To drive down cost and increase host utilization in Intel's Office and Enterprise private cloud, Intel IT has simplified private cloud capacity management by transitioning from allocation-based capacity management to performance-based capacity management. Focusing on host and application performance in conjunction with optimizing our private cloud host server configuration has significantly ...increased the number of virtual machines (VMs) and reduced our private cloud host server footprint and cost per VM.
We use three metrics to increase the density of our private cloud hosts (each for private cloud host and VM): total MHz CPU utilization, CPU Ready percentage, and maximum active memory utilization. Using these performance metrics to inform our capacity decisions has resulted in the following:
› An 11:1 average virtual CPU to physical CPU allocation ratio
› An average of 180-percent memory oversubscription
› A 19-percent increase in VM growth while decreasing our private cloud host server footprint by 23 percent during the fall 2014 to fall 2015 time period
› A licensing cost avoidance of between 34 percent and 76 percent
During the time we were using allocation-based performance management, we believed that our private cloud was memory-constrained and needed hosts with more memory. Transitioning to performance-based capacity management and increasing the level of our memory oversubscription led us to discover that CPU utilization—not memory—was the key metric that was preventing us from achieving higher private cloud host density.
Through performance testing, we found that a private cloud host with the 16-core, 2.3 GHz Intel® Xeon® processor E5-2698 v3 and 256 GB memory is the most cost-effective private cloud host configuration for our Office and Enterprise private cloud.