Sequential Access Patterns Research Articles

An Efficient Block Address Transformation Scheme in Block Layer for Flash-Based SSDs

Flash-based solid state drives (SSDs) are widely adopted in both industry and the academia since SSDs offer higher performance, lower latency, and lower power consumption compared with traditional hard disk drives (HDDs). Unfortunately, the performance of SSDs can be affected by I/O access patterns. For example, random I/O operation degrades the SSD performance compared with sequential I/O operation since the random I/O operation reduces the spatial locality and increases garbage collection (GC) overhead. To handle this issue, in this article, we propose an efficient block address transformation scheme in the block layer to improve both performance and portability. To do this, we first transform random access patterns to sequential access patterns by sequentializing the block addresses in the block layer. Second, we devise a mapping table for managing transformed block addresses. Third, we support correct read operations and transaction processing for updating the mapping table to avoid sacrificing the consistency. Finally, we provide a cleaning scheme to reclaim invalid data generated by the transformed sequential access. This scheme increases spatial locality, reduces GC overhead, and operates well on any file systems or devices. Experimental results show the proposed scheme improves performance by up to 2x, 1.86x, 2.15x, and 42.8% compared with existing scheme in the case of diverse file systems such as EXT4, XFS, BTRFS, and F2FS, respectively.

Access pattern-based high-performance main memory system for graph processing on single machines

With the increasing complexity of graph structures, the current real-world large-scale graphs are being represented by a considerable amount of vertex and edge data. Furthermore, the analysis of a large number of computing nodes has become a very complicated job that requires a large amount of hardware resources. Moreover, in large-scale graph processing, the vertex and edge data show random and sequential memory access patterns at the same time, and this is a major bottleneck in graph processing. In this paper, we present a high-capacity main memory system with an intelligent pattern-aware prefetching engine to overcome the scalability problem and the memory inefficiency of single-machine graph processing. The proposed intelligent pattern-aware prefetching engine is designed to predict and handle sequential or regular patterns and random-access patterns simultaneously. Experimental results demonstrated that the proposed model exhibited performance improvements of 60% over conventional DRAM models, approximately 40% over the existing prefetch models, and about 12.5% over the latest prefetch models.

An efficient cache management algorithm for streaming workloads

In this paper, we present a novel cache management algorithm for real-world streaming workloads. Streaming workloads are believed to exhibit very large and sequential access patterns, which has been the main consideration in designing media caching algorithms. However, legacy caching algorithms do not fully utilise fine-grained access patterns of streaming workloads and also tend to ignore human interactivity. In this paper, we present the least expectation first (LEF) algorithm, which manages a large number of block caches as two-level grouping. Specifically, we select caching and eviction targets based on the expected gain of the cached data blocks, thereby improving the cache hit ratio significantly. Experimental results show that the proposed algorithm performs better than well-known interval caching and LRU algorithms with respect to the hit ratio and the I/O bandwidth.

An efficient cache management algorithm for streaming workloads

In this paper, we present a novel cache management algorithm for real-world streaming workloads. Streaming workloads are believed to exhibit very large and sequential access patterns, which has been the main consideration in designing media caching algorithms. However, legacy caching algorithms do not fully utilise fine-grained access patterns of streaming workloads and also tend to ignore human interactivity. In this paper, we present the least expectation first (LEF) algorithm, which manages a large number of block caches as two-level grouping. Specifically, we select caching and eviction targets based on the expected gain of the cached data blocks, thereby improving the cache hit ratio significantly. Experimental results show that the proposed algorithm performs better than well-known interval caching and LRU algorithms with respect to the hit ratio and the I/O bandwidth.

Mapping Datasets to Object Storage System

Access libraries such as ROOT[1] and HDF5[2] allow users to interact with datasets using high level abstractions, like coordinate systems and associated slicing operations. Unfortunately, the implementations of access libraries are based on outdated assumptions about storage systems interfaces and are generally unable to fully benefit from modern fast storage devices. For example, access libraries often implement buffering and data layout that assume that large, single-threaded sequential access patterns are causing less overall latency than small parallel random access: while this is true for spinning media, it is not true for flash media. The situation is getting worse with rapidly evolving storage devices such as non-volatile memory and ever larger datasets. This project explores distributed dataset mapping infrastructures that can integrate and scale out existing access libraries using Ceph’s extensible object model, avoiding re-implementation or even modifications of these access libraries as much as possible. These programmable storage extensions coupled with our distributed dataset mapping techniques enable: 1) access library operations to be offloaded to storage system servers, 2) the independent evolution of access libraries and storage systems and 3) fully leveraging of the existing load balancing, elasticity, and failure management of distributed storage systems like Ceph. They also create more opportunities to conduct storage server-local optimizations specific to storage servers. For example, storage servers might include local key/value stores combined with chunk stores that require different optimizations than a local file system. As storage servers evolve to support new storage devices like non-volatile memory, these server-local optimizations can be implemented while minimizing disruptions to applications. We will report progress on the means by which distributed dataset mapping can be abstracted over particular access libraries, including access libraries for ROOT data, and how we address some of the challenges revolving around data partitioning and composability of access operations.

An Efficient MapReduce-Based Parallel Processing Framework for User-Based Collaborative Filtering

User-based collaborative filtering is one of the most-used methods for the recommender systems. However, it takes time to perform the method because it requires a full scan of the entire data to find the neighboring users of each active user, who have similar rating patterns. It also requires time-consuming computations because of the complexity of the algorithms. Furthermore, the amount of rating data in the recommender systems grows rapidly, as the number of users, items, and their rating activities tend to increase. Thus, a big data framework with parallel processing, such as Hadoop, is needed for the recommender systems. There are already many research studies on the MapReduce-based parallel processing method for collaborative filtering. However, most of the research studies have not considered the sequential-access restriction for executing MapReduce jobs and the minimization of the required full scan on the entire data on the Hadoop Distributed File System (HDFS), because HDFS sequentially access data on the disk. In this paper, we introduce an efficient MapReduce-based parallel processing framework for collaborative filtering method that requires only a one-time parallelized full scan, while adhering to the sequential access patterns on Hadoop data nodes. Our proposed framework contains a novel MapReduce framework, including a partial computation framework for calculating the predictions and finding the recommended items for an active user with such a one-way parallelized scan. Lastly, we have used the MovieLens dataset to show the validity of our proposed method, mainly in terms of the efficiency of the parallelized method.

An efficient method for batch updates in OPTICS cluster ordering

DBSCAN is one of the popular density-based clustering algorithms, but requires re-clustering the entire data when the input parameters are changed. OPTICS overcomes this limitation. In this paper, we propose a batch-wise incremental OPTICS algorithm which performs efficient insertion and deletion of a batch of points in a hierarchical cluster ordering, which is the output of OPTICS. Only a couple of algorithms are available in the literature on incremental versions of OPTICS. This can be attributed to the sequential access patterns of OPTICS. The existing incremental algorithms address the problem of incrementally updating the hierarchical cluster ordering for point-wise insertion/deletion, but these algorithms are only good for infrequent updates. The proposed incremental OPTICS algorithm performs batch-wise insertions/deletions and is suitable for frequent updates. It produces exactly the same hierarchical cluster ordering as that of classical OPTICS. Real datasets have been used for experimental evaluation of the proposed algorithm and results show remarkable performance improvement over the classical and other existing incremental OPTICS algorithms.

Embedded DRAM-Based Memory Customization for Low-Cost FFT Processor Design

In this paper, we present embedded dynamic random access memory (eDRAM)-based memory customization techniques for low-cost fast Fourier transform (FFT) processor design. The main idea is based on the observation that the FFT processor has regular and predictable memory access patterns, and it can be efficiently exploited for memory customization using eDRAM. The memory customization approaches are applied to both of the pipelined and memory-based FFT architectures. In the pipelined architecture, the read wordline (RWL) coupling write assist and data packing schemes are employed to reduce the redundant RWL and wordline driving, respectively, in column-interleaved memory arrays. The memory address decoder is also simplified with thermometer code by exploiting the sequential access patterns. For the memory-based architecture, the modified cached-memory structure is employed in addition to the techniques used in the pipelined FFT architecture. The hardware implementation results of 2k-point FFT with a 0.11- $\mu {\mathrm{ m}}$ CMOS technology show that the proposed eDRAM-based pipelined and cached-memory FFTs achieve 26.8% and 33.2% power savings over the static RAM-based FFT design, respectively.

A hybrid recommender system for e-learning based on context awareness and sequential pattern mining

The rapid evolution of the Internet has resulted in the availability of huge volumes of online learning resources on the web. However, many learners encounter difficulties in retrieval of suitable online learning resources due to information overload. Besides, different learners have different learning needs arising from their differences in learner’s context and sequential access pattern behavior. Traditional recommender systems such as content based and collaborative filtering (CF) use content features and ratings, respectively, to generate recommendations for learners. However, for accurate and personalized recommendation of learning resources, learner’s context and sequential access patterns should be incorporated into the recommender system. Traditional recommendation techniques do not incorporate the learner’s context and sequential access patterns in computing learner similarities and providing recommendations; hence, they are likely to generate inaccurate recommendations. Furthermore, traditional recommender systems provide unreliable recommendations in cases of high rating sparsity. In this paper, we propose a hybrid recommendation approach combining context awareness, sequential pattern mining (SPM) and CF algorithms for recommending learning resources to the learners. In our recommendation approach, context awareness is used to incorporate contextual information about the learner such as knowledge level and learning goals; SPM algorithm is used to mine the web logs and discover the learner’s sequential access patterns; and CF computes predictions and generates recommendations for the target learner based on contextualized data and learner’s sequential access patterns. Evaluation of our proposed hybrid recommendation approach indicated that it can outperform other recommendation methods in terms of quality and accuracy of recommendations.

A hybrid knowledge-based recommender system for e-learning based on ontology and sequential pattern mining

In recent years, there has been significant growth in the use of online learning resources by learners. However, due to information overload, many learners are experiencing difficulties in retrieving useful and relevant learning resources that meet their learning needs. Although existing recommender systems have recorded significant success in e-commerce domain, they still experience drawbacks in making accurate recommendations of learning resources in e-learning domain due to differences in learner characteristics such as learning style, knowledge level as well as learners’ sequential learning patterns. Most of the existing recommendation techniques do not consider differences in learner characteristics. This problem can be alleviated through incorporation of additional information about the learner into the recommendation process. Furthermore, many recommendation techniques experience cold-start and rating sparsity problems. In this paper, we propose a hybrid knowledge-based recommender system based on ontology and sequential pattern mining (SPM) for recommendation of e-learning resources to learners. In the proposed recommendation approach, ontology is used to model and represent the domain knowledge about the learner and learning resources whereas SPM algorithm discovers the learners’ sequential learning patterns. Our approach involves four steps: (1) creating ontology to represent knowledge about the learner and learning resources, (2) computing ratings similarity based on ontology domain knowledge and making predictions for the target learner, (3) generation of top N learning items by the collaborative filtering recommendation engine, and (4) application of SPM algorithm to the top N learning items to generate the final recommendations for the target learner. A number of experiments were carried out to evaluate the proposed hybrid recommender system and results show improved performance. Furthermore, the proposed hybrid approach can alleviate both the cold-start and data sparsity problems by making use of ontological domain knowledge and learner’s sequential access pattern respectively before the initial data to work on is available in the recommender system.

GPU-based Branchless Distance-Driven Projection and Backprojection.

Projection and backprojection operations are essential in a variety of image reconstruction and physical correction algorithms in CT. The distance-driven (DD) projection and backprojection are widely used for their highly sequential memory access pattern and low arithmetic cost. However, a typical DD implementation has an inner loop that adjusts the calculation depending on the relative position between voxel and detector cell boundaries. The irregularity of the branch behavior makes it inefficient to be implemented on massively parallel computing devices such as graphics processing units (GPUs). Such irregular branch behaviors can be eliminated by factorizing the DD operation as three branchless steps: integration, linear interpolation, and differentiation, all of which are highly amenable to massive vectorization. In this paper, we implement and evaluate a highly parallel branchless DD algorithm for 3D cone beam CT. The algorithm utilizes the texture memory and hardware interpolation on GPUs to achieve fast computational speed. The developed branchless DD algorithm achieved 137-fold speedup for forward projection and 188-fold speedup for backprojection relative to a single-thread CPU implementation. Compared with a state-of-the-art 32-thread CPU implementation, the proposed branchless DD achieved 8-fold acceleration for forward projection and 10-fold acceleration for backprojection. GPU based branchless DD method was evaluated by iterative reconstruction algorithms with both simulation and real datasets. It obtained visually identical images as the CPU reference algorithm.

An integrated prefetching/caching scheme in multimedia servers

Advances in storage and networking technologies have made on-demand multimedia services prevalent these days. To provide numerous concurrent users with such high-quality services, it is essential for storage systems to support sufficient I/O bandwidth for delivering multimedia data on time. Two data buffering techniques, prefetching and interval caching, have improved I/O performance by keeping data blocks in memory for future accesses when handling multimedia data with sequential access patterns. However, they have been addressed separately since it is challenging to determine whether it is better to prefetch or to cache each new stream depending on specific situations. Prefetching too many blocks or caching too long intervals may exhaust the cache space quickly. In this paper, we propose a scheme, called Integrated Prefetching/Caching (IPC), to simultaneously take benefit of both prefetching and interval caching using dynamic threshold values. The IPC schedules incoming streaming requests so that utilization of both cache space and disk bandwidth can be maximized. As a result, the IPC can continue to improve the performance without saturation as system resources are added. By simulation experiments, we show that IPC increases the number of concurrent streams significantly, compared to when either prefetching or caching is employed alone.

Improving I/O Performance Through an In-Kernel Disk Simulator

This paper presents two mechanisms that can significantly improve the I/O performance of both hard and solid-state drives for read operations: KDSim and REDCAP. KDSim is an in-kernel disk simulator that provides a framework for simultaneously simulating the performance obtained by different I/O system mechanisms and algorithms, and for dynamically turning them on and off, or selecting between different options or policies, to improve the overall system performance. REDCAP is a RAM-based disk cache that effectively enlarges the built-in cache present in disk drives. By using KDSim, this cache is dynamically activated/deactivated according to the throughput achieved. Results show that, by using KDSim and REDCAP together, a system can improve its I/O performance up to 88% for workloads with some spatial locality on both hard and solid-state drives, while it achieves the same performance as a ‘regular system’ for workloads with random or sequential access patterns.

New Encoding Method for Low Power Sequential Access ROMs

This paper propose a new ROM data encoding method that takes into account of a sequential access pattern to reduce the power consumption in ROMs used in applications such as FIR filters that access the ROM sequentially. In the proposed encoding method, the number of 1's, of which the increment leads to the increase of the power consumption, is reduced by applying an exclusive-or (XOR) operation to a bit pair composed of two consecutive bits in a bit line. The encoded data can be decoded by using XOR gates and D flip-flops, which are usually used in digital systems for synchronization and glitch suppression. By applying the proposed encoding method to coefficient ROMs of FIR filters designed by using various design methods, we can achieve average reduction of 43.7% over the unencoded original data in the power consumption, which is larger reduction than those achieved by previous methods.

Graph based Approach for Mining Frequent Sequential Access Patterns of Web pages

The Internet has impacted almost every aspect of our society. Since the number of web sites and web pages has grown rapidly, discovering and understanding web users’ surfing behavior are essential for the development of successful web monitoring and recommendation systems. To capture users’ web access behavior, one promising approach is web usage mining which discovers interesting and frequent user access patterns from web logs. Sequential Web page Access pattern mining has been a focused theme in data mining research for over a decade with wide range of applications. The aim of discovering frequent sequential access (usage) patterns in Web log data is to obtain information about the navigational behavior of the users. This can be used for advertising purposes, for creating dynamic user profiles etc. We propose a new approach for mining the web usage data by creating graph using web access sequence of sorted web log and mining the useful sequential access pattern.

Mining Sequential Access Pattern with Low Support From Large Pre-Processed Web Logs

Problem statement: To find frequently occurring Sequential patterns f rom web log file on the basis of minimum support provided. We introduced an efficient strategy for discovering Web usage mining is the application of sequential pattern min ing techniques to discover usage patterns from Web data, in order to understand and better serve the n eeds of Web-based applications. Approach: The approaches adopt a divide-and conquer pattern-growth principle. Our proposed method combined tree projection and prefix growth features from pattern- growth category with position coded feature from early-pruning category, all of these features are k ey characteristics of their respective categories, so we consider our proposed method as a pattern growth, early-pruning hybrid algorithm. Results: Our proposed Hybrid algorithm eliminated the need to st ore numerous intermediate WAP trees during mining. Since only the original tree was stored, it drastically cuts off huge memory access costs, which may include disk I/O cost in a virtual memory environment, especially when mining very long sequences with millions of records. Conclusion: An attempt had been made to our approach for improving efficiency. Our proposed method totally e liminates reconstructions of intermediate WAP- trees during mining and considerably reduces execut ion time.

Traversal Caches: A Framework for FPGA Acceleration of Pointer Data Structures

Field-programmable gate arrays (FPGAs) and other reconfigurable computing (RC) devices have been widely shown to have numerous advantages including order of magnitude performance and power improvements compared to microprocessors for some applications. Unfortunately, FPGA usage has largely been limited to applications exhibiting sequential memory access patterns, thereby prohibiting acceleration of important applications with irregular patterns (e.g., pointer-based data structures). In this paper, we present a design pattern for RC application development that serializes irregular data structure traversals online into a traversal cache, which allows the corresponding data to be efficiently streamed to the FPGA. The paper presents a generalized framework that benefits applications with repeated traversals, which we show can achieve between 7x and 29x speedup over pointer-based software. For applications without strictly repeated traversals, we present application-specialized extensions that benefit applications with highly similar traversals by exploiting similarity to improve memory bandwidth and execute multiple traversals in parallel. We show that these extensions can achieve a speedup between 11x and 70x on a Virtex4 LX100 for Barnes-Hut n-body simulation.

Principles of Holism for sequential twig pattern matching

Modern applications face the challenge of dealing with structured and semi-structured data. They have to deal with complex objects, most of them presenting some kind of internal structure, which often forms a hierarchy. Though XML documents are the most known, chemical compounds, CAD drawings, web-sites and many other applications have to deal with similar problems. In such environments, ordered and unordered tree pattern matching are the fundamental search operations. One of the main thrusts of research activities for tree pattern matching is the class of holistic approaches. Their ultimate goal is to evaluate a query twig as a whole by relying on sequential access patterns and non trivial auxiliary storage structures, typically stored in main memory. Based on the pre/post-order ranks of individual tree nodes, we establish strong theoretical bases as a foundation for correct and efficient holistic pattern matching algorithms. In particular, we define and prove sufficient and necessary conditions to minimize the amount of data retained in memory, thus introducing a correct and complete framework on which different holistic solutions can be compared. We also show how these rules can be applied for building algorithms for ordered and unordered tree-pattern matching. Thanks to the above theoretical achievements, each holistic algorithm gains in efficiency as it is directly implemented on the adopted numbering scheme, avoids expensive matching refinements and keeps memory requirements stable. An experimental analysis and comparison with previous approaches confirms the superiority of our approach tested on synthetic as well as real-life data sets.

RPP: Reference Pattern Based Kernel Prefetching Controller

Conventional kernel prefetching schemes have focused on taking advantage of sequential access patterns that are easy to detect. However, it is observed that, on random and even sequential references, they may cause performance degradation due to inaccurate pattern prediction and overshooting. To address these problems, we propose a novel approach to work with existing kernel prefetching schemes, called Reference Pattern based kernel Prefetching (RPP). The RPP can reduce negative effects of existing schemes by identifying one more reference pattern, i.e., looping, in addition to random and sequential patterns and delaying starting prefetching until patterns are confirmed to be sequential or looping.

Time-aware prefetching for on-demand video services in a residential service gateway

With the recent advances in network technology, the number of high-speed networked homes increases rapidly and the enhanced services such as on-demand video services become feasible in terms of market maturity. Another trend is that storage systems become network-accessible. One of the leading network-attached storage systems is the Fiber Channel Arbitrated Loop (FC-AL). As a residential service gateway, the FC-AL-based servers can stably provide high quality video (e.g., DVD quality MPEG-2 stream) with thousands of clients between external service providers and local clients. In addition, in densely populated areas such as New York City, they can be much more cost efficient. Using our end-to-end simulation experiments to combine all the components, we have observed that FC-AL-based streaming servers perform better than SCSI-based systems, but there is still room for performance improvement. We are motivated by the fact that, unlike in SCSI-based systems, all the disks in FC-AL-based severs utilize only a small portion of their caches to a similar degree due to FC-AL fairness arbitration algorithm. Thus, we propose an effective prefetching scheme to improve the performance by further utilizing the disk cache. We show how the proposed scheme can determine the maximum number of prefetched blocks depending on the disk block and cache size. It is also shown how to find the optimal number of blocks transmitted to the FC-AL from the disk cache per FC-AL arbitration. In addition, we describe the cache replacement policy to take full advantage of the sequential access pattern of video files, and explain how to support multiple loops. By analysis and simulation experiments, we show that our prefetching scheme is not able to only increase the total number of concurrent streams significantly by reducing the disk seek time, but it can also further utilize the FC-AL by reducing the overhead of arbitration.