Big Data Application and System Co-optimization in Cloud and HPC Environment

The emergence of big data requires powerful computational resources and memory subsystems that can be scaled efficiently to accommodate its demands. Cloud is a new well-established computing paradigm that can offer customized computing and memory resources to meet the scalable demands of big data applications. In addition, the flexible pay-as-you-go pricing model offers opportunities for using large scale of resources with low cost and no infrastructure maintenance burdens. High performance computing (HPC) on the other hand also has powerful infrastructure that has potential to support big data applications. In this dissertation, we explore the application and system co-optimization opportunities to support big data in both cloud and HPC environments. Specifically, we explore the unique features of both application and system to seek overlooked optimization opportunities or tackle challenges that are difficult to be addressed by only looking at the application or system individually. Based on the characteristics of the workloads and their underlying systems to derive the optimized deployment and runtime schemes, we divide the workflow into four categories: 1) memory intensive applications; 2) compute intensive applications; 3) both memory and compute intensive applications; 4) I/O intensive applications.When deploying memory intensive big data applications to the public clouds, one important yet challenging problem is selecting a specific instance type whose memory capacity is large enough to prevent out-of-memory errors while the cost is minimized without violating performance requirements. In this dissertation, we propose two techniques for efficient deployment of big data applications with dynamic and intensive memory footprint in the cloud. The first approach builds a performance-cost model that can accurately predict how, and by how much, virtual memory size would slow down the application and consequently, impact the overall monetary cost. The second approach employs a lightweight memory usage prediction methodology based on dynamic meta-models adjusted by the application's own traits. The key idea is to eliminate the periodical checkpointing and migrate the application only when the predicted memory usage exceeds the physical allocation. When applying compute intensive applications to the clouds, it is critical to make the applications scalable so that it can benefit from the massive cloud resources. In this dissertation, we first use the Kirchhoff law, which is one of the most widely used physical laws in many engineering principles, as an example workload for our study. The key challenge of applying the Kirchhoff law to real-world applications at scale lies in the high, if not prohibitive, computational cost to solve a large number of nonlinear equations. In this dissertation, we propose a high-performance deep-learning-based approach for Kirchhoff analysis, namely HDK. HDK employs two techniques to improve the performance: (i) early pruning of unqualified input candidates which simplify the equation and select a meaningful input data range; (ii) parallelization of forward labelling which execute steps of the problem in parallel. When it comes to both memory and compute intensive applications in clouds, we use blockchain system as a benchmark. Existing blockchain frameworks exhibit a technical barrier for many users to modify or test out new research ideas in blockchains. To make it worse, many advantages of blockchain systems can be demonstrated only at large scales, which are not always available to researchers. In this dissertation, we develop an accurate and efficient emulating system to replay the execution of large-scale blockchain systems on tens of thousands of nodes in the cloud. For I/O intensive applications, we observe one important yet often neglected side effect of lossy scientific data compression. Lossy compression techniques have demonstrated promising results in significantly reducing the scientific data size while guaranteeing the compression error bounds, but the compressed data size is often highly skewed and thus impact the performance of parallel I/O. Therefore, we believe it is critical to pay more attention to the unbalanced parallel I/O caused by lossy scientific data compression