• Best Practices for Avoiding Performance and Scalability Problems
  
  
• Configuring the Server for Performance and Scalability
  
  
• Performance and Scalability Expectations
  
  
  • Unicasting
    
    
  • Multicasting
    
    

For information about analyzing blockages during an installation, see Troubleshooting Performance Problems.

The following are best practices that you can use:

• Ensure that the network interface between the server and client has sufficient bandwidth. Consider gigabit network adapters on the physical server with Category 5e (Cat 5e) cabling to a switch that can handle a GB back-plane connection, with 100-MB ports on the front-plane (100 MB-clients, 1-GB back end to the server).
  
  
• Use high-quality Ethernet cabling. We recommend at least Cat 5 or Cat 5e is recommended throughout the physical network.
  
  
• Use network switches. Do not use a hub.
  
  
• Partition network segments to distribute the load across multiple servers.
  
  
• Keep network latency to a minimum to optimize TFTP transfers.
  
  
• Ensure that the disk that contains the RemoteInstall folder has enough throughput to meet the client demand. On small-scale solutions, this may mean getting a Serial Advanced Technology Attachment (SATA) drive that spins at 7,200 RPM or faster. On mid-scale solutions, this may mean getting multiple drives and configuring them using Redundant Array of Independent Drives (RAID) configuration. On large-scale solutions, this may mean investing in a hardware RAID array. The disk volume that contains RemoteInstall should be separate from the system volume.
  
  
• Ensure that there is sufficient memory on the server to handle the demands. This may mean upgrading a server from 32-bit (x86) to 64-bit (x64). (for details about how to evaluate whether this solution is worthwhile for you, see Performance and Scalability Expectations).
  
  
• Ensure that there is enough processor bandwidth on the server to handle the demands. If the server has a lot of processes or services that are running, you may need to distribute the processes and services or upgrade the server’s processor.
  
  

Performance is the speed of a single client installation. In tests, Windows Deployment Services performs on par with a network-based installation from a file share. As expected, factors such as image size, network speed, and disk speed on the client affect the installation times. Typical installations using the standard Windows Vista image took around 20 minutes from first client boot to desktop.

A key benefit of using Windows Deployment Services is the ability to deploy to several clients simultaneously. Again, many factors influence the solution's ability to scale, but the most important ones are the following (in order from most to least influential):

1. Network bandwidth. Windows Deployment Services performs best using a 1-GB-per-second network adapter. In tests, a server with a 100-MB-per-second network adapter could perform a maximum of 10 simultaneous installations, while keeping the installation time under an hour (regardless of the server RAM, disk speed, or processor speed). By contrast, a high-end server with a 1-GB-per-second network adapter could install Windows images on 75 simultaneous clients in 45 minutes.
   
   
2. RAM on the server. If the computer has enough available memory, it is possible to cache an entire image into memory. This reduces the number of disk read/write operations and, in turn, speeds up the process. If several different images are being deployed concurrently, you may need more RAM.
   
   
3. Disk speed on the server. Disk speed is another factor that can slow down deployments (even when you have the maximum amount of RAM). The install image must be read from the disk at least once, and a faster disk speed can accelerate this process.
   
   
4. Disk speed on the client. A blockage in the client computer's disk may keep it from achieving the shortest possible installation times.
   
   

This section outlines the approximate amounts of time that elapsed during the image apply and TFTP download phases.

• Unicasting
  
  
• Multicasting
  
  

Unicasting

The following table outlines the hardware configurations of the servers that were used during these scalability tests.

Server type	Network interface card	Hardware configuration
Low-end	100-MB network adapter	Single-processor x86 1 GB of RAM 5,400-RPM disk interface
Middle	100-MB network adapter	Single-processor x86 2 GB of RAM 7,200-RPM disk interface
High-end	1-GB network adapter	Dual-processor x64 4 GB of RAM 10,000-RPM disk interface

	Note
	Network configuration for the high-end server involved connecting the server’s GB network adapter to a GB switch, which was connected to 100-MB switches that supported a GB back-plane configuration.

Time elapsed during the image apply phase

This table shows the approximate time (in minutes) from start to finish that it took for all of the clients to apply an install image.

Number of clients	Low-end	Mid-range	High-end
1	25	25	25
10	61	55	25
25	125	117	25
50	235	220	35
75	355	330	45

Time elapsed during the TFTP download phase

The following table shows the time (in seconds) it took to download Boot.wim using TFTP.

Number of clients	Low-end	Mid-range	High-end
1	70	55	40
10	210	145	75
25	450	360	120
50	910	805	270
75	1515	1400	420

Time elapsed when the TFTP block was increased

The following table shows the effect on time (in seconds) of changing the default TFTP block size. The times are cumulative for the total number of clients simultaneously downloading the same boot image.

Network adapter	Number of clients	Default	4,000	8,000	16,000	32,000
GB	50	270	180	118	92	85
GB	75	422	267	171	126	125
100 MB	75	1,410			1,140

Multicasting

Microsoft performed tests to compare the installation times of multicast and unicast transmissions using the same hardware, software, and image set. The following table outlines the configurations of the servers and clients that were used during these tests. The boot and install images were taken from an x86-based version of Windows Server 2008. The size of the boot image was approximately 128 MB, and the size of the install image was approximately 1.32 GB. Note that the out-of-box experience (OOBE) and logon were automated by using an unattend file.

	Network interface card	Hardware configuration	Operating system
Server	1-Gbps network adapter	Dual Xenon processor 5150 2.67 Ghz 8 GB of RAM	64-bit version of Windows Server 2008
Client	100 megabits network adapter	Varied but capable of installing the x86-based version of Windows Vista

Multicast Installation

	25 clients	100 clients	300 clients
	Restart computer and start clock.	Restart computer and start clock.	Restart computer and start clock.
Time when the first client started download of boot image using TFTP	:23	:21	:23
Time when the last client finishes download of boot image using TFTP	1:02	2:40	7:16
Time when the first client started the multicast transfer	3:04	3:55	8:18
Time when the last client finished the multicast transfer	6:06	7:54	12:30
Total amount of time until the last client reached the desktop	19:47	22:40	27:40

Unicast Installation

	SMB 25 clients	SMB 100 clients	SMB 300 clients
	Restart computer and start clock.	Restart computer and start clock.	Restart computer and start clock.
Time when the first client started download of boot image using TFTP	:21	:22	:20
Time when the last client finished download of boot image using TFTP	:58	2:40	7:13
Time when the first client started image transfer using unicast/SMB	3:14	4:38	8:29
Time when the last client started image transfer using unicast/SMB	13:36	38:15	1:47:58
Total amount of time until the last client reached the desktop	20:59	45:37	1:55:15

Testing of Security Options with Multicast

The following table lists the times for the start and end of the multicast transfer of the install image, and the percentage of the CPU used for the multicast transfer, depending on the level of security that was enabled during the test. This test involved 25 client computers.

	No Security	Hashing (default)	Signing
Start of multicast transfer of the install image to the client	Clock started	Clock started	Clock started
End of multicast transfer of the install image to the client	2:19	2:27	31:05
Percentage of CPU used during the multicast transfer	~5%	~11%	~25%