Understanding Performance Interference of I O Workload in Virtualized Cloud Environments Summarized by Michael Riera 9 17 2011 University of Central Florida CDA5532 Agenda Purpose Xen I O Architecture Experiment Setup Experiment Results Base case Measurement Interference Analysis Throughput Interference Conclusion Purpose The purpose of this paper is to quantify the interference of multiple VMs on the throughput of an I O intensive workload through a series of SPEC benchmark files sizes of 1KB 4KB 10KB 30KB 50KB 70KB 100KB Xen Architecture Description Is a popular open source x86 virtual machine monitor VMM based on virtualization Technologies Supports para virtualization Hypercalls straight to the host operating system Lower overhead PCI Pass through Xen Architecture Description Example NIC Receive Packet Event Channel Xen VMM Decides whether the driver has access to HW Physical Interface Interface to Network card Bridge Removes from NIC Demux the network packet and delivers it to the appropriate VM Backend Raise a hypercall to the VMM requesting space on the guest I O Channel Netback and netfront exchange the page descriptor by page remapping over mechanism over I O descriptor ring Frontend Receives a packet as if it came from a NIC card Experiment Setup IBM ThinkCentre A52 Workstation 3 2Ghz Intel Pentium 4 16KB L1 Cache 2MB L2 Cache 2GB 400Mhz DDR memory Seagate 250GB Intel NIC E100 PRO 100 Network Client machines Connected 1Gb s lines Experiment Setup I O intensive workload Running two isolated guest VM1 VM2 VMM Sharing physical hardware Each VM has equal resources CPU Scheduling SMP Credit scheduler VM1 has an Apache HTTP server VM2 provides the web services Cache buffer data no disk reading are involved Experiment Setup Experiment Results Experiment Results Experiment Results CPU Utilization Measured the average CPU utilization of each VM including CPU usage of Dom0 VM1 and VM2 VMM events per second VMM adopts asynchronous hypercall mechanism to notify the VMs of system events VM switches Metric for number of times VMM need to context switch I O count Pages Exchange I O counts per execution Page exchange per execution VM State Execution state using CPU runnable state waiting for CPU block state waiting for IO Experiment Results Experiment Results Conclusion Competing network intensive loads High Cache misses High Wait times Less competing resources 1 100 Optimal Interference highly sensitive to multiple VMs Credit Scheduler to blame for large wait in the larger files compared to the smaller ones