Recent technological advances are causing an explosion in the production of sequential data. Biological sequences, web logs and time series are represented as strings. Currently, strings are stored, managed and queried in an ad-hoc fashion because they lack a standardized data model and query language. String queries are computationally demanding, especially when strings are long and numerous. Existing approaches cannot handle the growing number of strings produced by environmental, healthcare, bioinformatic, and space applications. There is a trade- off between performing analytics efficiently and scaling to thousands of cores to finish in reasonable times. In this thesis, we introduce a data model that unifies the input and output representations of core string operations. We define a declarative query language for strings where operators can be pipelined to form complex queries. A rich set of core string operators is described to support string analytics. We then demonstrate a database system for string analytics based on our model and query language. In particular, we propose the use of a novel data structure augmented by efficient parallel computation to strike a balance between preprocessing overheads and query execution times. Next, we delve into repeated motifs extraction as a core string operation for large-scale string analytics. Motifs are frequent patterns used, for example, to identify biological functionality, periodic trends, or malicious activities. Statistical approaches are fast but inexact while combinatorial methods are sound but slow. We introduce ACME, a combinatorial repeated motifs extractor. We study the spatial and temporal locality of motif extraction and devise a cache-aware search space traversal technique. ACME is the only method that scales to gigabyte- long strings, handles large alphabets, and supports interesting motif types with minimal overhead. While ACME is cache-efficient, it is limited by being serial. We devise a lightweight parallel space traversal technique, called FAST, that enables ACME to scale to thousands of cores. High degree of concurrency is achieved by partition- ing the search space horizontally and balancing the workload among cores with minimal communication overhead. Consequently, complex queries are solved in minutes instead of days. ACME is a versatile system that runs on workstations, clusters, and supercomputers. It is the first to utilize a supercomputer and scale to 16 thousand CPUs. Merely using more cores does not guarantee efficiency, because of the related overheads. To this end, we introduce an automatic tuning mechanism that suggests the appropriate number of cores to meet user constraints in terms of runtime while minimizing the financial cost of cloud resources. Particularly, we study workload frequency distributions then build a model that finds the best problem decomposition and estimates serial and parallel runtimes. Finally, we generalize our automatic tuning method as a general method, called APlug. APlug can be used in other applications and we integrate it with systems for molecular docking and multiple sequence alignment.

The complexity of SPARQL queries and RDF applications poses great challenges on distributed RDF management systems. SPARQL workloads are dynamic and con- sist of queries with variable complexities. Hence, systems that use static partitioning su↵er from communication overhead for workloads that generate excessive communi- cation. Concurrently, RDF applications are becoming more sophisticated, mandating analytical operations that extend beyond SPARQL queries. Being primarily designed and optimized to execute SPARQL queries, which lack procedural capabilities, exist- ing systems are not suitable for rich RDF analytics. This dissertation tackles the problem of accelerating SPARQL queries and RDF analytics on distributed shared-nothing RDF systems. First, a distributed RDF en- gine, coined AdPart, is introduced. AdPart uses lightweight hash partitioning for sharding triples using their subject values; rendering its startup overhead very low. The locality-aware query optimizer of AdPart takes full advantage of the partition- ing to (i) support the fully parallel processing of join patterns on subjects and (ii) minimize data communication for general queries by applying hash distribution of intermediate results instead of broadcasting, wherever possible. By exploiting hash- based locality, AdPart achieves better or comparable performance to systems that employ sophisticated partitioning schemes. To cope with workloads dynamism, AdPart is extended to dynamically adapt to workload changes. AdPart monitors the data access patterns and dynamically redis- tributes and replicates the instances of the most frequent patterns among workers.Consequently, the communication cost for future queries is drastically reduced or even eliminated. Experiments with synthetic and real data verify that AdPart starts faster than all existing systems and gracefully adapts to the query load. Finally, to support and accelerate rich RDF analytical tasks, a vertex-centric RDF analytics framework is proposed. The framework, named SPARTex, bridges the gap between RDF and graph processing. To do so, SPARTex: (i) implements a generic SPARQL operator as a vertex-centric program. The operator is coupled with an optimizer that generates e cient execution plans. (ii) It allows SPARQL to invoke vertex-centric programs as stored procedures. Finally, (iii) it provides a unified in- memory data store that allows the persistence of intermediate results. Consequently, SPARTex can e ciently support RDF analytical tasks consisting of complex pipeline of operators.

This work considers black-box Bayesian inference over high-dimensional parameter spaces. The well-known and widely respected adaptive Metropolis (AM) algorithm is extended herein to asymptotically scale uniformly with respect to the underlying parameter dimension, by respecting the variance, for Gaussian targets. The result- ing algorithm, referred to as the dimension-independent adaptive Metropolis (DIAM) algorithm, also shows improved performance with respect to adaptive Metropolis on non-Gaussian targets. This algorithm is further improved, and the possibility of probing high-dimensional targets is enabled, via GPU-accelerated numerical libraries and periodically synchronized concurrent chains (justified a posteriori). Asymptoti- cally in dimension, this massively parallel dimension-independent adaptive Metropolis (MPDIAM) GPU implementation exhibits a factor of four improvement versus the CPU-based Intel MKL version alone, which is itself already a factor of three improve- ment versus the serial version. The scaling to multiple CPUs and GPUs exhibits a form of strong scaling in terms of the time necessary to reach a certain convergence criterion, through a combination of longer time per sample batch (weak scaling) and yet fewer necessary samples to convergence. This is illustrated by e ciently sampling from several Gaussian and non-Gaussian targets for dimension d 1000.