Modeling virtualized application performance from hypervisor counters

Abstract

Managing a virtualized datacenter has grown more challenging, as each virtual machine's service level agreement (SLA) must be satisfied, when the service levels are generally inaccessible to the hypervisor. To aid in VM consolidation and service level assurance, we develop a modeling technique that generates accurate models of service level. Using only hypervisor counters as inputs, we train models to predict application response times and predict SLA violations. To collect training data, we conduct a simulation phase which stresses the application across many workloads levels, and collects each response time. Simultaneously, hypervisor performance counters are collected. Afterwards, the data is synchronized and used as training data in ensemble-based genetic programming for symbolic regression. This modeling technique is quite efficient at dealing with high-dimensional datasets, and it also generates interpretable models. After training models for web servers and virtual desktops, we test generalization across different content. In our experiments, we found that our technique could distill small subsets of important hypervisor counters from over 700 counters. This was tested for both Apache web servers and Windows-based virtual desktop infrastructures. For the web servers, we accurately modeled the breakdown points and also the service levels. Our models could predict service levels with 90.5% accuracy on a test set. On a untrained scenario with completely different contending content, our models predict service levels with 70% accuracy, but predict SLA violation with 92.7% accuracy. For the virtual desktops, on test scenarios similar to training scenarios, model accuracy was 97.6%. Our main contribution is demonstrating that a completely data-driven approach to application performance modeling can be successful. In contrast to many other works, our models do not use workload level or response times as inputs to the models, but nevertheless predicts service level accurately. Our approach also lets the models determine which inputs are important to a particular model's performance, rather than hand choosing a few inputs to train on.