A Quantitative Model for Predicting Cross-application Interference in Virtual Environments

TL;DR

This paper introduces a quantitative model that predicts cross-application interference in virtual environments by considering shared resource access, significantly improving prediction accuracy over previous qualitative approaches.

Contribution

It presents a novel model that accounts for multiple shared resources and access patterns to accurately predict interference in cloud-based HPC applications.

Findings

Prediction errors around 4-12%

Achieved less than 10% error in 96% of cases

Model outperforms qualitative approaches

Abstract

Cross-application interference can affect drastically performance of HPC applications when running in clouds. This problem is caused by concurrent access performed by co-located applications to shared and non-sliceable resources such as cache and memory. In order to address this issue, some works adopted a qualitative approach that does not take into account the amount of access to shared resources. In addition, a few works, even considering the amount of access, evaluated just the SLLC access contention as the root of this problem. However, our experiments revealed that interference is intrinsically related to the amount of simultaneous access to shared resources, besides showing that another shared resources, apart from SLLC, can also influence the interference suffered by co-located applications. In this paper, we present a quantitative model for predicting cross-application interference in virtual environments. Our proposed model takes into account the amount of simultaneous access to SLLC, DRAM and virtual network, and the similarity of application's access burden to predict the level of interference suffered by applications when co-located in a same physical machine. Experiments considering a real petroleum reservoir simulator and applications from HPCC benchmark showed that our model reached an average and maximum prediction errors around 4\% and 12\%, besides achieving an error less than 10\% in approximately 96\% of all tested cases.