Just "keeping the lights on" isn't enough at many businesses where high availability and high performance are expected. A virtualization administrator can improve VM responsiveness in numerous ways, such as tuning the virtual memory management configuration so it is optimized for the company's environment.
To help cook up a high-performing VMware vSphere virtual infrastructure, sysadmins can dig into Prasenjit Sarkar's vSphere High Performance Cookbook. Sarkar's book shares multiple step-by-step recipes on how to build and run a robust virtual infrastructure to ensure that VMs perform at the highest level.
Chapter 2, titled "Memory Performance Design," is available to download and read here (PDF). In this chapter, Sarkar focuses on different techniques and best practices for optimal CPU performance, such as trimming the number of VMs running inside a single host, and how to pinpoint why a virtual machine is demanding a disproportionate amount of memory.
There are more topics covered in this chapter regarding memory performance, including virtual memory reclamation techniques, key metrics to monitor and metrics not to use, identifying when memory is actually the problem and more. One of the main topics of the chapter overviews how to keep memory free for VMkernel, highlighted below.
Highlights from Chapter 2:
The amount of memory the VMkernel will try to keep free can be achieved through the MemMinFreePct parameter. MemMinFreePct determines the amount of memory that the VMkernel should keep free. vSphere 4.1 introduced a dynamic threshold of the Soft, Hard and Low state to set appropriate thresholds and prevent virtual machine performance issues, while protecting VMkernel. The different states, based on %pRAM which is still free, determines what type of memory reclamation techniques are being used.
For MemMinFreePct, using a default value of 6 percent can be inefficient when 256 gigabyte or 512 gigabyte systems are becoming more and more mainstream. A 6 percent threshold on a 512 gigabyte results in 30 gigabyte idling most of the time. However, not all customers use large systems; some prefer to scale out rather than to scale up. In this scenario, a 6 percent MemMinFreePct might be suitable. To have the best of both worlds, VMkernel uses a sliding scale to determine the MemMinFreePct threshold based on the amount of RAM installed in vSphere 5 hosts. Sliding scale is not applicable to vSphere 4.
This was first published in February 2014