Abstract

In big data world, Hadoop and other batch-processing tools are widely used to analyze data and get results in minutes. However, minutes of latency still cannot satisfy the proliferated needs for real-time decision in many fields such as live stock and trading feeds in financial services, telecommunications, sensor networks, online advertisement, etc. Distributed stream processing (DSP) systems aim to process, analyze and make decisions on-the-fly based on immense quantities of data streams being dynamically generated at high rates. As the rates of data streams may vary over time, DSP systems require an architecture that is elastic to handle dynamic load. Although many dynamic load balancing and autoscaling techniques for general pull-based distributed systems have been well studied, these solutions cannot be directly applied to DSP systems because DSP systems are push-based, they process data streams with different types of operators, each running on a cluster node. One research problem is to allocate data processing operators on nodes of clusters and balance the workload dynamically. Since the data volume and rate can be unpredictable, static mapping between operators and cluster resources often results in unbalanced operator load distribution. Furthermore, the problem of making DSP system scalable requires autoscaling at runtime. In this context, the operators need to be relocated among newly provisioned nodes. The contribution of this thesis is three folds. First, we proposes a software layer that is load-adaptive between a DSP engine and clusters. The architecture allows dynamic transferring of an operator to different cluster nodes at runtime and keeps the process transparent to developers. Second, an optimization method that combines correlation of resource utilization of nodes and capacity of clusters is proposed to balance load dynamically. Lastly, we design the autoscaling mechanism and algorithm to detect overload and provision nodes at runtime. We implement our design on S4, an open-source DSP engine first developed by Yahoo!. The implementation is evaluated by a top-N topic list application on Twitter streams using clusters on Amazon Web Services. The results demonstrate a 75.79% improvement on stream processing throughputs, and a 294.47% improvement on cluster resource utilization.