Effect of system load on video service metrics

Model selection, in order to learn the mapping between the kernel metrics of a machine in a server cluster and a service quality metric on a client’s machine, has been addressed by directly applying Linear Regression (LR) to the observations. The popularity of the LR approach is due to: 1) its implementation efficiency; 2) its low computational complexity; and finally, 3) it generally captures the data relatively accurately. LR, can however, produce misleading results if the LR model does not characterize the system: this deception is due in part to its accuracy. In the client-server service modeling literature LR is applied to the server and client metrics without treating the load on the system as the cause for the excitation of the system. By contrast, in this paper, we propose a generative model for the server and client metrics and a hierarchical model to explain the mapping between them, which is cognizant of the effects of the load on the system. Evaluations using real traces support the following conclusions: The system load accounts for ≥ 50% of the energy of a high proportion of the client and server metric traces –modeling the load is crucial; the load signal is localized in the frequency domain: we can remove the load by deconvolution; There is a significant phase shift between both the kernel and the service-level metrics, which, coupled with the load, heavily biases the results obtained from out-of-the-box LR without any system identification pre-processing.