Storage backends of parallel compute clusters are still based mostly on magnetic disks, while newer and faster storage technologies such as flash-based SSDs or non-volatile random access memory (NVRAM) are deployed within compute nodes. Including these new storage technologies into scientific workflows is unfortunately today a mostly manual task, and most scientists therefore do not take advantage of the faster storage media. One approach to systematically include nodelocal SSDs or NVRAMs into scientific workflows is to deploy ad hoc file systems over a set of compute nodes, which serve as temporary storage systems for single applications or longer-running campaigns. This paper presents results from the Dagstuhl Seminar 17202 "Challenges and Opportunities of User-Level File Systems for HPC" and discusses application scenarios as well as design strategies for ad hoc file systems using node-local storage media. The discussion includes open research questions, such as how to couple ad hoc file systems with the batch scheduling environment and how to schedule stage-in and stage-out processes of data between the storage backend and the ad hoc file systems. Also presented are strategies to build ad hoc file systems by using reusable components for networking and how to improve storage device compatibility. Various interfaces and semantics are presented, for example those used by the three ad hoc file systems BeeOND, GekkoFS, and BurstFS. Their presentation covers a range from file systems running in production to cutting-edge research focusing on reaching the performance limits of the underlying devices.

With the convergence of high-performance computing (HPC), big data and artificial intelligence (AI), the HPC community is pushing for "triple use" systems to expedite scientific discoveries. However, supporting these converged applications on HPC systems presents formidable challenges in terms of storage and data management due to the explosive growth of scientific data and the fundamental differences in I/O characteristics among HPC, big data and AI workloads. In this paper, we discuss the driving force behind the converging trend, highlight three data management challenges, and summarize our efforts in addressing these data management challenges on a typical HPC system at the parallel file system, data management middleware, and user application levels. As HPC systems are approaching the border of exascale computing, this paper sheds light on how to enable application-driven data management as a preliminary step toward the deep convergence of exascale computing ecosystems, big data, and AI.

It is hard for applications to make full utilization of the peak bandwidth of the storage system in highperformance computers because of I/O interferences, storage resource misallocations and complex long I/O paths. We performed several studies to bridge this gap in the Sunway storage system, which serves the supercomputer Sunway TaihuLight. To locate these issues and connections between them, an end-to-end performance monitoring and diagnosis tool was developed to understand I/O behaviors of applications and the system. With the help of the tool, we were about to find out the root causes of such performance barriers at the I/O forwarding layer and the parallel file system layer. An application-aware I/O forwarding allocation framework was used to address the I/O interferences and resource misallocations at the I/O forwarding layer. A performance-aware data placement mechanism was proposed to mitigate the impact of I/O interferences and performance variations of storage devices in the PFS. Together, applications obtained much better I/O performance. During the process, we also proposed a lightweight storage stack to shorten the I/O path of applications with N-N I/O pattern. This paper summarizes these studies and presents the lessons learned from the process.

Burst buffer has become a major component to meet the I/O performance requirement of HPC bursty traffic. This paper proposes Gfarm/BB that is a file system for a burst buffer efficiently exploiting node-local storage systems. Although node-local storages improve storage performance, they are only available during the job allocation. Gfarm/BB should have better access and metadata performance while it should be constructed on-demand before the job execution. To improve the read and write performance, it exploits the file descriptor passing and remote direct memory access (RDMA). It improves the metadata performance by omitting the persistency and the redundancy since it is a temporal file system. Using RDMA, writes and reads bandwidth are improved by 1.7x and 2.2x compared with IP over InfiniBand (IPoIB), respectively. It achieves 14 700 operations per second in the directory creation performance, which is 13.4x faster than the fully persistent and redundant case. The construction of Gfarm/BB takes 0.31 seconds using 2 nodes. IOR benchmark and ARGOT-IO application I/O benchmark show the scalable performance improvement by exploiting the locality of node-local storages. Compared with BeeOND, Gfarm/BB shows 2.6x and 2.4x better performance in IOR write and read benchmarks, respectively, and it shows 2.5x better performance in ARGOT-IO.

Many scientific fields increasingly use high-performance computing (HPC) to process and analyze massive amounts of experimental data while storage systems in today's HPC environments have to cope with new access patterns. These patterns include many metadata operations, small I/O requests, or randomized file I/O, while general-purpose parallel file systems have been optimized for sequential shared access to large files. Burst buffer file systems create a separate file system that applications can use to store temporary data. They aggregate node-local storage available within the compute nodes or use dedicated SSD clusters and offer a peak bandwidth higher than that of the backend parallel file system without interfering with it. However, burst buffer file systems typically offer many features that a scientific application, running in isolation for a limited amount of time, does not require. We present GekkoFS, a temporary, highly-scalable file system which has been specifically optimized for the aforementioned use cases. GekkoFS provides relaxed POSIX semantics which only offers features which are actually required by most (not all) applications. GekkoFS is, therefore, able to provide scalable I/O performance and reaches millions of metadata operations already for a small number of nodes, significantly outperforming the capabilities of common parallel file systems.

Modern High-Performance Computing (HPC) systems are adding extra layers to the memory and storage hierarchy, named deep memory and storage hierarchy (DMSH), to increase I/O performance. New hardware technologies, such as NVMe and SSD, have been introduced in burst buffer installations to reduce the pressure for external storage and boost the burstiness of modern I/O systems. The DMSH has demonstrated its strength and potential in practice. However, each layer of DMSH is an independent heterogeneous system and data movement among more layers is significantly more complex even without considering heterogeneity. How to efficiently utilize the DMSH is a subject of research facing the HPC community. Further, accessing data with a high-throughput and low-latency is more imperative than ever. Data prefetching is a well-known technique for hiding read latency by requesting data before it is needed to move it from a high-latency medium (e.g., disk) to a low-latency one (e.g., main memory). However, existing solutions do not consider the new deep memory and storage hierarchy and also suffer from under-utilization of prefetching resources and unnecessary evictions. Additionally, existing approaches implement a client-pull model where understanding the application's I/O behavior drives prefetching decisions. Moving towards exascale, where machines run multiple applications concurrently by accessing files in a workflow, a more data-centric approach resolves challenges such as cache pollution and redundancy. In this paper, we present the design and implementation of Hermes:a new, heterogeneous-aware, multi-tiered, dynamic, and distributed I/O buffering system. Hermes enables, manages, supervises, and, in some sense, extends I/O buffering to fully integrate into the DMSH. We introduce three novel data placement policies to efficiently utilize all layers and we present three novel techniques to perform memory, metadata, and communication management in hierarchical buffering systems. Additionally, we demonstrate the benefits of a truly hierarchical data prefetcher that adopts a server-push approach to data prefetching. Our evaluation shows that, in addition to automatic data movement through the hierarchy, Hermes can significantly accelerate I/O and outperforms by more than 2x state-of-the-art buffering platforms. Lastly, results show 10%-35% performance gains over existing prefetchers and over 50% when compared to systems with no prefetching.

Technology enhancements and the growing breadth of application workflows running on high-performance computing (HPC) platforms drive the development of new data services that provide high performance on these new platforms, provide capable and productive interfaces and abstractions for a variety of applications, and are readily adapted when new technologies are deployed. The Mochi framework enables composition of specialized distributed data services from a collection of connectable modules and subservices. Rather than forcing all applications to use a one-size-fits-all data staging and I/O software configuration, Mochi allows each application to use a data service specialized to its needs and access patterns. This paper introduces the Mochi framework and methodology. The Mochi core components and microservices are described. Examples of the application of the Mochi methodology to the development of four specialized services are detailed. Finally, a performance evaluation of a Mochi core component, a Mochi microservice, and a composed service providing an object model is performed. The paper concludes by positioning Mochi relative to related work in the HPC space and indicating directions for future work.

Scientific applications at exascale generate and analyze massive amounts of data. A critical requirement of these applications is the capability to access and manage this data efficiently on exascale systems. Parallel I/O, the key technology enables moving data between compute nodes and storage, faces monumental challenges from new applications, memory, and storage architectures considered in the designs of exascale systems. As the storage hierarchy is expanding to include node-local persistent memory, burst buffers, etc., as well as disk-based storage, data movement among these layers must be efficient. Parallel I/O libraries of the future should be capable of handling file sizes of many terabytes and beyond. In this paper, we describe new capabilities we have developed in Hierarchical Data Format version 5 (HDF5), the most popular parallel I/O library for scientific applications. HDF5 is one of the most used libraries at the leadership computing facilities for performing parallel I/O on existing HPC systems. The state-of-the-art features we describe include:Virtual Object Layer (VOL), Data Elevator, asynchronous I/O, full-featured single-writer and multiple-reader (Full SWMR), and parallel querying. In this paper, we introduce these features, their implementations, and the performance and feature benefits to applications and other libraries.

With the advancement of new information technologies, a revolution is being taken place to bring the industry into a new era of intelligent manufacturing. One of the key requirements of intelligent manufacturing is the interoperability of industrial applications. However, it is challenging to realize the interoperability for legacy industrial applications due to 1) the deficient semantic information of data transmitted over heterogeneous communication protocols, 2) the difficulty to understand the complex process of business logic with no source code, and 3) the high cost and potential risk of reengineering the applications. To address the issues, in this paper, we propose an approach named SmartPipe to exposing existing functionalities of an industrial application as APIs without source code while simultaneously allowing the application to remain unchanged. We design a behavioral runtime model (BRM) as the self-representation of the industrial applications, based on which a computational reflection framework is designed to flexibly construct the model and generate APIs that encapsulate specific functionalities. We validate SmartPipe on a real industrial application that controls the spin-draw winding machine. Results show that our approach is effective and more suitable for industrial scenes compared with traditional approaches.

Internet of Things (IoT) applications have massive client connections to cloud servers, and the number of networked IoT devices is remarkably increasing. IoT services require both low-tail latency and high concurrency in datacenters. This study aims to determine whether an order of magnitude improvement is possible in tail latency and concurrency in mainstream systems by proposing a hardware-software codesigned labeled network stack (LNS) for future datacenters. The key innovation is a cross-layered payload labeling mechanism that distinguishes different requests by payload across the full network stack, including application, TCP/IP, and Ethernet layers. This type of design enables prioritized data packet processing and forwarding along the full datapath, such that latency-insensitive requests cannot significantly interfere with high-priority requests. We build a prototype datacenter server to evaluate the LNS design against a commercial X86 server and the mTCP research, using a cloud-supported IoT application scenario. Experimental results show that the LNS design can provide an order of magnitude improvement in tail latency and concurrency. A single datacenter server node can support over 2 million concurrent long-living connections for IoT devices as a 99-percentile tail latency of 50 ms is maintained. In addition, the hardware-software codesign approach remarkably reduces the labeling and prioritization overhead and constrains the interference of high-priority requests to low-priority requests.

This paper presents CirroData, a high-performance SQL-on-Hadoop system designed for Big Data analytics workloads. As a home-grown enterprise-level online analytical processing (OLAP) system with more than seven-year research and development (R&D) experiences, we share our design details to the community about how to achieve high performance in CirroData. Multiple optimization techniques have been discussed in the paper. The effectiveness and the efficiency of all these techniques have been proved by our customers' daily usage. Benchmark-level studies, as well as several real application case studies of CirroData, have been presented in this paper. Our evaluations show that CirroData can outperform various types of counterpart database systems in the community, such as "Spark+Hive", "Spark+HBase", Impala, DB-X/Y, Greenplum, HAWQ, and others. CirroData can achieve up to 4.99x speedup compared with Greenplum, HAWQ, and Spark in the standard TPC-H queries. Application-level evaluations demonstrate that CirroData outperforms "Spark+Hive" and "Spark+HBase" by up to 8.4x and 38.8x, respectively. In the meantime, CirroData achieves the performance speedups for some application workloads by up to 20x, 100x, 182.5x, 92.6x, and 55.5x as compared with Greenplum, DB-X, Impala, DB-Y, and HAWQ, respectively.

Both resource efficiency and application QoS have been big concerns of datacenter operators for a long time, but remain to be irreconcilable. High resource utilization increases the risk of resource contention between co-located workload, which makes latency-critical (LC) applications suffer unpredictable, and even unacceptable performance. Plenty of prior work devotes the effort on exploiting effective mechanisms to protect the QoS of LC applications while improving resource efficiency. In this paper, we propose MAGI, a resource management runtime that leverages neural networks to monitor and further pinpoint the root cause of performance interference, and adjusts resource shares of corresponding applications to ensure the QoS of LC applications. MAGI is a practice in Alibaba datacenter to provide on-demand resource adjustment for applications using neural networks. The experimental results show that MAGI could reduce up to 87.3% performance degradation of LC application when co-located with other antagonist applications.

An analysis of real-world operational data of Tianhe-1A (TH-1A) supercomputer system shows that chilled water data not only can reflect the status of a chiller system but also are related to supercomputer load. This study proposes AquaSee, a method that can predict the load and cooling system faults of supercomputers by using chilled water pressure and temperature data. This method is validated on the basis of real-world operational data of the TH-1A supercomputer system at the National Supercomputer Center in Tianjin. Datasets with various compositions are used to construct the prediction model, which is also established using different prediction sequence lengths. Experimental results show that the method that uses a combination of pressure and temperature data performs more effectively than that only consisting of either pressure or temperature data. The best inference sequence length is two points. Furthermore, an anomaly monitoring system is set up by using chilled water data to help engineers detect chiller system anomalies.