Intel® Xeon® Scalable Processor - Virtual Desktop Infrastructure with Intel® Speed Select Technology

1. Introduction

The Virtual Desktop Infrastructure (VDI) workload has become more significant with the growth of remote working and distributed workforce.  

The centralization of resources through VDI, provides an opportunity to reduce the total cost of ownership (TCO) for data centers that host VDI sessions. Balancing the uneven distribution of system resources used during VDI sessions can improve the TCO. The issue is that some nodes and cores being allocated to a VDI session need more system resources than are available. Other nodes serving virtual machines (VMs) need fewer system resources and they have a surplus.

To a large extent, Intel® Speed Select Technology solves this problem by allocating system resources to VMs that need them. This is done without compromising the performance of the applications running in the VM.

This guide describes how VDI as a workload can take advantage of the Intel Speed Select Technology, including being optimized for VM capacity without an additional bill of material, to improve the total cost of ownership.

1.1 Terminology

Table 1. Terminology

1.2 Reference Documentation

Table 2. Reference Documentation

2. Overview

The number of people working remotely has increased due to change in the way we work, for example, the hybrid working arrangement, greater flexibility in the workforce, or work from anywhere. The VDI workload enables greater flexibility and freedom so that workers can be productive working from almost anywhere. VDI is a form of desktop virtualization. Virtualization runs inside a virtual machine (VM) that is hosted on a central server. Users are given access to the VM through VDI sessions. For example, a Task worker is someone who uses a single application and needs a small VM.  A Knowledge worker is someone who uses word processors, presentation software, spreadsheets, and other similar tools. They need a larger VM. A Power user is mostly a content creator who may be making videos or other types of content that require more computing resources, memory, and other system resources and would benefit from an even larger VM. The size of the VM changes depending on the application being run inside the VM. Different types of users will have different VM parameter settings based on their user profile. Currently, a typical enterprise IT environment may have a server cluster that serves different types of users. However, since all of the servers in the server cluster have been configured with the same processing capability, memory, and storage configurations, every VM ends up being treated equally. Therefore, some VMs may have a surplus of resources while others are lacking.  

3. Why DVI is Different from Other Virtual Workloads?

The key difference between VDI and other virtualized workloads is the number of VMs that are spawned out of one node. A typical node in VDI deals with hundreds of virtual machines together. System resources are shared among all of these virtual machines and the applications running inside them. VDI sessions also have specific service level agreement (SLA) commitments that must be met.  These may be compromised by a subset of VDI sessions that we call “rogue” sessions. These rogue sessions consume too many resources and starve other VDI sessions until they cannot meet the SLA commitments. This happens when rogue sessions perform computationally intensive tasks that consume more of the CPU energy so that performance suffers in other VDI sessions running on other cores.

System resources such as computing power, memory, storage, and networking must be carefully prioritized and distributed. The allocation of system resources is very important to help ensure that the user’s experience using their virtual desktop is the same as using their laptop, desktop, or other fully-powered, end-user device.

Most of the hypervisor vendors, size the VMs with number of CPU cycles available. This calculation is based on the base frequency of that platform.  A typical Compute Selection during capacity planning derives Calculated Compute Frequency (CCF) of a given platform. CCF is the total CPU cycles of the system, calculated as follows:

CCF = Base Frequency of the CPU x Core Count per Socket x No. of Sockets x Hyper-Threading-Factor

For example, when using the 2nd Generation Intel® Xeon® Scalable Processor 6230, CCF will be calculated as:

CCF = 2100 (base frequency) x 20 (number of cores per socket) x 2 (number of sockets) MHz x 1.4 (HT factor) = 117600

On the same node, if the VMs are sized for different user profiles, then currently, the number of cores is the only sizing factor for getting a different CPU cycle, which results in uneven utilization. VMs hosting applications for some user profiles have too many allocated cycles and others do not have enough. System resources could be better utilized if VMs running Worker applications could share their extra cycles with VMs running Power user applications that need the extra power.

4. Intel® Speed Select Technology

The Intel Speed Select Technology is a collection of power and performance features that gives more granular control over CPU performance. It can be tuned for advanced prioritization of cores and to help obtain more performance out of the same server. The Intel Speed Select Technology features supported by the 3rd Generation Intel® Xeon® Scalable processor aims to allow more control over processor performance in order to support optimized the TCO. Intel® Speed Select Technology - Base Frequency (Intel® SST-BF) and Intel® Speed Select Technology - Core Power (Intel® SST-CP) technology can maintain a higher base frequency on a subset of processor cores and a lower base frequency on the remaining processor cores. Intel® Speed Select Technology - Turbo Frequency (Intel® SST-TF) can maintain a higher base frequency on a subset of processor cores and a lower base frequency on the rest of the processor cores. Intel® SST can be used by VDI users to help boost performance for certain cores either by disabling other cores or by lowering their frequency. Various features within Intel® Speed Select Technology (Intel® SST) can be used to optimize VDI workloads. Two usage models are described in this guide: Intel® Speed Select Technology – Performance Profile (Intel SST-PP) and Intel SST-BF.

4.1 System Level Uniform VDI Instances

Intel SST-PP allows multiple optimized performance profiles per SKU via a static boot-time configuration or a dynamic runtime configuration. The figure below shows a CPU with 8 cores having a lower performance characteristic as a base Intel SST-PP configuration. When switched to a different configuration with only 6 cores available, the per core performance improves resulting in the VMs running on those cores getting a performance boost. In this usage model, the entire CPU and server are reconfigured to use a specific Intel SST-PP profile. Having only one type of VDI VM instance deployed takes full advantage of the optimized CPU performance¹.

#!/bin/sh

i=0

while [ $i -lt 160 ]

do

        vsish -e get /hardware/msr/pcpu/$i/addr/0x774

        i=`expr $i + 1`


The sample output of this command is a list of cores with a frequency

0x8000ff01

0x8000ff01

0x8000ff01

0x8000ff16

0x8000ff16