High-performance load forecasting on large-scale VM-type thin client system using data stream processing technology

This paper describes a load forecasting method for a large-scale VM-type thin client system. The purpose of this method is to deploy the client environment VMs for load balancing and energy-saving. For load forecasting, it is necessary to collect and analyze load data generated continuously and in large quantities, the so called "data stream," from all VMs. However, existing concentrated batch processing technology with a database is not suitable for processing such a data stream in real time due to the database access overhead. To address this issue, we propose an incremental forecasting algorithm and a distributed processing architecture. The former serially processes data on the route of data flow without storing it in the database. The latter distributes the processing load by hierarchizing the system. We also develop a prototype using our data stream processing framework and evaluate it. Our simulated experiments confirm that our proposed method achieves forecast processing on a large-scale VM-type thin client system in real time.