Characterization of Cloud-Based Applications for Different Memory Resource Sharing Scenarios

A typical Cloud-Based application is scale-out, latency sensitive, and usually involves processing of enormous amounts of data. Web Searches, Streaming services, online transactions, etc., run on massive Warehouse-Scale Computers that are powered by processors designed to run traditional scale-up personal computer applications.

Since Architectural decisions for the processors that run the Internet today were made with a desktop computer in mind, memory resources are often overprovisioned and underutilized. In this work, five different Cloud-Based applications from the CloudSuite benchmark suite are characterized according to their interaction with memory resources and their level of data sharing. Threads of each workload are distributed among nodes to study how different scenarios of Last Level Cache and Memory Controller Channel sharing can affect performance. Furthermore, the impact of co-running applications is also taken into consideration, and a scheduling algorithm intended to run on top of the Operating System scheduler is proposed.