Tactic: Traffic Aware Cloud for Tiered Infrastructure Consolidation

Large-scale enterprise applications are deployed as distributed applications. These applications consist of many inter-connected components with heterogeneous roles and complex dependencies. Each component typically consumes 5-15% of the server capacity. Deploying each component as a separate virtual machine (VM) allows us to consolidate the low-utilized VMs into the same infrastructure, enhancing the physical resource utilization. The application performance is often specified by the service level agreement (SLA) in terms of the 95th percentile response time.;We propose Tactic, an automated application performance management system to find a VM placement that satisfies the SLA. Conceptually, Tactic finds a placement that provides sufficient resource capacity as well as reduce the resource contention due to requests arriving concurrently. We estimate the probability that the VMs receive the requests simultaneously in order to quantify the amount of resource contention and approximate the achievable response time of each VM. To estimate the overall application response time, we combine the response time of the individual VMs based on their interactions.;We automatically derive the interactions among the VMs based on their dependencies. The execution order of the dependent VMs is also derived to estimate the response time of each VM. We introduce the basic dependency primitives to represent both dependencies and the execution order among the VMs. We use the primitives to construct the dependency model which includes both the relationship of the VMs and the execution order needed for predicting the overall application response time. If the predicted response time does not meet the SLA, we need to find another placement candidate.;We propose a contention-aware placement algorithm that reduces the resource contention for the high-impact VMs. The impact is measured by the ratio of the change in the application response time to the change in the individual VM's response time. The algorithm identifies the VM with the highest impact using the dependency model. Compared to the random placement approach, our contention-aware approach can reduce the number of iterations required to find the placement that satisfies the SLA. With Tactic, we can automatically extract the dependencies from a complex application. We use the request traffic to infer the dependency model for the response time prediction. Tactic satisfies the 95th percentile response time.;We evaluate Tactic in realistic deployment scenarios using Drupal content management system and MongoDB distributed database. The results show that Tactic can accurately predict the 95 th percentile response time with less than 8% error for both applications. Tactic can also identify a placement with up 35% lower overall response time on Drupal, compared to a utilization-based approach.