Browsing by Subject "Distributed Systems"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Multi-Tenant Geo-Distributed Data Analytics
    (2019-07) Jonathan, Albert
    Geo-distributed data analytics has gained much interest in recent years due to the need for extracting insights from geo-distributed data. Traditionally, data analytics has been done within a cluster/data center environment. However, analyzing geo-distributed data using existing cluster-based systems typically cannot satisfy the timeliness requirement of most applications and result in wasteful resource consumption due to the fundamental differences of the environments, especially due to the scarce, highly heterogeneous, and dynamic nature of the wide-area resources: compute power and network bandwidth. This thesis addresses the challenges faced by geo-distributed data analytics systems in ensuring high-performance and reliable execution of multiple data analytics applications/queries. Specifically, the focus is on sharing resources across multiple users, applications, and computing frameworks. Sharing resources is attractive as it increases resource utilization and reduces operational cost. However, ensuring high-performance execution of multiple applications in a shared environment is challenging as they may compete for the same resources, especially in a wide-area environment with scarce resources. Furthermore, dynamics such as workload variation, resource variation, stragglers, and failures are inevitable in large-scale distributed systems. These can cause large resource perturbation that significantly affect the performance of query executions. This thesis makes the following contributions. First, we present a resource sharing technique across multiple geo-distributed data analytics frameworks. The main challenge here is how to elastically partition resources while allowing high locality scheduling to each individual framework, which is critical to the execution performance of geo-distributed analytics queries. We then address the problem of how to identify and exploit common executions across multiple queries to mitigate wasteful resource consumption. We demonstrate that traditional multi-query optimization may degrade the overall query execution performance due to its lack of support for network awareness. Finally, we highlight the importance of adaptability in ensuring reliable query execution in the presence of dynamics, both for single and multiple query executions. We propose a systematic approach that can selectively determine which queries to adapt and how to adapt them based on the types of queries, dynamics, and optimization goals.
  • Thumbnail Image
    Item
    Optimizing Timeliness, Accuracy, and Cost in Geo-Distributed Data-Intensive Computing Systems
    (2016-12) Heintz, Benjamin
    Big Data touches every aspect of our lives, from the way we spend our free time to the way we make scientific discoveries. Netflix streamed more than 42 billion hours of video in 2015, and in the process recorded massive volumes of data to inform video recommendations and plan investments in new content. The CERN Large Hadron Collider produces enough data to fill more than one billion DVDs every week, and this data has led to the discovery of the Higgs boson particle. Such large scale computing is challenging because no one machine is capable of ingesting, storing, or processing all of the data. Instead, applications require distributed systems comprising many machines working in concert. Adding to the challenge, many data streams originate from geographically distributed sources. Scientific sensors such as LIGO span multiple sites and generate data too massive to process at any one location. The machines that analyze these data are also geo-distributed; for example Netflix and Facebook users span the globe, and so do the machines used to analyze their behavior. Many applications need to process geo-distributed data on geo-distributed systems with low latency. A key challenge in achieving this requirement is determining where to carry out the computation. For applications that process unbounded data streams, two performance metrics are critical: WAN traffic and staleness (i.e., delay in receiving results). To optimize these metrics, a system must determine when to communicate results from distributed resources to a central data warehouse. As an additional challenge, constrained WAN bandwidth often renders exact computation infeasible. Fortunately, many applications can tolerate inaccuracy, albeit with diverse preferences. To support diverse applications, systems must determine what partial results to communicate in order to achieve the desired staleness-error tradeoff. This thesis presents answers to these three questions--where to compute, when to communicate, and what partial results to communicate--in two contexts: batch computing, where the complete input data set is available prior to computation; and stream computing, where input data are continuously generated. We also explore the challenges facing emerging programming models and execution engines that unify stream and batch computing.
  • Thumbnail Image
    Item
    Transaction and data consistency models for cloud applications
    (2014-02) Padhye, Vinit A.
    The emergence of cloud computing and large-scale Internet services has given rise to new classes of data management systems, commonly referred to as NoSQL systems. The NoSQL systems provide high scalability and availability, however they provide only limited form of transaction support and weak consistency models. There are many applications that require more useful transaction and data consistency models than those currently provided by the NoSQL systems. In this thesis, we address the problem of providing scalable transaction support and appropriate consistency models for cluster based as well as geo-replicated NoSQL systems. The models we develop in this thesis are founded upon the snapshot isolation (SI) model which has been recognized as attractive for scalability. In supporting transactions on cluster-based NoSQL systems, we introduce a notion of decoupled transaction management in which transaction management functions are decoupled from storage system and integrated with the application layer. We present two system architectures based on this concept. In the first system architecture all transaction management functions are executed in a fully decentralized manner by the application processes. The second architecture is based on a hybrid approach in which the conflict detection functions are performed by a dedicated service. Because the SI model can lead to non-serializable transaction executions, we investigate two approaches for ensuring serializability. We perform a comparative evaluation of the two architectures and approaches for guaranteeing serializability and demonstrate their scalability. For transaction management in geo-replicated systems, we propose an SI based transaction model, referred to as causal snapshot isolation (CSI), which provides causal consistency using asynchronous replication. The causal consistency model provides more useful consistency guarantees than the eventual consistency model. We build upon the CSI model to provide an efficient transaction model for partially replicated databases, addressing the unique challenges raised due to partial replication in supporting snapshot isolation and causal consistency. Through experimental evaluations, we demonstrate the scalability and performance of our mechanisms.