Parallel Algorithms and Systems for Applications in Data Analytics

并行计算研究的主要目标是创建使能技术，用于解决自然科学，工程，医学，商业和社会科学中的数据密集型和/或计算困难的问题。我建议的研究课题是：（1）并行算法和系统的实时数据聚合高速数据现代数据分析系统严重依赖于数据聚合，通常实现为二进制关联聚合查询（例如，总和或最大值）在一个指定的子集的数据项存储在数据库中。与通常仅访问数据库的一小部分（例如，更新客户记录）的传统交易处理系统的查询相比，用于数据分析的聚合查询可能需要聚合数据库的大部分（例如，计算作为时间的函数的某种类型的物品的总销售额）。这可能会导致大型数据集的严重性能问题。此外，连续监控高速数据流（例如股票交易数据流）中的新事件的应用程序需要能够在流数据到达时实时分析它。我们建议研究使用混合并行架构（由具有CPU和GPU的计算节点组成的集群/云）对高速数据进行实时数据聚合。(2)自动调整YARN许多数据分析应用程序都构建在Hadoop map-reduce和Spark等关键技术之上，这两种技术都使用YARN作为资源管理器。在给定的硬件平台上安装这样的系统涉及调整许多参数。手动调优通常会导致次优和脆弱的性能，因为对于一个作业（输入数据集）最优的参数可能不太适合另一个作业。自动调整并行系统是便携式系统，可自动适应不同和/或变化的硬件配置和输入数据集。我们建议研究如何为云架构自动调优YARN。(3)蛋白质分析的并行算法和系统蛋白质-蛋白质相互作用（PPI）是定义细胞生物学的基本分子相互作用。PPI被认为涉及药物发现的重要靶点，并与许多细胞状况和疾病有关。用一组给定的PPI（药物靶点）设计合成蛋白质被称为蛋白质工程。我们建议开发一种新的并行蛋白质工程系统，用于大规模混合并行架构（由具有CPU和GPU的计算节点组成的集群/云）。

The main goal of parallel computing research is to create enabling technology for solving data intensive and/or computationally hard problems in the Natural Sciences, Engineering, Medical Sciences, Business, and Social Sciences. My proposed research topics are: (1) Parallel Algorithms And Systems For Real-Time Data Aggregation On High Velocity DataModern data analytics systems rely heavily on data aggregation typically implemented as binary associative aggregation queries (for example, sum or max) over a specified subset of the data items stored in the database. In contrast to queries for traditional transaction processing systems which typically access only a small portion of the database (e.g. update a customer record), aggregation queries for data analytics may need to aggregate large portions of the database (e.g. calculate the total sales of a certain type of items as a function of time). This can lead to significant performance issues for large data sets. In addition, applications that continuously monitor new events in high velocity data streams (e.g. stock exchange data streams) require the ability to analyze streaming data as it arrives, in real-time. We propose to study the use of hybrid parallel architectures (clusters/clouds comprised of compute nodes with CPUs and GPUs) for real-time data aggregation on high velocity data.(2) Auto-Tuning YARNMany data analytics applications are built on top of key technologies such as Hadoop map-reduce and Spark, both of which use YARN as resource manager. The installation of such systems on a given hardware platform involves tuning many parameters. Manual tuning often results in sub-optimal and brittle performance because parameters that are optimal for one job (input data set) may not be well suited to another. Auto-tuned parallel systems are portable systems that adapt automatically to different and/or changing hardware configurations and input data sets. We propose to study how to auto-tune YARN for cloud architectures. (3) Parallel Algorithms And Systems For Protein AnalyticsProtein-protein interactions (PPIs) are essential molecular interactions that define the biology of a cell. PPIs are thought to involve important targets for drug discovery and are linked to a number of cellular conditions and diseases. Designing synthetic proteins with a given set of PPIs (drug targets) is called protein engineering. We propose to develop a new parallel protein engineering system for large scale hybrid parallel architectures (clusters/clouds comprised of compute nodes with CPUs and GPUs).