dataflow Modeling for HPC Applications.2020. |
| 条目包含的文件 | 条目无相关文件。 | |||||
| 个性服务 |
| 推荐该条目 |
| 保存到收藏夹 |
| 导出为Endnote文件 |
| 谷歌学术 |
| 谷歌学术中相似的文章 |
| [Calotoiu Alexandru]的文章 |
| 百度学术 |
| 百度学术中相似的文章 |
| [Calotoiu Alexandru]的文章 |
| 必应学术 |
| 必应学术中相似的文章 |
| [Calotoiu Alexandru]的文章 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
