60 likes | 213 Views
Energy Efficient Computing. Prof. Tajana Šimunić Rosing UCSD. Power/Thermal Management in Virtualized Environments. Apps. Apps. Apps. OS. OS. OS. Guest 1. Guest 2. Guest n. Objectives: CPU Scheduling I/O Access/Management Power/Thermal Management?. Hypervisor. Hardware. I/O.
E N D
Energy Efficient Computing Prof. Tajana Šimunić Rosing UCSD
Power/Thermal Management in Virtualized Environments Apps Apps Apps OS OS OS Guest 1 Guest 2 Guest n • Objectives: • CPU Scheduling • I/O Access/Management • Power/Thermal Management? Hypervisor Hardware I/O CPU n/w HDD CPU0 CPU1 CPU2 CPUn
Approach • Control knobs • DVFS • Switch CPUs to low frequencies when idle/underutilized • DPM • Control CPU and I/O device low power states • Guest Scheduling • Co-locate guests to free up resources • Eg. free up sockets/memories in cc-NUMA machines • Guest Migration • Migrate guests to free resources/machines • Challenge • Minimize impact on performance!
Approach • Identify guest performance characteristics Guest n • I/O Intensiveness • Maintain metrics for I/O accesses per guest Hypervisor • CPU Intensiveness • Use CPU performance counters to estimate avg IPC I/O Intensive? CPU Intensive? Hardware I/O CPU n/w HDD CPU0 CPU1 CPU2 CPUn
Characteristics driven policies • Scheduling • Co-locateguests with orthogonal characteristics • Eg I/O intensive and CPU intensive guests can share CPU socket without impacting each others performance • DVFS/DPM policies • Based on the utilization metrics maintained per guest • Migration • Migration policies to balance guests across physical servers • Thermal Management: Balance I/O and CPU intensive guests
Implementation • Use Xen 3.3 as the hypervisor • Includes CPUfreq driver for DVFS • Has a basic ondemand governor • Implement a guest characteristics based governor • Includes idle governor for DPM • Enhance it to include guest characteristics • Enhance the base scheduler to perform clustering/balancing based on characteristics • Add a distributed module in the scheduler to communicate with other hypervisors and perform load balancing across physical servers