Coupled Storage System for Efficient Management of Self-Describing Data Formats

在过去的几十年里，社会比以往任何时候都更加依赖信息技术的技术进步。特别是在科学研究领域，这确实使解决日益复杂的问题成为可能，这些问题现在需要超级计算机的计算能力。处理问题的复杂性不断增加，计算能力不断提高，导致数据量迅速增加；全球产生的数据量大约每两年翻一番，导致指数级数据泛滥。由于存储和网络技术的发展相当缓慢，这就造成了一个严重的问题。其结果是计算设备和存储设备之间的性能差距越来越大，从而导致存储瓶颈。这对于高性能计算中的大型系统来说尤其如此。为了缓解这种情况，使用了不同存储设备的层次结构，一方面满足对高容量的需求，另一方面满足对高速度和可靠性的需求。通过结合不同存储技术的优势，在降低购置、运维成本的同时，显著提高了整体性能。然而，对于未来的百亿亿级系统，困难将会变得更糟，为了利用系统的能力，需要进行关键的改进。现有的输入/输出（I/O）堆栈导致了额外的性能和管理问题。所产生的数据通常使用自描述数据格式存储，以便于科学界内部的交换和分析。该项目的目标是探索用于这些格式的耦合存储系统的好处。它将引入一种新的混合方法，利用来自高性能计算和数据库系统领域的存储技术，其中每种技术将根据其各自的优点和缺点使用。通过将存储系统与自描述数据格式紧密耦合，可以利用结构化信息选择合适的存储技术和存储层。由于这些信息目前是不可用的，存储系统必须采用启发式方法，这通常会导致次优性能以及不必要和昂贵的数据移动。此外，存储系统将支持可适应的I/O语义，以根据应用程序和数据格式需求调整其性能。总之，这些特性将支持全新的数据管理方法，并提供显著的性能改进。科学用户的现有工作流程将通过专用的数据分析界面得到支持。所有更改都将经过彻底测试，以确保与现有应用程序和接口的向后兼容性。因此，在CoSEMoS之上运行应用程序不需要任何修改，这有助于保留过去在科学软件开发方面的投资。

Over the last decades, societies came to rely more than ever on technological progress in information technology. Especially in the area of scientific research, this does enable the possibility to solve increasingly complex problems, which nowadays require the computational power of supercomputers. The rising complexity of the processed problems as well as the growth of computation power leads to rapidly increasing data volumes; the globally produced data volume doubles approximately every two years, leading to an exponential data deluge. This imposes a serious problem as the development of the storage and network technologies is considerably slower. The result is a widening gap between the performance of computing and storage devices, resulting in a storage bottleneck. This is especially true for large-scale systems found in high-performance computing. To ease this situation, a hierarchy of different storage devices is used to suffice the demand for high capacity on the one hand and for high velocity as well as reliability on the other hand. By combining the advantages of different storage technologies, the overall performance is significantly increased while inducing lesser costs for acquisition, operation and maintenance. However, for future exascale systems, the difficulties will get even worse, requiring critical improvements in order to exploit the systems' capabilities. The existing input/output (I/O) stack leads to additional performance and management issues.The produced data is typically stored using self-describing data formats to facilitate exchange and analysis within the scientific community. The project goal is to explore the benefits of a coupled storage system for these formats. It will introduce a novel hybrid approach leveraging storage technologies from the fields of high-performance computing and database systems, where each technology will be used according to its respective strengths and weaknesses. By coupling the storage system tightly with self-describing data formats, it can make use of structural information for selecting appropriate storage technologies and tiers. As such information is currently not available, storage systems have to employ heuristics, which often lead to suboptimal performance as well as unnecessary and expensive data movements. Moreover, the storage system will support adaptable I/O semantics to tune its performance according to application and data format requirements. Together, these features will enable completely new data management methods and provide significant performance improvements. Existing workflows of scientific users will be supported through a dedicated data analysis interface. All changes will be thoroughly tested to ensure backwards compatibility with existing applications and interfaces. Consequently, no modifications will be necessary to run applications on top of CoSEMoS, which helps preserve past investments in scientific software development.