HotStorage '13 Workshop Program

While the world's stored data is growing exponentially—doubling roughly every two years—hardware grows exponentially cheaper. Long term up-front storage acquisition practices create prodigious idle capacity purchased at premium prices. Scale-out storage promises to address this problem by turning static, monolithic systems into growing, distributed, systems. Scale-out must clear many hurdles to become a seamless replacement for all monolithic storage, and peta-scale scale-out storage systems create new data manageability challenges. 

Peter Godman brings 20 years of industry systems experience to Qumulo. As VP of Engineering and CEO of Corensic, Peter brought the world's first thin-hypervisor based product to market. As Director of Software Engineering at Isilon, he led development of several major releases of Isilon's award-winning OneFS distributed filesystem and was inventor of 18 patented technologies. Peter studied math and computer science at MIT.

File systems do not properly isolate faults that occur within them. As a result, a single fault may affect multiple clients adversely, making the entire file system unavailable. We introduce a new file system abstraction, called file pod, to allow applications tomanage failure and recovery polices explicitly for a group of files. Based on this abstraction, we propose the isolation file system, which provides fine-grained fault isolation and quick recovery.

Mobile applications are becoming increasingly datacentric — often relying on cloud services to store, share, and analyze data. App developers have to frequently manage the local storage on the device (e.g., SQLite databases, file systems), as well as data synchronization with cloud services. Developers have to address common issues such as data packaging, handling network failures, supporting disconnected operations, propagating changes, and detecting and resolving conflicts. To free mobile developers from this burden, we are building Simba , a platform to rapidly develop and deploy datacentric mobile apps. Simba provides a unified storage and synchronization API for both structured data and unstructured objects. Apps can specify a data model spanning tables and objects, and atomically sync such data with the cloud without worrying about network disruptions. Simba is also frugal in consuming network resources.

We propose a new approach, called dynamic program and erase scaling (DPES), for improving the endurance of NAND flash memory. The DPES approach is based on our key finding that the NAND endurance is dependent on the erase voltage as well as the number of P/E cycles. Since the NAND endurance has a near-linear dependence on the erase voltage, lowering the erase voltage is an effective way of improving the NAND endurance. By modifying NAND chips to support multiple write modes with different erase voltages, DPES enables a flash software to exploit the new tradeoff between the NAND endurance and write speed. In this paper, we present a novel NAND endurance model which accurately captures the tradeoff relationship between the endurance and write speed under dynamic program and erase scaling. Based on our NAND endurance model, we have implemented the first DPES-aware FTL, called autoFTL, which improves the NAND endurance with a negligible degradation in the overall write throughput. Our experimental results using various I/O traces show that autoFTL can improve the maximum number of P/E cycles by 45% over an existing DPES-unaware FTL with less than 0.2% decrease in the overall write throughput.

Flash memory has been an active topic of research in recent years, but hard information about the parameters and behavior of both flash chips and SSDs has been difficult to obtain for those outside of the industry. In this paper several misconceptions found in the literature are addressed, in order to enable future researchers to avoid some of the errors found in prior work.

We examine the following topics: flash device parameters such as page and erase block size, speed, and reliability, as well as flash translation layer (FTL) requirements and behavior under random and sequential I/O. We have endeavored to find public sources for our claims, and provide experimental evidence in several cases. In doing so, we provide previously unpublished results showing the viability of random writes on commodity SSDs when restricted to a sufficiently small portion of the logical address space.

Parity protection at system level is typically employed to compose reliable storage systems. However, careful consideration is required when SSD based systems employ parity protection. First, additional writes are required for parity updates. Second, parity consumes space on the device, which results in write amplification from less efficient garbage collection at higher space utilization.

This paper analyzes the effectiveness of SSD based RAID and discusses the potential benefits and drawbacks in terms of reliability. A Markov model is presented to estimate the lifetime of SSD based RAID systems in different environments. In a single array, our preliminary results show that parity protection provides benefit only with considerably low space utilizations and low data access rates. However, in a large system, RAID improves data lifetime even when we take write amplification into account.

As disk manufacturers compete to build ever larger and cheaper disks, the possibility of RAID failures becomes more significant for larger and larger disk arrays, creating opportunities for products beyond RAID 6. In this paper, we present the design and implementation of RAIDq, a software-friendly, multiple-parity RAID. RAIDq uses a linear code with efficient encoding and decoding algorithms and addresses a wide range of general cases of RAID that are of practical interest. However, RAIDq does have a limit on how many data disks it can support, which we will analyze in this paper. A second benefit of RAIDq is that it includes existing RAID 5 and 6 as special cases and hence is 100% backward compatible. This allows RAIDq to reuse the efficient coding algorithms and implementations of RAID 5 and 6. Last but not least, RAIDq is optimized for software implementation, as its encoding only involves simple XOR and multiplication by several fixed elements in a finite field. Thanks to the popularity of RAID 6, such operations have been highly optimized on modern processors, of which RAIDq can take advantage, as corroborated by our experiment results.

In this panel, we will discuss the very controversial and timely topic of software-defined storage. What does it mean? How do we get it? Do we already have it?

Given the adoption of virtualization and the recent advent of software-defined networking, many people believe that the remaining technology needed to construct a truly agile datacenter is software-defined storage. However, every vendor has its own definition of what it means and how to construct it. Some even claim to already have it. In this panel discussion we hope to shed some light on this topic and elaborate on the requirements and challenges to construct a storage layer for the next generation of datacenters.

Physical level backups offer increased performance in terms of throughput and scalability as compared to logical backup models, while still maintaining logical consistency. As the trend toward virtualization grows, virtual machine backups (a form of physical backup) are even more important, while becoming easier to perform. The downside is that physical backup generally requires more storage, because of file system meta-data and unallocated blocks. Deduplication is becoming widely accepted and many believe that it will favor logical backup, but this has not been well studied and the relative cost of physical vs. logical on deduplicating storage is not known. In this paper, we take a data-driven approach using user data to quantify the storage costs and contributing factors of physical backups over numerous generations. Based on our analysis, we show how physical backups can be as storage efficient as logical backups, while also giving good backup performance.

In a virtualized cloud cluster, frequent snapshot backup of virtual disks improves hosting reliability; however, it takes significant memory resource to detect and remove duplicated content blocks among snapshots. This paper presents a low-cost deduplication solution scalable for a large number of virtual machines. The key idea is to separate duplicate detection from the actual storage backup instead of using inline deduplication, and partition global index and detection requests among machines using fingerprint values. Then each machine conducts duplicate detection partition by partition independently with minimal memory usage. Another optimization is to allocate and control buffer space for exchanging detection requests and duplicate summaries among machines. Our evaluation shows that the proposed multi-stage scheme uses a small amount of memory while delivering a satisfactory backup throughput.

The PCI Express Solid State Disks (PCIe SSDs) blur the difference between block and memory access semantic devices. Since these SSDs leverage PCIe bus as storage interface, their interfaces are different from conventional memory system interconnects as well as thin storage interfaces. This leads to a new SSD architecture and storage software stack design. Unfortunately, there are not many studies focusing on the system characteristics of these emerging PCIe SSD platforms. In this paper, we quantitatively analyze the challenges faced by PCIe SSDs in getting flash memory closer to CPU and study two representative PCIe SSD architectures (from-scratch SSD and bridge-based SSD) using state-of-the-art real SSDs from two different vendors. Our experimental analysis reveals that 1) while the from-scratch SSD approach offers remarkable performance improvements, it requires enormous host-side memory and computation resources which may not be acceptable in many computing systems; 2) the performance of the from-scratch SSD significantly degrades in a multi-core system; 3) redundant flash software and controllers should be eliminated from the bridge-based SSD architecture; and 4) latency of PCIe SSDs significantly degrade with their storage-level queueing mechanism. Finally, we discuss system implications including potential PCIe SSD applications such as all-flash array.

Numeric time series data has unique storage requirements and access patterns that can benefit from specialized support, given its importance in Big Data analyses. Popular frameworks and databases focus on addressing other needs, making them a suboptimal fit. This paper describes the support needed for numeric time series, suggests an architecture for efficient time series storage, and illustrates its potential for satisfying key requirements.