PROJECT TITLE :

Multi-tier Workload Consolidations in the Cloud Profiling, Modeling and Optimization

ABSTRACT:

It is becoming increasingly important to cut down on tail latency in order to improve the experience that users have when using a service. Cloud applications that are user-facing and sensitive to latency typically consist of multiple interactive tiers (such as Web, App, and Database) running in separate virtual machines (VMs) with complex interaction patterns. However, previous methods for VM consolidation frequently ignored the interactions that can occur between virtual machines (VMs) located in different tiers, which led to subpar performance for applications. We take a fresh look at the issue of user-perceived tail latency as a lens through which to examine the consolidation of multi-tier interactive workloads in this article. We propose an innovative consolidation method that is based on profiling in order to satisfy tail latency requirements while simultaneously reducing the number of physical machines that are used. In order to accomplish such a goal, we must first carry out large-scale profiling experiments in a KVM virtualized private cluster using a variety of consolidation settings so that we can determine the empirical performance values. Interference with co-located virtual machines (VMs) and interaction between tiers are two key factors that we take into consideration when analyzing the effect of multi-tier workloads on tail latency. We model the consolidation of workloads across multiple tiers as an optimization problem, complete with a variety of objectives and constraints, and we derive the consolidation schedule from this model. We put the proposed models into action and assess them, in addition to making comparisons with other approaches (i.e., without profiling or without considering interaction influence). Extensive experimental results show that the proposed method is capable of reducing tail latency by up to 5X compared to the method without profiling and by up to 1.3X compared to the method without considering the interaction influence between different tiers. Both of these comparisons are based on the fact that the proposed method takes into consideration the interaction influence between different tiers.