A hybrid neuro--wavelet predictor for QoS control and stability
This work addresses resource over-provisioning and service stability for online services, though it appears incremental as it combines existing wavelet analysis and recurrent neural network methods.
The paper tackled the problem of predicting user request peaks in distributed systems to enable proactive resource allocation, achieving a root mean squared error below 0.06% for short-term forecasts.
For distributed systems to properly react to peaks of requests, their adaptation activities would benefit from the estimation of the amount of requests. This paper proposes a solution to produce a short-term forecast based on data characterising user behaviour of online services. We use \emph{wavelet analysis}, providing compression and denoising on the observed time series of the amount of past user requests; and a \emph{recurrent neural network} trained with observed data and designed so as to provide well-timed estimations of future requests. The said ensemble has the ability to predict the amount of future user requests with a root mean squared error below 0.06\%. Thanks to prediction, advance resource provision can be performed for the duration of a request peak and for just the right amount of resources, hence avoiding over-provisioning and associated costs. Moreover, reliable provision lets users enjoy a level of availability of services unaffected by load variations.