Modeling Memory Bandwidth Interference in Cloud Data Centers via Deep Learning

While large scale data centers have become more and more popular, performance interference among different workloads brings challenges to the guarantee of QoS. Although quite a number of studies have been made on handling such interference caused by resource sharing, there still lacks work on modeling the memory bandwidth interference which could obviously affect the workload performance. Due to the lack of isolation techniques on memory bandwidth and the dynamics of memory access patterns, workload interference on memory bandwidth is more complicated than the interference on other shared resources. In this paper, we propose a novel Memory Bandwidth Interference Model (MBIM) to estimate the interference of memory bandwidth in co-located cloud workloads. To capture the interference precisely, our MBIM harnesses the architectural structure of the workload deployment and utilizes micro-architectural metrics. Our MBIM consists of four parts: (1) Container static feature selection. (2) Dynamic load information integration. (3) Hierarchical information generation using multi-layer perceptron. (4) Memory bandwidth interference estimation model based on deep learning. Extensive experiments demonstrate that our MBIM can improve average accuracy by 14.9% in MSE, 10.7% in MAPE, and 6.7% in SMAPE of the estimated memory bandwidth interference. To the best of our knowledge, this is the first work of modeling memory bandwidth interference in cloud data centers.