Big data queries are increasing in complexity and the performance of data analytics is of growing importance. To this end, Big Data on high-performance computing (HPC) infrastructure is becoming a pathway to high-performance data analytics. The state of performance studies on this convergence between Big Data and HPC, however, is limited and ad hoc. A systematic performance study is thus timely and forms the core of this research.
This thesis investigates the challenges involved in developing Big Data applications with significant computations and strict latency guarantees on multicore HPC clusters. Three key areas it considers are thread models, affinity, and communication mechanisms. Thread models discuss the challenges of exploiting intra-node parallelism on modern multicore chips, while affinity looks at data locality and Non-Uniform Memory Access (NUMA) effects. Communication mechanisms investigate the difficulties of Big Data communications. For example, parallel machine learning depends on collective communications, unlike classic scientific simulations, which mostly use neighbor communications. Minimizing this cost while scaling out to higher parallelisms requires non-trivial optimizations, especially when using high-level languages such as Java or Scala. The investigation also includes a discussion on performance implications of different programming models such as dataflow and message passing used in Big Data analytics. The optimizations identified in this research are incorporated in developing the Scalable Parallel Interoperable Data Analytics Library (SPIDAL) in Java, which includes a collection of multidimensional scaling and clustering algorithms optimized to run on HPC clusters.
Besides presenting performance optimizations, this thesis explores a novel scheme for characterizing Big Data benchmarks. Fundamentally, a benchmark evaluates a certain performance-related aspect of a given system. For example, HPC benchmarks such as LINPACK and NAS Parallel Benchmark (NPB) evaluate the floating-point operations (flops) per second through a computational workload. The challenge with Big Data workloads is the diversity of their applications, which makes it impossible to classify them along a single dimension. Convergence Diamonds (CDs) is a multifaceted scheme that identifies four dimensions of Big Data workloads. These dimensions are problem architecture, execution, data source and style, and processing view.
The performance optimizations together with the richness of CDs provide a systematic guide to developing high-performance Big Data benchmarks, specifically targeting data analytics on large, multicore HPC clusters.
|Advisor:||Fox, Geoffrey C.|
|Commitee:||Lumsdaine, Andrew, Qiu, Judy, Tang, Haixu|
|School Location:||United States -- Indiana|
|Source:||DAI-B 78/04(E), Dissertation Abstracts International|
|Subjects:||Information science, Computer science|
|Keywords:||Big data, HPC, Java, Multicore, Performance|
Copyright in each Dissertation and Thesis is retained by the author. All Rights Reserved
The supplemental file or files you are about to download were provided to ProQuest by the author as part of a
dissertation or thesis. The supplemental files are provided "AS IS" without warranty. ProQuest is not responsible for the
content, format or impact on the supplemental file(s) on our system. in some cases, the file type may be unknown or
may be a .exe file. We recommend caution as you open such files.
Copyright of the original materials contained in the supplemental file is retained by the author and your access to the
supplemental files is subject to the ProQuest Terms and Conditions of use.
Depending on the size of the file(s) you are downloading, the system may take some time to download them. Please be