WP#84 - TRADE-OFFS OF PROCESSOR POWER MANAGEMENT FUNCTIONS OF SERVERS

28 April, 2021 | White Paper

CPU Power Management functions (P-states and C-states or their equivalent) offer hardware available, software initiated functions to reduce CPU voltage and frequency when workload demands are low or absent to reduce server energy use.  Depending on the amount of time a server operates below 25% utilization or is idle, with longer periods generating higher energy savings, power demand can be reduced up to 50%  with the full implementation of P-states and C-states.  These power demand reductions improve SERT® overall efficiency scores by up to 30%.

 

Power management functions come at a cost of reduced server performance and increased response time (exit latency).  Data from SERT measurements on a single configuration with power management turned on resulted in performance reductions at the 100% utilization level in the Hybrid ssj worklet of 5.6% for an EPYC™ CPU and 14.3% for a Power9™ CPU compared to power management off.  For response time, the literature shows that P-state transition times and C-state exit latencies cause response delays that are problematic for some workloads and applications.  As an example, operating system (OS) managed power management profiles can interfere with virtualization programs like VMware® impeding performance rates on the virtualized images.

 

Power management can be a benefit or a problem depending on the data center operations, workloads and applications.  While the implementation of a power management profile can be beneficial in many instances, the optimum P-state and C-state settings and the choice of controlling software – BIOS, hypervisor or operating system – will depend on the specific use case.  In some cases, such as high speed financial trading, network providers or high performance computing, power management functions will be need to be turned off to ensure the performance and response times demanded by those workloads.


Topics: CPU; Power Management;