| Title | : | Detecting Performance Interference in Containerized Cloud Environments |
| Speaker | : | Kartik Joshi (IITM) |
| Details | : | Wed, 25 Jul, 2018 3:00 PM @ A M Turing Hall |
| Abstract: | : | Services are often deployed in cloud environments in the form of containers due to low initialization times and overheads. Cloud instances are usually virtual machines (VMs) which utilize the same underlying hardware resources of the physical machine. Hypervisors such as Xen, KVM, etc., partition the hardware resources among VMs and isolate the applications of different users. Some of the resources such as last-level cache, memory bus, etc., cannot be partitioned and contention for these resources can create unpredictability in the performance of the applications in the VMs. This performance interference can be very severe for latency-sensitive and user-facing applications. Existing works to detect performance interference are either too expensive in terms of offline profiling or feasible only from the perspective of the infrastructure owner. In this work, we present two subscriber-centric mechanisms to detect performance interference in containerized cloud environments. In the first mechanism called Sherlock, we profile the service for a short duration to determine its ideal behaviour using linear regression. Any significant deviation from the ideal behaviour caused due to cache interference is detected and the subscriber is notified. In the second mechanism called Watson, we monitor the replicated instances of the service in parallel and predict the ideal performance of the service at runtime. It identifies the services whose performance vary significantly from the ideal and notifies the cloud subscriber. As part of this work, we also set up a BOSS MOOL OS based OpenStack private cloud. We discuss some of the major challenges faced along with their solutions. We also discuss how this infrastructure opens up new directions for research and academic purposes. Experiments on the private cloud with a real-world web service and data service show that our proposed techniques achieve high detection accuracy. |
