Supercomputing application simulations on hundreds of thousands of cores produce vast amounts of data that need to be analyzed on smaller-scale clusters to glean insights. The process is referred to as an end-to-end workflow. Extant workflow systems are stymied by the storage wall, resulting from both the disk-based parallel file system (PFS) failing to keep pace with the compute and memory subsystems as well as the inefficiencies in end-to-end workflow processing. In the post-petaflop era, supercomputers are provisioned with flash devices, as an intermediary between compute nodes and the PFS, enabling novel paradigms not just for expediting I/O, but also for the in-situ analysis of the simulation output data on the flash device. An array of such active flash elements allows us to fundamentally rethink the way data analysis workflows interact with storage systems. By blending the flash storage array and data analysis together in a seamless fashion, we create an analysis workflow-aware storage system, AnalyzeThis. Our guiding principle is that analysis-awareness be deeply ingrained in each and every layer of the storage system—active flash fabric, analysis object abstraction layer, scheduling layer within the storage, and an easy-to-use file system interface—thereby elevating data analyses as first-class citizens. Together, these concepts transform AnalyzeThis into a potent analytics-aware appliance. / Master of Science

Master of Science / Department of Computing and Information Sciences / Daniel A. Andresen / The recovery-driven design of the file system has been one of the most challenging fields over the major trends in operating systems. This field has assumed considerable importance in the past decades as the disk sizes have been increasing without a comparable increase in the disk I/O bandwidth and seek time. The rapid increase in the storage size is expected to become constant in the future due to the growing market demand and the continuous database size increment of many companies and major businesses. Due to the same reason, the cost of the average file system checking time has increased without a significant improvement in the disk I/O bandwidth and seek time performance. Operating system bugs, power outages, and hardware failures which result in a file system crash were the main reasons behind the innovation of novel recovery approaches such as Journaling and soft-updates. Although such approaches avoided complete file system checking by checking solely inconsistencies in file system metadata, it become inevitable for them to check the entire file system for inconsistencies because of the previously entioned types of problems. One of the emerging recovery-driven designs which considers minimizing file system checking cost is the Chunkfs files ystem. Chunkfs file system introduces an innovative look into the file system design by dividing the file system layout into smaller chunks, each one of which represents a smaller scale file system by itself. In our work we probed an alternative recovery-driven design which is considerably inspired by the Chunkfs concepts and follows the same design guidelines. This recovery-riven design is introduced by adding a new feature to the file system which best utilizes the existing underlying design through considering the block groups as individual chunks, confining their files and directories spanning across different block groups by means of special controlled continuation links. These links provide a fault isolation means by circumscribing the checking of the file system to only these block groups which appear to be dirty after a crash, a method resulting in a moderate reduction in file system checking cost. We also probed different metrics of metadata sizes, and the probable cost of files and directories expansion across the different block groups.

Title: Analysis of a File System Using the Verifying C Compiler Author: Bc. David Škorvaga Department: Department of Distributed and Dependable Systems Supervisor: RNDr. Jan Kofroň, Ph.D. Abstract: Formal verification is a way to improve reliability of software systems. One approach of formal verification is focused on proving correctness of annotat- ed source code of an established programming language. Verifying C Compiler (VCC) is a verifier for concurrent C that accepts an annotated code in C language and automatically verifies its correctness with respect to the given annotation. There have been successful attempts to verify some critical systems, including the operating system kernel. Another critical part of operating system is its file system. In the thesis, we choose FatFs file system, a simple device-independent implementation of the FAT file system. We specify a part of it using the VCC annotation and successfully verify its correctness. Keywords: Formal Verification, File System, VCC

This dissertation presents novel approaches to the use of non-volatile memory devices in building storage systems. There are many types of non-volatile memory devices, and they usually have better performance than regular magnetic hard disks in terms of throughput and latency. This dissertation focused on two of them, NAND flash memory and Phase Change Memory (PCM). This work consisted of two parts. The first part was to design a high-performance hybrid storage system employing Solid State Drives that are build out of NAND flash memory and Hard Disk Drives. In this hybrid system, we proposed two different policies to improve its performance. One is to exploit the fact that the performances of Solid State Drive and Hard Disk Drive are asymmetric and the other is to exploit concurrency on multiple devices. We implemented prototypes in Linux and evaluate both policies in multiple workloads and multiple configurations. The results showed that the proposed approaches improve the performance significantly, and adapt to different configurations of the system under different workloads. The second part was to implement a file system on a special class of memory devices, Storage Class Memory (SCM), which is both byte addressable and also nonvolatile, e.g. PCM. We claimed that both the existing regular file systems and the memory based file systems are not suitable for SCM, and proposed a new file system, called SCMFS, which is implemented on the virtual address space. In SCMFS, we utilized the existing memory management module in the operating system to do the block management. Our design keeps address space within a file contiguous to reduce the block management software. The simplicity of SCMFS not only makes it easy to implement, but also improves the performance. We implemented a prototype of SCMFS in Linux and evaluated its performance through multiple benchmarks.

Thesis (M.S.)--Worcester Polytechnic Institute. / Keywords: LAN; Linux; NFS. Includes bibliographical references (p. 91-92).

Diversas aplicações executadas em ambientes de cluster necessitam de uma área de armazenamento permanente com alta capacidade e que forneça uma visão homogênea dos dados entre todos os nós. Esta área compartilhada é comumente implementada através de um sistema de arquivos distribuído, permitindo o acesso através da abstração mais comum para gerenciamento de dados. A disparidade entre poder de processamento e desempenho de dispositivos de armazenamento atuais, no entanto, torna tais sistemas um ponto crítico ao desempenho de aplicações paralelas que lidam com grandes volumes de dados. Ambientes de cluster podem apresentar execução concorrente de aplicações em conjuntos independentes de máquinas. Desta forma, uma grande quantidade de clientes com características distintas farão acessos ao sistema de arquivos compartilhado. Em tais casos, o dimensionamento do sistema de armazenamento distribuído nem sempre poderá prover o desempenho necessário à execução das aplicações com os recursos inicialmente a ele destinados. O presente trabalho propõe uma estratégia de reconfiguração dinâmica para o sistema de arquivos dNFSp. Esta estratégia leva em consideração o comportamento temporal presente em aplicações paralelas para inserir servidores de dados exclusivos a aplicações com alta demanda de I/O. Com a utilização de servidores exclusivos, torna-se possível isolar aplicações com comportamentos que causam grande perda de desempenho no sistema como um todo. Foi desenvolvida uma ferramenta de monitoramento do desempenho junto aos servidores do dNFSp, de forma a solicitar novos servidores apenas quando a interação das fases de I/O das aplicações em execução interferirem no desempenho do sistema. Esses novos servidores são solicitados para um sistema escalonador de recursos para cluster, de forma a utilizar nós livres para o armazenamento de dados. Os resultados mostram que a estratégia proposta é capaz de detectar a saturação do sistema de armazenamento e iniciar os servidores exclusivos, levando a um ganho de desempenho para as aplicações em execução. A reconfiguração dinâmica também mostrou-se capaz de evitar baixos desempenhos causados pela interação dos períodos de inatividade de aplicações temporais e pela utilização de diversos processos de I/O em um único nó. / Several applications executed in cluster environments need a shared storage area with high capacity and a homogeneous view of the stored data to all processing nodes. This area is commonly implemented as a distributed file system, allowing the access to data through the well-known file abstraction. The great gap in performance of processors and storage devices, on the other hand, makes such system a critical point to the performance of parallel applications. A common trait of large cluster environments is the concurrent execution of applications. In this scenario, many clients with distinct behaviors will compete to access the shared storage system. The number of I/O resources originally dedicated to this shared storage may provide unsatisfactory performance to the applications in this case. This work proposes a dynamic reconfiguration strategy for the dNFSp file system. This strategy takes into consideration the temporal behavior of distributed applications to launch dedicated I/O resources to the more I/O-demanding applications. The exclusiveness of data servers allows for the isolation of access patterns that cause contention on the system, resulting in improved performance for all executing applications. We developed a tool to monitor the performance of the storage servers in order to launch new servers only when the interaction of I/O phases from running applications cause each other’s performance to drop. These resources are required to a batch scheduler system present on the cluster, allowing the use of computing nodes for temporary data storage. The results show that the proposed metrics lead to the detection of performance saturation on the file system and the start of the dedicated resources, resulting in increased I/O performance. The reconfiguration has also been able to avoid some low performance situations caused by interactions of inactivity periods from temporal applications and by utilization of several I/O processes on the same cluster node.

Diversas aplicações executadas em ambientes de cluster necessitam de uma área de armazenamento permanente com alta capacidade e que forneça uma visão homogênea dos dados entre todos os nós. Esta área compartilhada é comumente implementada através de um sistema de arquivos distribuído, permitindo o acesso através da abstração mais comum para gerenciamento de dados. A disparidade entre poder de processamento e desempenho de dispositivos de armazenamento atuais, no entanto, torna tais sistemas um ponto crítico ao desempenho de aplicações paralelas que lidam com grandes volumes de dados. Ambientes de cluster podem apresentar execução concorrente de aplicações em conjuntos independentes de máquinas. Desta forma, uma grande quantidade de clientes com características distintas farão acessos ao sistema de arquivos compartilhado. Em tais casos, o dimensionamento do sistema de armazenamento distribuído nem sempre poderá prover o desempenho necessário à execução das aplicações com os recursos inicialmente a ele destinados. O presente trabalho propõe uma estratégia de reconfiguração dinâmica para o sistema de arquivos dNFSp. Esta estratégia leva em consideração o comportamento temporal presente em aplicações paralelas para inserir servidores de dados exclusivos a aplicações com alta demanda de I/O. Com a utilização de servidores exclusivos, torna-se possível isolar aplicações com comportamentos que causam grande perda de desempenho no sistema como um todo. Foi desenvolvida uma ferramenta de monitoramento do desempenho junto aos servidores do dNFSp, de forma a solicitar novos servidores apenas quando a interação das fases de I/O das aplicações em execução interferirem no desempenho do sistema. Esses novos servidores são solicitados para um sistema escalonador de recursos para cluster, de forma a utilizar nós livres para o armazenamento de dados. Os resultados mostram que a estratégia proposta é capaz de detectar a saturação do sistema de armazenamento e iniciar os servidores exclusivos, levando a um ganho de desempenho para as aplicações em execução. A reconfiguração dinâmica também mostrou-se capaz de evitar baixos desempenhos causados pela interação dos períodos de inatividade de aplicações temporais e pela utilização de diversos processos de I/O em um único nó. / Several applications executed in cluster environments need a shared storage area with high capacity and a homogeneous view of the stored data to all processing nodes. This area is commonly implemented as a distributed file system, allowing the access to data through the well-known file abstraction. The great gap in performance of processors and storage devices, on the other hand, makes such system a critical point to the performance of parallel applications. A common trait of large cluster environments is the concurrent execution of applications. In this scenario, many clients with distinct behaviors will compete to access the shared storage system. The number of I/O resources originally dedicated to this shared storage may provide unsatisfactory performance to the applications in this case. This work proposes a dynamic reconfiguration strategy for the dNFSp file system. This strategy takes into consideration the temporal behavior of distributed applications to launch dedicated I/O resources to the more I/O-demanding applications. The exclusiveness of data servers allows for the isolation of access patterns that cause contention on the system, resulting in improved performance for all executing applications. We developed a tool to monitor the performance of the storage servers in order to launch new servers only when the interaction of I/O phases from running applications cause each other’s performance to drop. These resources are required to a batch scheduler system present on the cluster, allowing the use of computing nodes for temporary data storage. The results show that the proposed metrics lead to the detection of performance saturation on the file system and the start of the dedicated resources, resulting in increased I/O performance. The reconfiguration has also been able to avoid some low performance situations caused by interactions of inactivity periods from temporal applications and by utilization of several I/O processes on the same cluster node.

Diversas aplicações executadas em ambientes de cluster necessitam de uma área de armazenamento permanente com alta capacidade e que forneça uma visão homogênea dos dados entre todos os nós. Esta área compartilhada é comumente implementada através de um sistema de arquivos distribuído, permitindo o acesso através da abstração mais comum para gerenciamento de dados. A disparidade entre poder de processamento e desempenho de dispositivos de armazenamento atuais, no entanto, torna tais sistemas um ponto crítico ao desempenho de aplicações paralelas que lidam com grandes volumes de dados. Ambientes de cluster podem apresentar execução concorrente de aplicações em conjuntos independentes de máquinas. Desta forma, uma grande quantidade de clientes com características distintas farão acessos ao sistema de arquivos compartilhado. Em tais casos, o dimensionamento do sistema de armazenamento distribuído nem sempre poderá prover o desempenho necessário à execução das aplicações com os recursos inicialmente a ele destinados. O presente trabalho propõe uma estratégia de reconfiguração dinâmica para o sistema de arquivos dNFSp. Esta estratégia leva em consideração o comportamento temporal presente em aplicações paralelas para inserir servidores de dados exclusivos a aplicações com alta demanda de I/O. Com a utilização de servidores exclusivos, torna-se possível isolar aplicações com comportamentos que causam grande perda de desempenho no sistema como um todo. Foi desenvolvida uma ferramenta de monitoramento do desempenho junto aos servidores do dNFSp, de forma a solicitar novos servidores apenas quando a interação das fases de I/O das aplicações em execução interferirem no desempenho do sistema. Esses novos servidores são solicitados para um sistema escalonador de recursos para cluster, de forma a utilizar nós livres para o armazenamento de dados. Os resultados mostram que a estratégia proposta é capaz de detectar a saturação do sistema de armazenamento e iniciar os servidores exclusivos, levando a um ganho de desempenho para as aplicações em execução. A reconfiguração dinâmica também mostrou-se capaz de evitar baixos desempenhos causados pela interação dos períodos de inatividade de aplicações temporais e pela utilização de diversos processos de I/O em um único nó. / Several applications executed in cluster environments need a shared storage area with high capacity and a homogeneous view of the stored data to all processing nodes. This area is commonly implemented as a distributed file system, allowing the access to data through the well-known file abstraction. The great gap in performance of processors and storage devices, on the other hand, makes such system a critical point to the performance of parallel applications. A common trait of large cluster environments is the concurrent execution of applications. In this scenario, many clients with distinct behaviors will compete to access the shared storage system. The number of I/O resources originally dedicated to this shared storage may provide unsatisfactory performance to the applications in this case. This work proposes a dynamic reconfiguration strategy for the dNFSp file system. This strategy takes into consideration the temporal behavior of distributed applications to launch dedicated I/O resources to the more I/O-demanding applications. The exclusiveness of data servers allows for the isolation of access patterns that cause contention on the system, resulting in improved performance for all executing applications. We developed a tool to monitor the performance of the storage servers in order to launch new servers only when the interaction of I/O phases from running applications cause each other’s performance to drop. These resources are required to a batch scheduler system present on the cluster, allowing the use of computing nodes for temporary data storage. The results show that the proposed metrics lead to the detection of performance saturation on the file system and the start of the dedicated resources, resulting in increased I/O performance. The reconfiguration has also been able to avoid some low performance situations caused by interactions of inactivity periods from temporal applications and by utilization of several I/O processes on the same cluster node.

Historically, cryptographic file systems have been several times slower than non-cryptographic file systems. This paper describes the design and implementation of a fast Cryptographic Log-structured File System on OpenBSD. We experimentally demonstrate that our pro-totype file system performs close to the Fast File System (FFS) and the Log-structured File System (LFS). To increase performance, our file system performs most encryption and decryption work during disk read and write operations. This is possible thanks to the SEAL encryption algorithm, a software optimized stream cipher that allows the en-cryption work to be performed prior to the actual data is available. We believe that our cryptographic file system design is ideal for optimal read and write performance on locally stored confidential data. / Denna uppsats beskriver utvecklingen av ett kryptografiskt log-strukturerat filsystem och vi visar genom experiment att dess prestanda är jämförbar med lokala filsystem. / Karl Knutsson Skiftesgatan 40 332 35 Gislaved Sweden

International Telemetering Conference Proceedings / October 20-23, 2003 / Riviera Hotel and Convention Center, Las Vegas, Nevada / There is a gulf that separates measurement and telemetry applications from the full benefits of Internet style communication. Whereas the Web provides ubiquitous infrastructure for the distribution of file-based “static” data, there is no general Web solution for real-time streaming data. At best, there are proprietary products that target consumer multimedia and resort to custom point-to-point data connections. This paper considers an extension of the static file paradigm to a dynamic file and introduces a streaming data solution integrated with the existing file-based infrastructure of the Web. The solution approach appears to maximize platform and application independence leading to improved application interoperability potential for large or complex measurement and telemetry networks.