Frequent itemset mining is an important building block in many data mining applications like market basket analysis, recommendation, web-mining, fraud detection, and gene expression analysis. In many of them, the datasets being mined can easily grow up to hundreds of gigabytes or even terabytes of data. Hence, efficient algorithms are required to process such large amounts of data. In recent years, there have been many frequent-itemset mining algorithms proposed, which however (1) often have high memory requirements and (2) do not exploit the large degrees of parallelism provided by modern multiprocessor systems. The high memory requirements arise mainly from inefficient data structures that have only been shown to be sufficient for small datasets. For large datasets, however, the use of these data structures force the algorithms to go out-of-core, i.e., they have to access secondary memory, which leads to serious performance degradations. Exploiting available parallelism is further required to mine large datasets because the serial performance of processors almost stopped increasing. Algorithms should therefore exploit the large number of available threads and also the other kinds of parallelism (e.g., vector instruction sets) besides thread-level parallelism. In this work, we tackle the high memory requirements of frequent itemset mining twofold: we (1) compress the datasets being mined because they must be kept in main memory during several mining invocations and (2) improve existing mining algorithms with memory-efficient data structures. For compressing the datasets, we employ efficient encodings that show a good compression performance on a wide variety of realistic datasets, i.e., the size of the datasets is reduced by up to 6.4x. The encodings can further be applied directly while loading the dataset from disk or network. Since encoding and decoding is repeatedly required for loading and mining the datasets, we reduce its costs by providing parallel encodings that achieve high throughputs for both tasks. For a memory-efficient representation of the mining algorithms’ intermediate data, we propose compact data structures and even employ explicit compression. Both methods together reduce the intermediate data’s size by up to 25x. The smaller memory requirements avoid or delay expensive out-of-core computation when large datasets are mined. For coping with the high parallelism provided by current multiprocessor systems, we identify the performance hot spots and scalability issues of existing frequent-itemset mining algorithms. The hot spots, which form basic building blocks of these algorithms, cover (1) counting the frequency of fixed-length strings, (2) building prefix trees, (3) compressing integer values, and (4) intersecting lists of sorted integer values or bitmaps. For all of them, we discuss how to exploit available parallelism and provide scalable solutions. Furthermore, almost all components of the mining algorithms must be parallelized to keep the sequential fraction of the algorithms as small as possible. We integrate the parallelized building blocks and components into three well-known mining algorithms and further analyze the impact of certain existing optimizations. Our algorithms are already single-threaded often up an order of magnitude faster than existing highly optimized algorithms and further scale almost linear on a large 32-core multiprocessor system. Although our optimizations are intended for frequent-itemset mining algorithms, they can be applied with only minor changes to algorithms that are used for mining of other types of itemsets.

Network-on-chip (NoC) based multi-processor systems-on-chip (MPSoCs) are promising candidates for future multi-processor embedded platforms, which are expected to be composed of hundreds of heterogeneous processing elements (PEs) to potentially provide high performances. However, together with the performances, the systems complexity will increase, and new high level design techniques will be needed to efficiently model, simulate, debug and synthesize them. System-level design (SLD) is considered to be the next frontier in electronic design automation (EDA). It enables the description of embedded systems in terms of abstract functions and interconnected blocks. A promising complementary approach to SLD is the use of models of computation (MoCs) to formally describe the execution semantics of functions and blocks through a set of rules. However, also when this formalization is used, there is no clear way to synthesize system-level models into software (SW) and hardware (HW) towards a NoC-based MPSoC implementation, i.e., there is a lack of system design automation (SDA) techniques to rapidly synthesize and prototype system-level models onto heterogeneous NoC-based MPSoCs. In addition, many of the proposed solutions require large overhead in terms of SW components and memory requirements, resulting in complex and customized multi-processor platforms. In order to tackle the problem, a novel model-based SDA flow has been developed as part of the thesis. It starts from a system-level specification, where functions execute according to the synchronous MoC, and then it can rapidly prototype the system onto an FPGA configured as an heterogeneous NoC-based MPSoC. In the first part of the thesis the HeartBeat model is proposed as a model-based technique which fills the abstraction gap between the abstract system-level representation and its implementation on the multiprocessor prototype. Then details are provided to describe how this technique is automated to rapidly prototype the modeled system on a flexible platform, permitting to adjust the system specification until the designer is satisfied with the results. Finally, the proposed SDA technique is improved defining a methodology to automatically explore possible design alternatives for the modeled system to be implemented on a heterogeneous NoC-based MPSoC. The goal of the exploration is to find an implementation satisfying the designer's requirements, which can be integrated in the proposed SDA flow. Through the proposed SDA flow, the designer is relieved from implementation details and the design time of systems targeting heterogeneous NoC-based MPSoCs on FPGA is significantly reduced. In addition, it reduces possible design errors proposing a completely automated technique for fast prototyping. Compared to other SDA flows, the proposed technique targets a bare-metal solution, avoiding the use of an operating system (OS). This reduces the memory requirements on the FPGA platform comparing to related work targeting MPSoC on FPGA. At the same time, the performance (throughput) of the modeled applications can be increased when the number of processors of the target platform is increased. This is shown through a wide set of case studies implemented on FPGA. / <p>QC 20140609</p>