Enhancing performance of networked systems in the cloud by considering the characteristics of resources and workloads

Abstract

Cloud computing has experienced tremendous success in today’s market. In the cloud, various networked systems such as intrusion detection systems, firewalls, and containers are running and consist of diverse cloud-based services. The performance of the networked systems matters because their performance directly affects the operating cost and revenue of the cloud-based service providers. However, delivering the high performance of the systems is nontrivial because they would operate slowly if the characteristic of the resource and workload is not considered in their design. In this dissertation, we argue that one can effectively enhance the performance of the networked systems in the cloud by considering the dependency relationship in workload and resource usage characteristics. In particular, we argue that the performance of the systems can be improved efficiently by designing in consideration of the dependency relationships hidden in program code or data flow. Our methodology falls into three categories. First, we dramatically improved the throughput of networked middleboxes through a design that considers the instruction-level pipelining of CPU and dependency relationship between characters in string patterns. Second, we efficiently reduced the latency of mobile applications by leveraging the low latency of the edge cloud and the dependency relationship between protocol messages. Finally, we reduced the tail latency of microservices by considering the shared nature of resource and microservice chain.