Workshop Program

Cloud-of-clouds storage exploits diversity of cloud storage vendors to provide fault tolerance and avoid vendor lock-ins. Its inherent diversity property also enables us to offer keyless data security via dispersal algorithms. However, the keyless security of existing dispersal algorithms relies on the embedded random information, which breaks data deduplication of the dispersed data. To simultaneously enable keyless security and deduplication, we propose a novel dispersal approach called convergent dispersal, which replaces original random information with deterministic cryptographic hash information that is derived from the original data but cannot be inferred by attackers without knowing the whole data. We develop two convergent dispersal algorithms, namely CRSSS and CAONT-RS. Our evaluation shows that CRSSS and CAONT-RS provide complementary performance advantages for different parameter settings.

Modern applications expand to fill the space available to them, exploiting local storage to improve performance by caching, prefetching and precomputing data. In virtualized settings, this behavior compromises storage elasticity owing to a rigid contract between the hypervisor and the guest OS: once space is allocated to a virtual disk and used by an application, it cannot be reclaimed by the hypervisor. In this paper, we propose a new guest filesystem called Harmonium that exploits the ephemeral or derivative nature of application data. Each file in Harmonium optionally has a motif that describes how the file can be reconstructed via computation, network accesses, or operations on other files. Harmonium expands files from their motifs when space is available, and contracts them back to their motifs when it is scarce. Given a target size, the system selects files to expand or contract based on the load on the CPU, network, and storage, as well as expected access patterns. As a result, Harmonium enables elastic cloud storage, allowing the hypervisor to dynamically balance storage across multiple VMs.

The persistent storage options in smartphones employ journaling or double-write to enforce atomicity, consistency and durability, which introduces significant overhead to system performance. Our in-depth examination of the issue leads us to believe that much of the overhead would be unnecessary if we rethink the volatility of memory considering the battery-backed characteristics of DRAM in modern-day smartphones. With this rethinking, we propose quasi Non-Volatile Memory (qNVRAM), a new design that makes the DRAM in smartphones quasi non-volatile, to help remove the performance overhead of enforcing persistency. We assess the feasibility and effectiveness of our design by implementing a persistent page cache in SQLite. Our evaluation on a real Android smartphone shows that qNVRAM speeds up the insert, update and delete transactions by up to 16:33x, 15:86x and 15:76x respectively.

Modern storage systems are becoming more complex, combining different storage technologies with different behaviors. Performance alone is not enough to characterize storage systems: energy efficiency, durability, and more are becoming equally important. We posit that one must evaluate storage systems from a monetary cost perspective as well as performance. We believe that cost should consider the workloads used over the storage systems’ expected lifetime. We designed and developed a versatile hybrid storage system under Linux that combines HDD and SSD. The SSD can be used as cache or as primary storage for hot data. Our system includes tunable parameters to enable trading off performance, energy use, and durability. We built a cost model and evaluated our system under a variety of workloads and parameters, to illustrate the importance of cost evaluations of storage systems.

We propose a radical re-architecture of the traditional operating system storage stack to move the kernel off the data path. Leveraging virtualized I/O hardware for disk and flash storage, most read and write I/O operations go directly to application code. The kernel dynamically allocates extents, manages the virtual to physical binding, and performs name translation. The benefit is to dramatically reduce the CPU overhead of storage operations while improving application flexibility.

High performance storage layer is vital for allowing interactive ad hoc SQL analytics (OLAP style) over Big Data. The paper makes a case for leveraging flash in the Big Data stack to speed up queries. State-of-the-art Big Data layouts and algorithms are optimized for hard disks (i.e., sequential access is emphasized over random access) and result in suboptimal performance on flash given its drastically different performance characteristics. While existing columnar and row-columnar layouts are able to reduce disk IO compared to row-based layouts, they still end up reading significant columnar data irrelevant to the query as they only employ coarse-grained, intra-columnar data skipping which doesn’t work across all queries. FlashQueryFile’s specialized columnar data layouts, selection, and projection algorithms fully exploit fast random accesses and high internal I/O parallelism of flash to allow fast and I/O-efficient query processing and fine-grained, intra-columnar data skipping to minimize data read per query. FlashQueryFile results in 11X-100X TPC-H query speedup and 38%-99.08% reduction in data read compared to flash-based HDD-optimized row-columnar data layout and its associated algorithms.

Despite years of experience, countless implementations, and increased importance in the modern era, many basic facets of storage systems remain problematic. In this talk, I'll highlight two fundamental problems found at the interface between applications and storage, and suggest new directions that help bridge the gap between what applications need and what current storage systems provide.

Remzi Arpaci-Dusseau is a professor of Computer Sciences at the University of Wisconsin-Madison. He and his wife Andrea co-lead a research group that has been active in the systems community for many years; their work has had academic impact (including nine Best Paper awards) and practical impact (including the transaction checksum in Linux ext4 and fast file system checking in FreeBSD). Remzi co-chaired USENIX ATC '04, FAST '07, OSDI '10, and will co-chair SOCC '14. Remzi has won numerous teaching awards and is co-author (with his wife) of a free online operating systems book (available at http://www.ostep.org); chapters of the book have been downloaded over 1 million times in the past few years.

The term sequential I/O is widely used in systems research with the intuitive understanding that it means consecutive access. From a survey of the literature, though, this intuitive understanding has translated into numerous, inconsistent definitions. Since sequential I/O is such a fundamental concept in systems research, we believe that a sequentiality metric should allow us to compare access patterns in a meaningful way.

We explore access properties that could be incorporated into potential metrics for sequential I/O including: access size, gaps between accesses, multi-stream, and inter-arrival time. We then analyze hundreds of largescale storage traces and discuss how potential metrics compare. Interestingly, we find I/O traces considered highly sequential by one metric can be highly random to another metric. We further demonstrate that many plausible metrics are weakly correlated, though metrics weighted by size have more consistency. While there may not be a single metric for sequential I/O that is best in all cases, we believe systems researchers should more carefully consider, and state, which definition they use.

In this paper, we evaluate the efficacy, in a Hadoop setting, of two coding schemes, both possessing an inherent double replication of data. The two coding schemes belong to the class of regenerating and locally regenerating codes respectively, and these two classes are representative of recent advances made in designing codes for the efficient storage of data in a distributed setting. In comparison with triple replication, double replication permits a significant reduction in storage overhead, while delivering good MapReduce performance under moderate work loads. The two coding solutions under evaluation here, add only moderately to the storage overhead of double replication, while simultaneously offering reliability levels similar to that of triple replication.

One might expect from the property of inherent data duplication that the performance of these codes in executing a MapReduce job would be comparable to that of double replication. However, a second feature of this class of code comes into play here, namely that under both coding schemes analyzed here, multiple blocks from the same coded stripe are required to be stored on the same node. This concentration of data belonging to a single stripe negatively impacts MapReduce execution times. However, much of this effect can be undone by simply adding a larger number of processors per node. Further improvements are possible if one tailors the Map task scheduler to the codes under consideration. We present both experimental and simulation results that validate these observations.

This paper makes a case for the multi-streamed solidstate drive (SSD). It offers an intuitive storage interface for the host system to inform the SSD about the expected lifetime of data being written. We show through experimentation with a real multi-streamed SSD prototype that the worst-case update throughput of a Cassandra NoSQL DB system can be improved by nearly 56%. We discuss powerful use cases of the proposed SSD interface.

Solid State Disks (SSDs) have risen to prominence as an I/O accelerator with low power consumption and high energy efficiency. In this paper, we question some common assumptions regarding SSDs’ operating temperature, dynamic power, and energy consumption through extensive empirical analysis. We examine three different real high-end SSDs that respectively employ multiple channels, cores, and flash chips. Our evaluations reveal that dynamic power consumption of many-resource SSD is, on average, 5x and 4x worse than an enterprise-scale SSD and HDD, respectively. This work also addresses SSD overheating problem and power throttling issues, which result in significant performance degradation. Lastly, we offer an evidence that HW/SW optimization studies are needed to improve energy efficiency in future SSDs.