Modeling of Diverse Computing Requirements in Azure Task Traces
Mulei Ma, Chenyu Gong, Yang Yang, Yue Gao, Kai Li · 2023
In the cloud data centers of large technology companies, there are various resources such as computing, communication and storage, which provide stable and quality-guaranteed services for subscribers. To get a comprehensive understanding of real load trends and to observe the characteristics of user tasks consuming resources, we analyzed the workload of Microsoft Azure Virtual Machines (VMs). Specifically, we analyzed the workload of all VMs in the Azure cluster within one month in 2017 and 2019. The existing public data provides complete information for each schema. Through data analysis and fitting, we revealed the consistency and difference in delay and memory usage of workload tasks during the time span from 2017 to 2019. In addition, we found the distribution characteristics of the task computing requirement, i.e., it conforms to a large scale exponential distribution superimposed with a small scale Sample function. We propose two modeling approaches to fit the 2017 and 2019 traces and find that the complex one is able to reduce the Sum of Squares due to Error (SSE) by 72.9% and 66.9% compared to the simple one. The task characteristics we discovered can help researchers understand the workload and provide a modeling basis for the simulation framework.