The test environment for this study consisted of a native (physical) Exchange server environment and a virtual Exchange server environment. Each environment supported 16,000 active mailboxes that were running a LoadGen heavy user profile. Each environment contained a mix of MAPI, Outlook Web App, and Exchange ActiveSync Client Access types. (For a full description of the workload, see the "Test Methodology" section.)

Note:
In this white paper, native refers to a Windows installation without Hyper-V enabled; root refers to the parent partition inside a Windows configuration with Hyper-V enabled; and guest virtual machine refers to the child partition that is hosted on the root (parent) partition of Windows.

Visual representation of hardware that was used in the test environment

The native Exchange server environment consisted of two Client Access servers, two Hub Transport servers, four Mailbox servers, and two domain controllers all running a 64-bit version of Windows Server 2008. All physical Exchange servers were deployed on Hewlett-Packard (HP) DL380 G5 servers that had appropriate memory and the disk configurations shown in Table 1. The Mailbox servers were connected to HP Direct Attached Storage (DAS) that had both 300-GB SAS (MSA60s) and 146-GB SAS (MSA70s) disk types represented. (See the "Storage Configuration" section for additional details.)

Table 1 Summary of hardware that was used in the native environment

Return to top

The virtual server environment consisted of two physical root servers that were running a 64-bit version of Windows Server 2008 and the release to manufacturing (RTM) version of Windows Server 2008 Hyper-V. (There were no Hyper-V hotfixes at the time of this test.) For more information about the RTM version of Windows Server 2008 Hyper-V, see Microsoft Knowledge Base article 950050, Description of the update for the release version of the Hyper-V technology for Windows Server 2008. 

All physical root servers were deployed on HP DL580 G5 servers with appropriate memory and the disk configurations shown in Table 2. Microsoft System Center Virtual Machine Manager was used to convert each native Exchange server into a guest virtual machine. One virtual Client Access server, one virtual Hub Transport server, and two virtual Mailbox servers were deployed on each physical Hyper-V root server. The original DAS was then transferred to the root servers and configured as SCSI pass-through storage. All storage and Exchange server configurations were identical in the native and virtual environments. The two domain controllers were not in the scope of this test, and they remained physical servers.

Table 2 Summary of hardware that was used in the virtual environment

Return to top

Two common HP DAS solutions were tested as part of this study. The first solution used two HP MSA70 storage enclosures with 48 146-GB (2.5") hard disk drives per Mailbox server. The second solution used two HP MSA60 storage enclosures with 24 300-GB SAS (3.5") hard disk drives per Mailbox server. Both solutions were tested with 4,000 mailboxes per server and an initial mailbox size of 200 MB. For more information about the storage solutions that were used in this study, see Table 3.

Table 3 Summary of storage solutions

The server configuration for load testing consisted of sixteen servers (load servers) running a 64-bit version of Windows Server 2008 and Microsoft Exchange Load Generator (v8.02.0045). All load servers were deployed on HP DL380 G4 servers with the memory and the disk configurations shown in Table 4.

Table 4 Summary of hardware used for load generation

Return to top

The latest BIOS versions for the HP ProLiant DL380 G5 and DL580 G5 servers that were used in this study support advanced power regulation settings as defined below. All servers were set to OS Control Mode. (The default configuration for these servers is HP Dynamic Power Savings Mode.)

BIOS advanced power regulation settings included:

• HP Dynamic Power Savings Mode: Automatically varies processor speed and power usage based on processor utilization. Allows for reducing overall power consumption with little or no effect on performance. Does not require operating system support.
  
• HP Static Low Power Mode: Reduces processor speed and power usage. Guarantees a reduced maximum power usage for the system. Performance effects will be bigger for environments that have more processor utilization.
  
• HP Static High Performance Mode: Processors will always run in their maximum power/performance state regardless of the operating system power management policy.
  
• OS Control Mode: Processors always run in their maximum power/ performance state unless the operating system is using a power management policy.
  

In addition to the BIOS settings, all native Exchange servers and Hyper-V root servers had the following Windows Server 2008 power configuration settings applied. These settings change P-state parameters so that they drop to a lower frequency state more aggressively during light loads. We expected this to provide better power utilization during the evening period of our 24-hour test.

powercfg /setpossiblevalue /sub_processor /procperf 2 binary 640364000000a0860100a08601001e00000032000000

powercfg /setactive scheme_balanced

For more information about P-State parameters, see the Windows Server Performance Team Blog article, Configuring Windows Server 2008 Power Parameters for Increased Power Efficiency.

Return to top

Power consumption for all Exchange servers and DAS servers was monitored by using Smart-Watt power monitoring devices. These devices were in-line between the server power supply and the rack power distribution unit and collected average watt consumption over five-minute intervals. The monitoring devices read the power consumption millions of times per minute and produced an average for the five-minute period. Each power read included a time stamp, the number of watts being consumed, and the total watt hours consumed to date. The data from all devices was collected and stored in a Microsoft SQL Server database for easy analysis. For more information about Smart-Watt devices, see the Smart Works Web site.

This testing was conducted at the Microsoft Enterprise Engineering Center (EEC), a state-of-the-art data center on the Microsoft main campus in Redmond, Washington. With over $100M in hardware and networking equipment from top manufacturers, the EEC can replicate almost any production environment. Many of our most important customers visit the EEC to see how prereleases of Microsoft products will perform in a production environment before the customers deploy these products in their own environments. By visiting the EEC, customers can influence the final release, reduce uncertainty, and reduce their deployment costs.

Customer validation testing at the EEC also lets Microsoft product groups listen directly to customers, learn about real-world applications of our technologies, and work with customers to find defects in the software before it’s released to the market.

Each Mailbox server was configured with 4,000 mailboxes for a total of 16,000 mailboxes in the environment. The mailboxes were initialized with an average size of 200 MB using Load Generator (LoadGen).

Because the main goal of this study was to examine power utilization from a typical enterprise Exchange environment, we selected an Exchange user protocol mix that would represent an average enterprise environment and sufficiently load all Exchange server roles (see Table 5).

The client protocol mix was 75 percent MAPI (Outlook 2007) and 25 percent Outlook Web Access. Exchange ActiveSync was run against 25 percent of the mailboxes. The incoming e-mail message rate was set to seven messages per second.

Table 5 Summary of Exchange user profiles tested

Return to top

When measured over a 24-hour period, the native Exchange servers consumed between 210 and 274 average watts. This data is summarized in Table 7. During the 24-hour period, the environment with eight physical servers consumed a total of 1,962 average watts. The projected annual power consumption for the native environment is 17,184 kilowatt hours (kWh)/year.

Note:
To calculate kW/year, the following equation was used: KWh/year = ((Total Average Watts * 24) / 1000) * 365

Table 7 Power utilization for the native Exchange server environment

Return to top

A physical Hyper-V root server configured with four guest virtual machines running Exchange 2007 roles (including one Client Access server, one Hub Transport server, and two Mailbox servers) consumed 491 average watts, measured over a 24-hour period. This data is summarized in Table 8. The total average number of watts consumed by the virtual environment, measured over a 24-hour period, was 982. The projected annual power consumption for the virtual environment was 8,602 kWh/year.

Table 8 Power utilization for the Hyper-V host server

The HP MSA60 storage solution, which consisted of 96 spindles in eight enclosures, consumed 672 average watts. This data is summarized in Table 9. The projected annual power consumption for this storage solution was 5,888 kWh/year.

Note:
The actual test environment consisted of four MSA60s and four MSA70s. For comparison, we calculated power consumption as if the complete solution used MSA60s.

Table 9 Power utilization for the native Exchange server environment using MSA60s

The HP MSA70 storage solution, which consisted of 192 spindles in eight enclosures, consumed 952 average watts. This data is summarized in Table 10. The projected annual power consumption for this storage solution was 8,340 kWh/year.

Note:
The actual test environment consisted of four MSA60s and four MSA70s. For comparison, we calculated power consumption as if the complete solution used MSA70s.

Table 10 Power utilization for the native Exchange server environment using MSA70s

Note:
Because identical storage solutions were used for both native and virtual Exchange environments, there was no difference in storage-related power consumption between native and virtual environments in this study.

Return to top

After these tests were finished, the data was analyzed and compared. The native Exchange 2007 servers, without accounting for storage utilization, consumed 1,962 average watts during the 24-hour test period. After the eight native Exchange 2007 servers were converted into eight virtual Exchange 2007 servers running on two physical Hyper-V root servers, the two root servers consumed 982 average watts during the 24-hour test period. This virtualization scenario resulted in a 50 percent reduction in power consumption as shown in the following figure. This translates into a projected savings of 8,582 kWh/year.

The MSA60 storage solution that was tested consumed 672 average watts during the 24-hour test period. If the storage power consumption is added to the server power consumption, 2,634 average watts were consumed by the native server environment, and 1,654 average watts were consumed by the virtual server environment. When the MSA60 storage solution was included, the virtualization scenario had a 37 percent reduction in power consumption.

The MSA70 storage solution consumed 948 average watts during the 24-hour test period. If storage power consumption is added to server power consumption, 2,910 average watts were consumed by the native server environment and 1,930 average watts were consumed by the virtual server environment. When the MSA70 storage solution was included, the virtualization scenario had a 34 percent reduction in power consumption.

Comparison of average watts consumed by native and virtual environments

Return to top