Capacity Planning - Why Bother?

1.0  Introduction

This Xtravirt article discusses Capacity Planning (Workload Profiling) and the reasons why it is an important first step to any virtualization exercise.

Prompted by the frustrations of trying to convince customers of the importance of fully understanding their environment before embarking on a virtualization project, this papers aim is to clear up misconceptions about the purpose and importance of this exercise.

2.0  What is Capacity Planning?

Capacity Planning will be referred to as “Workload Profiling” as this name better represents the core activity of the exercise. Workload profiling is where a software tool such as VMware Capacity Planning, PlateSpin Recon or Lanamark is installed in a customer’s existing environment and configured to systematically gather performance and inventory data over a period of time, typically several weeks. At the end of the exercise the data is analysed and a report is generated showing the usage of the environment. This data is then used as the basis for the virtualization design phase.

The key point to note is that it’s not simply about determining how much new server hardware is required to support the new virtual environment (i.e. Capacity Planning), it’s also about having the detailed information to be able to make informed design decisions during the design phase.

3.0  How it Works

Workload profiling software typically requires one or more collection servers which periodically queries the networked servers or workstations (in the case of VDI[1]) using agentless techniques such as WMI or remote registry calls (for Windows computers). This information is stored in a database ready for analysis at the end of the monitoring period.

The following metrics will typically be collected. This is not an exhaustive list:
-     CPU usage
-     Memory usage
-     Network IO
-     Disk Usage
-     Disk IO
-     Hardware / Software Inventory

1. Workload Profiling for VDI is relatively new, however some providers such as Lanamark are providing specific tools for this

4.0  The Benefits

There is a wealth of information that is captured during the exercise, and some of the key benefits are highlighted below.  These centre around achieving high accuracy for:

• Specification of new server hardware
• Networking design
• Storage design
• Virtual hardware specification

Each of these are now discussed in further detail.

4.1 Specification of new Server Hardware

This is the traditional reason for workload profiling and reflects the commonly applied industry term ‘Capacity Planning’. Whatever software tool is used it should provide a consolidation planning report.  In simple terms this will analyse the total CPU, memory, disk and network usage across all the monitored servers. It then will then examine scenarios which include one or more customer selected server hardware specifications, and Hypervisor platforms, then calculate how many hosts (and what specification) are required to run your existing workloads based on current performance. The software may also make recommendation as to what virtual servers should be placed on which hosts; however, with dynamic load balancing features such as VMware DRS / VMotion / Live Migration, this becomes less important.

It should also be possible to carry out different ‘what-if’ scenarios to experiment with different hardware specifications and number of hosts until a suitable hardware design is found. Our experience from the field is that customers who don’t perform this exercise tend to over specify the hardware requirements of their hosts. Memory in particular appears to be a good example, particularly when allowing for VMware’s transparent page sharing technologies and memory over-commit abilities. A real life example is a customer who bought 8 HP DL380 Hosts with 2 x Quad core Intel processors. Their own guessimate was to provide each server with 64GB RAM. After a capacity planning exercise it was apparent that only ~25GB RAM per Host was required, so 32 GB was recommended. This equated to a saving of 256 GB RAM, more than enough to pay for the workload profiling exercise!

The response from some customers is that they’ll use this as extra capacity for growth, however what they may fail to consider is that they must balance the CPU and memory utilisation.  After all, there is little point having 64 GB RAM if the processors are at maximum capacity when only 25GB is being used!

Note: There are many factors that can affect this but often a Dual Quad core server with 32 GB RAM seems to be a good balance between CPU and Memory usage in a ‘typical’ VMware environment.

4.2 Virtual Networking Design

Performance statistics showing average and peak network IO allows the virtualization architect to make informed design decisions regarding the amount and specification of the networking hardware. This can also assist in designing the virtual switches and allows them to examine their network backbone to ensure that it can handle the expected peak loads.

4.3 Storage Design

There are 2 areas to be concerned with when it comes to storage:

• Storage Throughput (IOPS)
• Storage Capacity

Storage Throughput (IOPS)

The primary consideration when it comes to designing storage should be throughput.  There is no point obtaining excess spare space from cheap slower disks if the performance is unacceptable. Storage throughput is often measured in IO transactions per second (IOPS).  This is the amount of read or write operations performed per second by a particular disk or disk array. Each hard disk is capable of supplying a standard amount of IOPS depending upon spindle speed, eg: 10k, 15k. The more disks that can be accessed in an array the higher the total IOPS. Fewer, albeit faster, hard disks may in fact give you worse performance than using many slower disks. The cumulative peak IOPS data gathered during the workload profiling exercise will assist in finding the best match for storage and LUN design.

Storage Capacity

Knowing the total disk usage allows for accurate purchasing of storage. If the workload profiling exercise is carried out for several weeks then trend analysis can also be carried out to show expected growth. Without this data it is very difficult to accurately estimate how much storage is required now, and in the future.

4.4 Virtual Hardware Specification

There should always be some logic behind selection of new virtual hardware specification especially when converting existing physical servers. There is an overhead to the underlying host when allocating memory or vCPUs, so it’s a good idea to be as accurate as possible. Understanding the ‘actual’ usage of the current workloads is the only real way to effectively size the new virtual machines. Basing this decision on the existing physical memory and processor usage will almost certainly result in an over allocation of resources in most cases, and an inefficient use of the host’s limited resources.

5.0  Summary

Workload profiling is an important exercise that should be considered by businesses of any size that are embarking on a virtualization project. The licensing model of the workload profiling providers scale with size, therefore smaller and larger businesses still achieve comparable value. The cost of the workload profiling exercise is often offset by tangible savings from purchasing the correct hardware and also in the resultant quality of the design and time taken to achieve it.

Whilst knowing what hardware is required is an important element, the information and due diligence gained from these exercises are key project influencers.

6.0  Useful References

Leading Capacity Planning / Workload Profiling tools:

VMware Capacity Planning, http://www.vmware.com/products/capacity-planner/
Platespin Recon, http://www.platespin.com/Products/recon/
Lanamark, http://www.lanamark.com/