Dealing with Heterogeneity and Poor Isolation in Public
Dealing with Heterogeneity and Poor Isolation in Public Clouds Venkatanathan Varadarajan, Thawan Kooburat, Benjamin Farley, Kevin Bowers§, Ari Juels§, Thomas Ristenpart and Michael M. Swift AMD Problem: Simple Pricing Model Intel – E 5430 Same Hardware Type Pay per hour of usage Intel – E 5507 Intra-CPU Heterogeneity 1. Multi-tenancy and poor isolation: 3 x-6 x performance degradation 2. Variations in motherboard & peripherals For the same price! CPU Heterogeneity 1. Public clouds deploy varying hardware over time 2. New machines faster & healthy For the same price! Fig-1: Amazon EC 2 machine distribution across zones Fig-2: EC 2 local disk across architectures Sources of Performance Variability 1. Heterogeneous hardware: New machines perform better, 2. Multi-tenancy: Sharing same machine with multiple tenant VMs. Solution 1: Placement Gaming Fig-3: EC 2 local disk on three different E 5507 s Solution 2: Resource-Freeing Attacks Goal: Get onto best-performing machines for workload Constraints: No access to cloud scheduler, no migration Cloud Provider API: Start and Stop VM operations Goal: Reduce contention for (poorly isolated) target resource Constraints: No access to hypervisor resource scheduler Public Interface: Exposed by the webserver Approach: Customer-Controlled Placement 1. Use additional VMs: Explore for better machines 2. On-the-fly migration: Optimistically replace slow VMs RFA Idea Introduce new work to Shift victim resource usage away from the target resource towards the bottleneck resource Placement Model A - Service instances An RFA Example Setting 1. Two webservers fighting over network bandwidth 2. Each gets half bandwidth compared to running alone B - Exploratory instances Steps in Placement Gaming Costs 1. Start A+B instances 1. Migration costs 2. Stop B slowest instances after unit time 2. B exploratory instances 3. Replace slow instances with new ones Results: A=10, B=0 for 12 hours on EC 2 Steps in an RFA 1. Send compute-intensive CGI requests 2. Create a CPU bottleneck 3. Free up target network resource Results: Under Xen VMM References Ø "More for Your Money: Exploiting Performance Heterogeneity in Public Clouds", Farley, B. , V. Varadarajan, K. Bowers, A. Juels, T. Ristenpart, and M. Swift in Symposium on Cloud Computing (SOCC). Ø "Resource-freeing attacks: improve your cloud performance (at your neighbor's expense)", V. Varadarajan, T. Kooburat, B. Farley, T. Ristenpart, and M. Swift in Proceedings of the 2012 ACM conference on Computer and Communications Security (CCS). § Part of this work was a collaboration with RSA labs
- Slides: 1