Search

Complex storage stacks providing data compression, indexing, and analytics help leverage the massive amounts of data generated today to derive insights. It is challenging to perform this computation, however, while fully utilizing the underlying storage media. This is because, while storage servers with large core counts are widely available, single-core performance and memory bandwidth per core grow slower than the core count per die. Computational storage offers a promising solution to this problem by utilizing dedicated compute resources along the storage processing path. We present DeltaFS Indexed Massive Directories (IMDs), a new approach to computational storage. DeltaFS IMDs harvest available (i.e., not dedicated) compute, memory, and network resources on the compute nodes of an application to perform computation on data. We demonstrate the efficiency of DeltaFS IMDs by using them to dynamically reorganize the output of a real-world simulation application across 131,072 CPU cores. DeltaFS IMDs speed up reads by 1,740× while only slightly slowing down the writing of data during simulation I/O for in situ data processing.

Fail-slow hardware is an under-studied failure mode. We present a study of 114 reports of fail-slow hardware incidents, collected from large-scale cluster deployments in 14 institutions. We show that all hardware types such as disk, SSD, CPU, memory, and network components can exhibit performance faults. We made several important observations such as faults convert from one form to another, the cascading root causes and impacts can be long, and fail-slow faults can have varying symptoms. From this study, we make suggestions to vendors, operators, and systems designers.

We are approaching a point in time when it will be infeasible to catalog and query data after it has been generated. This trend has fueled research on in-situ data processing (i.e. operating on data as it is streamed to storage). One important example of this approach is in-situ data indexing. Prior work has shown the feasibility of indexing at scale as a two-step process. First, one partitions data by key across the CPU cores of a parallel job. Then each core indexes its subset as data is persisted. Online partitioning requires transferring data over the network so that it can be indexed and stored by the core responsible for the data. This approach is becoming increasingly costly as new computing platforms emphasize parallelism instead of individual core performance that is crucial for communication libraries and systems software in general. In addition to indexing, scalable online data partitioning is also useful in other contexts such as load balancing and efficient compression.We present FilterKV, an efficient data management scheme for fast online data partitioning of key-value (KV) pairs. FilterKV reduces the total amount of data sent over the network and to storage. We achieve this by: (a) partitioning pointers to KV pairs instead of the KV pairs themselves and (b) using a compact format to represent and store KV pointers. Results from LANL show that FilterKV can reduce total write slowdown (including partitioning overhead) by up to 3x across 4096 CPU cores.

Analysis of large-scale simulation output is a core element of scientific inquiry, but analysis queries may experience significant I/O overhead when the data is not structured for efficient retrieval. While in-situ processing allows for improved time-to-insight for many applications, scaling in-situ frameworks to hundreds of thousands of cores can be difficult in practice. The DeltaFS in-situ indexing is a new approach for in-situ processing of massive amounts of data to achieve efficient point and small-range queries. This paper describes the challenges and lessons learned when scaling this in-situ processing function to hundreds of thousands of cores. We propose techniques for scalable all-to-all communication that is memory and bandwidth efficient, concurrent indexing, and specialized LSM-Tree formats. Combining these techniques allows DeltaFS to control the cost of in-situ processing while maintaining 3 orders of magnitude query speedup when scaling alongside the popular VPIC particle-in-cell code to 131,072 cores.

Checkpointing is the predominant storage driver in today's petascale supercomputers and is expected to remain as such in tomorrow's exascale supercomputers. Users typically prefer to checkpoint into a shared file yet parallel file systems often perform poorly for shared file writing. A powerful technique to address this problem is to transparently transform shared file writing into many exclusively written as is done in ADIOS and PLFS. Unfortunately, the metadata to reconstruct the fragments into the original file grows with the number of writers. As such, the current approach cannot scale to exaflop supercomputers due to the large overhead of creating and reassembling the metadata. In this paper, we develop and evaluate algorithms by which patterns in the PLFS metadata can be discovered and then used to replace the current metadata. Our evaluation shows that these patterns reduce the size of the metadata by several orders of magnitude, increase the performance of writes by up to 40 percent, and the performance of reads by up to 480 percent. This contribution therefore can allow current checkpointing models to survive the transition from petato exascale.

In this paper we introduce the Indexed Massive Directory, a new technique for indexing data within DeltaFS. With its design as a scalable, server-less file system for HPC platforms, DeltaFS scales file system metadata performance with application scale. The Indexed Massive Directory is a novel extension to the DeltaFS data plane, enabling in-situ indexing of massive amounts of data written to a single directory simultaneously, and in an arbitrarily large number of files. We achieve this through a memory-efficient indexing mechanism for reordering and indexing data, and a log-structured storage layout to pack small writes into large log objects, all while ensuring compute node resources are used frugally. We demonstrate the efficiency of this indexing mechanism through VPIC, a widely-used simulation code that scales to trillions of particles. With DeltaFS, we modify VPIC to create a file for each particle to receive writes of that particle's output data. Dynamically indexing the directory's underlying storage allows us to achieve a 5000x speedup in single particle trajectory queries, which require reading all data for a single particle. This speedup increases with application scale while the overhead is fixed at 3% of available memory.

Many computing sites, LANL being one of them, have a requirement for long-term retention of mostly cold data. Although the main function of this storage tier is capacity, it does also have a bandwidth requirement. For many years, tape was the best economic solution for this requirement. However, over time, data sets have grown larger more quickly than tape bandwidth has improved. We have now entered a regime in which disk is the more economically efficient medium for this storage tier. Also more and more, data dominates the computing world. There is a "sea" of data out there in many different formats such as file, object, and Key-value that needs to be efficiently managed and effectively used. In this paper, we introduce a new hybrid storage system named MarFS. MarFS is a Near-POSIX File System using scale-out commercial/cloud for data and many POSIX file systems for metadata services. MarFS is an approach to support a data lake for HPC that sits on industry based commodity storage hardware and is a software layer that provides a global namespace and near POSIX semantics. MarFS provides the capability to serve as an umbrella over a variety of underling storage layers. In this paper, we present the system architecture of the proposed MarFS near-POISX file system, we conduct early functional performance testing cases on MarFS's software components, and finally we address current deployment status and future development works of the MarFS.

Power is becoming an increasingly important concern for large supercomputer centers. However, to date, there have been a dearth of studies of power usage 'in the wild'-on production supercomputers running production workloads. In this paper, we present the initial results of a project to characterize the power usage of the three Top500 supercomputers at Los Alamos National Laboratory: Cielo, Roadrunner, and Luna #15, #19, and #47, respectively, on the June 2012 Top500 list. Power measurements taken both at the switchboard level and within the compute racks are presented and discussed. Some noteworthy results of this study are that 1 variability in power consumption differs across architectures, even when running a similar workload and 2 Los Alamos National Laboratory's scientific workload draws, on average, only 70-75% of LINPACK power and only 40-55% of nameplate power, implying that power capping may enable a substantial reduction in power and cooling infrastructure while impacting comparatively few applications. Copyright © 2013 John Wiley & Sons, Ltd.

PurposeThe purpose of this paper is to describe a process for extracting and matching author names from large collections of bibliographic metadata using the Hadoop implementation of MapReduce. It considers the challenges and risks associated with name matching on such a large‐scale and proposes simple matching heuristics for the reduce process. The resulting semantic graphs of authors link names to publications, and include additional features such as phonetic representations of author last names. The authors believe that this achieves an appropriate level of matching at scale, and enables further matching to be performed with graph analysis tools.Design/methodology/approachA topically‐focused collection of metadata records describing peer‐reviewed papers was generated based upon a search. The matching records were harvested and stored in the Hadoop Distributed File System (HDFS) for processing by hadoop. A MapReduce job was written to perform coarse‐grain author name matching, and multiple papers were matched with authors when the names were very similar or identical. Semantic graphs were generated so that the graphs could be analyzed to perform finer grained matching, for example by using other metadata such as subject headings.FindingsWhen performing author name matching at scale using MapReduce, the heuristics that determine whether names match should be limited to the rules that yield the most reliable results for matching. Bad rules will result in lots of errors, at scale. MapReduce can also be used to generate or extract other data that might help resolve similar names when stricter rules fail to do so. The authors also found that matching is more reliable within a well‐defined topic domain.Originality/valueLibraries have some of the same big data challenges as are found in data‐driven science. Big data tools such as hadoop can be used to explore large metadata collections, and these collections can be used as surrogates for other real world, big data problems. MapReduce activities need to be appropriately scoped so as to yield good results, while keeping an eye out for problems in code which can be magnified in the output from a MapReduce job.

High performance computing fault tolerance depends on scalable parallel file system performance. For more than a decade scalable bandwidth has been available from the object storage systems that underlie modern parallel file systems, and recently we have seen demonstrations of scalable parallel metadata using dynamic partitioning of the namespace over multiple metadata servers. But even these scalable parallel file systems require significant numbers of dedicated servers, and some workloads still experience bottlenecks. We envision exascale parallel file systems that do not have any dedicated server machines. Instead a parallel job instantiates a file system namespace service in client middleware that operates on only scalable object storage and communicates with other jobs by sharing or publishing namespace snapshots. Experiments shows that our serverless file system design, DeltaFS, performs metadata operations orders of magnitude faster than traditional file system architectures.

Erasure code storage systems are becoming popular choices for cloud storage systems due to cost-effective storage space saving schemes and higher fault-resilience capabilities. Both erasure code encoding and decoding procedures are involving heavy array, matrix, and table-lookup compute intensive operations. Multi-core, many-core, and streaming SIMD extension are implemented in modern CPU designs. In this paper, we study the power consumption and energy efficiency of erasure code computing using traditional Intel x86 platform and Intel Streaming SIMD extension platform. We use a breakdown power consumption analysis approach and conduct power studies of erasure code encoding process on various storage devices. We present the impact of various storage devices on erasure code based storage systems in terms of processing time, power utilization, and energy cost. Finally we conclude our studies and demonstrate the Intel x86's Streaming SIMD extensions computing is a cost-effective and favorable choice for future power efficient HPC cloud storage systems.

This document will be organized into sections that are defined by the requirements for a file system that presents a near-POSIX (Portable Operating System Interface) interface to the user, but whose data is stored in whatever form is most efficient for the type of data being stored. After defining the requirement the design for meeting the requirement will be explained. Finally there will be sections on configuring and administering this file system. More and more, data dominates the computing world. There is a “sea” of data out there in many different formats that needs to be managed and used. “Mar” means “sea” in Spanish. Thus, this product is dubbed MarFS, a file system for a sea of data.

Erasure code storage systems are becoming popular choices for cloud storage systems due to cost-effective storage space saving schemes and higher fault-resilience capabilities. Both erasure code encoding and decoding procedures are involving heavy array, matrix, and table-lookup compute intensive operations. Multi-core, many-core, and streaming SIMD extension are implemented in modern CPU designs. In this paper, we study the power consumption and energy efficiency of erasure code computing using traditional Intel x86 platform and Intel Streaming SIMD extension platform. We use a breakdown power consumption analysis approach and conduct power studies of erasure code encoding process on various storage devices. We present the impact of various storage devices on erasure code based storage systems in terms of processing time, power utilization, and energy cost. Finally we conclude our studies and demonstrate the Intel x86's Streaming SIMD extensions computing is a cost-effective and favorable choice for future power efficient HPC cloud storage systems.

For archiving large datasets in high-performance computing facilities, tape technology has a long history of providing inexpensive capacity. However, as the memory-size of supercomputers continues to grow geometrically, the cost of tape bandwidth is becoming more important. The projected costs for tape-drives, robotics, and maintenance, are creating challenges for tape-based archives. The advent of erasure-coded object storage, driven by the "cloud storage" industry, might make it practical to implement archives using disks, or hybrid disk-and-tape systems. We used linear optimization techniques to investigate when and how this transition might best be made, taking into consideration our significant investment in tape technology. Our models introduce a technique to systematically relax constraints on the relationship between tape-capacity and tape-bandwidth, which governs a trade-off between cost and performance. We ran parameter studies that support some preliminary conclusions about paths forward for archive infrastructure at LANL.

Petascale computing infrastructures for scientific discovery make petascale demands on information storage capacity, performance, concurrency, reliability, availability, and manageability.The Petascale Data Storage Institute focuses on the data storage problems found in petascale scientific computing environments, with special attention to community issues such as interoperability, community buy-in, and shared tools.The Petascale Data Storage Institute is a collaboration between researchers at Carnegie Mellon University, National Energy Research Scientific Computing Center, Pacific Northwest National Laboratory, Oak Ridge National Laboratory, Sandia National Laboratory, Los Alamos National Laboratory, University of Michigan, and the University of California at Santa Cruz.

The I/O bottleneck in high-performance computing is becoming worse as application data continues to grow. In this work, we explore how patterns of I/O within these applications can significantly affect the effectiveness of the underlying storage systems and how these same patterns can be utilized to improve many aspects of the I/O stack and mitigate the I/O bottleneck. We offer three main contributions in this paper. First, we develop and evaluate algorithms by which I/O patterns can be efficiently discovered and described. Second, we implement one such algorithm to reduce the metadata quantity in a virtual parallel file system by up to several orders of magnitude, thereby increasing the performance of writes and reads by up to 40 and 480 percent respectively. Third, we build a prototype file system with pattern-aware prefetching and evaluate it to show a 46 percent reduction in I/O latency. Finally, we believe that efficient pattern discovery and description, coupled with the observed predictability of complex patterns within many high-performance applications, offers significant potential to enable many additional I/O optimizations.

HDF5 is a data model, library and file format for storing and managing data. It is designed for flexible and efficient I/O for high volume and complex data. Natively, it uses a single-file format where multiple HDF5 objects are stored in a single file. In a parallel HDF5 application, multiple processes access a single file, thereby resulting in a performance bottleneck in I/O. Additionally, a single-file format does not allow semantic post processing on individual objects outside the scope of the HDF5 application. We have developed a new plugin for HDF5 using its Virtual Object Layer that serves two purposes: 1) it uses PLFS to convert the single-file layout into a data layout that is optimized for the underlying file system, and 2) it stores data in a unique way that enables semantic post-processing on data. We measure the performance of the plugin and discuss work leveraging the new semantic post-processing functionality enabled. We further discuss the applicability of this approach for exascale burst buffer storage systems.

In the petascale era, the storage stack used by the extreme scale high performance computing community is fairly homogeneous across sites. On the compute edge of the stack, file system clients or IO forwarding services direct IO over an interconnect network to a relatively small set of IO nodes. These nodes forward the requests over a secondary storage network to a spindle-based parallel file system. Unfortunately, this architecture will become unviable in the exascale era. As the density growth of disks continues to outpace increases in their rotational speeds, disks are becoming increasingly cost-effective for capacity but decreasingly so for bandwidth. Fortunately, new storage media such as solid state devices are filling this gap; although not cost-effective for capacity, they are so for performance. This suggests that the storage stack at exascale will incorporate solid state storage between the compute nodes and the parallel file systems. There are three natural places into which to position this new storage layer: within the compute nodes, the IO nodes, or the parallel file system. In this paper, we argue that the IO nodes are the appropriate location for HPC workloads and show results from a prototype system that we have built accordingly. Running a pipeline of computational simulation and visualization, we show that our prototype system reduces total time to completion by up to 30%.

The largest-scale high-performance (HPC) systems are stretching parallel file systems to their limits in terms of aggregate bandwidth and numbers of clients. To further sustain the scalability of these file systems, researchers and HPC storage architects are exploring various storage system designs. One proposed storage system design integrates a tier of solid-state burst buffers into the storage system to absorb application I/O requests. In this paper, we simulate and explore this storage system design for use by large-scale HPC systems. First, we examine application I/O patterns on an existing large-scale HPC system to identify common burst patterns. Next, we describe enhancements to the CODES storage system simulator to enable our burst buffer simulations. These enhancements include the integration of a burst buffer model into the I/O forwarding layer of the simulator, the development of an I/O kernel description language and interpreter, the development of a suite of I/O kernels that are derived from observed I/O patterns, and fidelity improvements to the CODES models. We evaluate the I/O performance for a set of multiapplication I/O workloads and burst buffer configurations. We show that burst buffers can accelerate the application perceived throughput to the external storage system and can reduce the amount of external storage bandwidth required to meet a desired application perceived throughput goal.

There yet exist no truly parallel file systems. Those that make the claim fall short when it comes to providing adequate concurrent write performance at large scale. This limitation causes large usability headaches in HPC. Users need two major capabilities missing from current parallel file systems. One, they need low latency interactivity. Two, they need high bandwidth for large parallel IO; this capability must be resistant to IO patterns and should not require tuning. There are no existing parallel file systems which provide these features. Frighteningly, exascale renders these features even less attainable from currently available parallel file systems. Fortunately, there is a path forward.

Exascale will present many challenges to the HPC community, but the primary problem will likely be power consumption. Current petascale systems already use a significant fraction of the power that an exascale system will be allotted. In this paper, we show measurements for real I/O power use in three large systems. We show that I/O power use is proportionally fairly low per machine, between 4.4 and 5.5% of the total consumption. We use these measurements to motivate a burst-buffer checkpointing solution for power-efficient I/O at exascale. We estimated this solution to use approximately 6.6% of the exascale machine power budget, which is on par with today's systems.

Mean Time To failure, MTTF, is a commonly accepted metric for reliability. In this paper we present a novel approach to achieve the desired MTTF with minimum redundancy. We analyze the failure behavior of large scale systems using failure logs collected by Los Alamos National Laboratory. We analyze the root cause of failures and present a choice of specific hardware and software components to be made fault-tolerant, through duplication, to achieve target MTTF at minimum expense. Not all components show similar failure behavior in the systems. Our objective, therefore, was to arrive at an ordering of components to be incrementally selected for protection to achieve a target MTTF. We propose a model for MTTF for tolerating failures in a specific component, system-wide, and order components according to the coverage provided. Systems grouped based on hardware configuration showed similar improvements in MTTF when different components in them were targeted for fault-tolerance.

Present and future Archive Storage Systems have been challenged to (a) scale to very high bandwidths, (b) scale in metadata performance, (c) support policy-based hierarchical storage management capability, (d) scale in supporting changing needs of very large data sets, (e) support standard interface, and (f) utilize commercial-off-the-shelf (COTS) hardware. Parallel file systems have also been demanded to perform the same manner but at one or more orders of magnitude faster in performance. Archive systems continue to improve substantially comparable to file systems in their design due to the need for speed and bandwidth, especially metadata searching speeds such as more caching and less robust semantics. Currently, the number of extreme highly scalable parallel archive solutions is very limited especially for moving a single large striped parallel disk file onto many tapes in parallel. We believe that a hybrid storage approach of using COTS components and an innovative software technology can bring new capabilities into a production environment for the HPC community. This solution is much faster than the approach of creating and maintaining a complete end-to-end unique parallel archive software solution. We relay our experience of integrating a global parallel file system and a standard backup/archive product with an innovative parallel software code to construct a scalable and parallel archive storage system. Our solution has a high degree of overlap with current parallel archive products including (a) doing parallel movement to/from tape for a single large parallel file, (b) hierarchical storage management, (c) ILM features, (d) high volume (non-single parallel file) archives for backup/archive/content management, and (e) leveraging all free file movement tools in Linux such as copy, move, ls, tar, etc. We have successfully applied our working COTS Parallel Archive System to the current world’s first petaflop/s computing system, LANL’s Roadrunner machine, and demonstrated its capability to address requirements of future archival storage systems. Now this new Parallel Archive System is used on the LANL’s Turquoise Network

Parallel applications running across thousands of processors must protect themselves from inevitable system failures. Many applications insulate themselves from failures by checkpointing. For many applications, checkpointing into a shared single file is most convenient. With such an approach, the size of writes are often small and not aligned with file system boundaries. Unfortunately for these applications, this preferred data layout results in pathologically poor performance from the underlying file system which is optimized for large, aligned writes to non-shared files. To address this fundamental mismatch, we have developed a virtual parallel log structured file system, PLFS. PLFS remaps an application’s preferred data layout into one which is optimized for the underlying file system. Through testing on PanFS, Lustre, and GPFS, we have seen that this layer of indirection and reorganization can reduce checkpoint time by an order of magnitude for several important benchmarks and real applications without any application modification.

In this paper we present a cost-effective, high bandwidth server I/O network architecture, named PaScal (Parallel and Scalable). We use the PaScal server I/O network to support data-intensive scientific applications running on very large-scale Linux clusters. PaScal server I/O network architecture provides (1) bi-level data transfer network by combining high speed interconnects for computing inter-process communication (IPC) requirements and low-cost gigabit Ethernet interconnect for global IP based storage/file access, (2) bandwidth on demand I/O network architecture without re-wiring and reconfiguring the system, (3) multi-path routing scheme, (4) reliability improvement through reducing large number of network components in server I/O network, and (5) global storage/file systems support in heterogeneous multi-cluster and grids environments. We have compared the PaScal server I/O network architecture with the federated server I/O network architecture (FESIO). Concurrent MPI-I/O performance testing results and deployment cost comparison demonstrate that the PaScal server I/O network architecture can outperform the FESIO network architecture in many categories: cost-effectiveness, scalability, and manageability and ease of large-scale I/O network.

This paper presents the design and implementation of a new I/O networking infrastructure, named PaScal (parallel and scalable I/O networking framework). PaScal is used to support high data bandwidth IP based global storage systems for large scale Linux clusters. PaScal has several unique properties. It employs (1) Multi-level switch-fabric interconnection network by combining high speed interconnects for computing inter-process communication (IPC) requirements and low-cost Gigabit Ethernet interconnect for global IP based storage/file access, (2) A bandwidth on demand scaling I/O networking architecture, (3) open-standard IP networks (routing and switching), (4) multipath routing for load balancing and failover, (5) open shortest path first (OSPF) routing software, and (6) Supporting a global file system in multi-cluster and multi-platform environments. We describe both the hardware and software components of our proposed PaScal. We have implemented the PaScal I/O infrastructure on several large-size Linux clusters at LANL. We have conducted a sequence of parallel MPI-IO assessment benchmarks on LANL's Pink 1024 node Linux cluster and the Panasas global parallel file system. Performance results from our parallel MPI-IO benchmarks on the Pink cluster demonstrate that the PaScal I/O Infrastructure is robust and capable of scaling in bandwidth on large-size Linux clusters.