This paper presents a practical solution to a problem facing high-fan-in, high-bandwidth synchronized TCP workloads in datacenter Ethernets—the TCP incast problem. In these networks, receivers can experience a drastic reduction in application throughput when simultaneously requesting data from many servers using TCP. Inbound data overfills small switch buffers, leading to TCP timeouts lasting hundreds of milliseconds. For many datacenter workloads that have a barrier synchronization requirement (e.g., filesystem reads and parallel data-intensive queries), throughput is reduced by up to 90%. For latency-sensitive applications, TCP timeouts in the datacenter impose delays of hundreds of milliseconds in networks with round-trip-times in microseconds. Our practical solution uses high-resolution timers to enable microsecond-granularity TCP timeouts. We demonstrate that this technique is effective in avoiding TCP incast collapse in simulation and in real-world experiments. We show that eliminating the minimum retransmission timeout bound is safe for all environments, including the wide-area.

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.

Abstract—The Hadoop Distributed File System (HDFS) is designed to store very large data sets reliably, and to stream those data sets at high bandwidth to user applications. In a large cluster, thousands of servers both host directly attached storage and execute user application tasks. By distributing storage and computation across many servers, the resource can grow with demand while remaining economical at every size. We describe the architecture of HDFS and report on experience using HDFS to manage 25 petabytes of enterprise data at Yahoo!.

rose by any other name...

Acknowledgements: We would like to thank several people who made significant contributions in improving this paper. Ruth Klundt put in a significant effort and time to run our experimental evaluation at Sandia National Labs, especially getting it working few days before a deadline; thanks to Lee Ward who offered us Sandia’s resources.

Over the past five years, large-scale storage installations have required fault-protection beyond RAID-5, leading to a flurry of research on and development of erasure codes for multiple disk failures. Numerous open-source implementations of various coding techniques are available to the general public. In this paper, we perform a head-to-head comparison of these implementations in encoding and decoding scenarios. Our goals are to compare codes and implementations, to discern whether theory matches practice, and to demonstrate how parameter selection, especially as it concerns memory, has a significant impact on a code’s performance. Additional benefits are to give storage system designers an idea of what to expect in terms of coding performance when designing their storage systems, and to identify the places where further erasure coding research can have the most impact.

RAID-6 storage systems protect k disks of data with two parity disks so that the system of k + 2 disks may tolerate the failure of any two disks. Coding techniques for RAID-6 systems are varied, but an important class of techniques are those with minimum density, featuring an optimal combination of encoding, decoding and modification complexity. The word size of a code impacts both how the code is laid out on each disk’s sectors and how large k can be. Word sizes which are powers of two are especially important, since they fit precisely into file system blocks. Minimum density codes exist for many word sizes with the notable exception of eight. This paper fills that gap by describing new codes for this important word size. The description includes performance properties as well as details of the discovery process. 1.

As the world moves to digital storage for archival purposes, there is an increasing demand for systems that can provide secure data storage in a cost-effective manner. By identifying common chunks of data both within and between files and storing them only once, deduplication can yield cost savings by increasing the utility of a given amount of storage. Unfortunately, deduplication exploits identical content, while encryption attempts to make all content appear random; the same content encrypted with two different keys results in very different ciphertext. Thus, combining the space efficiency of deduplication with the secrecy aspects of encryption is problematic. We have developed a solution that provides both data security and space efficiency in single-server storage and distributed storage systems. Encryption keys are generated in a consistent manner from the chunk data; thus, identical chunks will always encrypt to the same ciphertext. Furthermore, the keys cannot be deduced from the encrypted chunk data. Since the information each user needs to access and decrypt the chunks that make up a file is encrypted using a key known only to the user, even a full compromise of the system cannot reveal which chunks are used by which users.

Data-intensive file systems, developed for Internet services and popular in cloud computing, provide high reliability and availability by replicating data, typically three copies of everything. Alternatively high performance computing, which has comparable scale, and smaller scale enterprise storage systems get similar tolerance for multiple failures from lower overhead erasure encoding, or RAID, organizations. DiskReduce is a modification of the Hadoop distributed file system (HDFS) enabling asynchronous compression of initially triplicated data down to RAID-class redundancy overheads. In addition to increasing a cluster’s storage capacity as seen by its users by up to a factor of three, DiskReduce can delay encoding long enough to deliver the performance benefits of multiple data copies. 1.

Abstract—Peta-scale scientific applications running on High End Computing (HEC) platforms can generate large volumes of data. For high performance storage and in order to be useful to science end users, such data must be organized in its layout, indexed, sorted, and otherwise manipulated for subsequent data presentation, visualization, and detailed analysis. In addition, scientists desire to gain insights into selected data characteristics ‘hidden ’ or ‘latent ’ in the massive datasets while data is being produced by simulations. PreDatA, short for Preparatory Data Analytics, is an approach for preparing and characterizing data while it is being produced by the large scale simulations running on peta-scale machines. By dedicating additional compute nodes on the peta-scale machine as staging nodes and staging simulation’s output data through these nodes, PreDatA can exploit their computational power to perform selected data manipulations with lower latency than attainable by first moving data into file systems and storage. Such in-transit manipulations are supported by the PreDatA middleware through RDMAbased data movement to reduce write latency, application-specific operations on streaming data that are able to discover latent data characteristics, and appropriate data reorganization and metadata annotation to speed up subsequent data access. As a result, PreDatA enhances the scalability and flexibility of current I/O stack on HEC platforms and is useful for data pre-processing, runtime data analysis and inspection, as well as for data exchange between concurrently running simulation models. Performance evaluations with several production peta-scale applications on Oak Ridge National Laboratory’s Leadership Computing Facility demonstrate the feasibility and advantages of the PreDatA approach. I.