Virtual Hard Disk Performance

Windows Server 2008 / Windows Server 2008 R2 / Windows 7

Authors:
Liang Yang
Anthony F. Voellm

1
This document is provided “as-is.” Information and views expressed in this document, including URL and other
Internet Web site references, may change without notice. You bear the risk of using it.

This document does not provide you with any legal rights to any intellectual property in any Microsoft product. You
may copy and use this document for your internal, reference purposes.

© 2010 Microsoft Corporation. All rights reserved.

Microsoft, Windows, Windows Server, Hyper-V are either registered trademarks or trademarks of Microsoft
Corporation in the United States and/or other countries.

The names of actual companies and products mentioned herein may be the trademarks of their respective owners.

2
Contents
1. Introduction.........................................................................................................................................4
2. Performance Measurement Methodologies.......................................................................................6
A. Test Details......................................................................................................................................6
Choice of Targets: Raw Disk vs. Regular File vs. VHD...........................................................................6
Windows hypervisor impact on performance measurement..............................................................7
Choice of Workloads............................................................................................................................7
Choice of Performance Metrics...........................................................................................................8
Choice of Target File Size.....................................................................................................................9
Choice of Data Block Size.....................................................................................................................9
Choice of VHD Attach Mode................................................................................................................9
Reducing variation in measurements................................................................................................10
B. Test tools.......................................................................................................................................10
C. Server Platform and Operating System..........................................................................................10
D. Storage Platform Settings..............................................................................................................10
Storage Platform Hardware Settings.................................................................................................10
Storage Platform Software Settings...................................................................................................11
E. Virtual Hard Disk (VHD) Details......................................................................................................12
F. File System and Volume Settings...................................................................................................14
3. Comparison of Windows Server 2008 and Windows Server 2008 R2 VHD Performance..................15
G. Fixed Virtual Hard Disk creation speed..........................................................................................15
H. Fixed Sized VHD Performance Comparison...................................................................................16
SQL Server Log Workload..................................................................................................................16
Exchange Server Workload................................................................................................................17
I. Dynamically Expanding VHD Performance Comparison................................................................18
Media Streaming Workload...............................................................................................................19
Online Transaction Processing Workload..........................................................................................20
J. Differencing VHD Performance Comparison..................................................................................21
Web Server Log Workload.................................................................................................................22
Decision Support System Database Workload...................................................................................23
Performance Impact of Differencing Chain Length............................................................................24
4. Comparison of Native and Virtual Machine VHD Performance.........................................................25
K. Hyper-V Virtual Machine Performance Testing Settings................................................................25
5. Comparison of VHD type performance [fixed sized vs. dynamically expanding vs. differencing]......28
6. How to choose your Hyper-V and VHD Storage Container Format...................................................30
7. Summary of supported and practical limits.......................................................................................32
8. Closing...............................................................................................................................................34
9. Acknowledgements...........................................................................................................................35
10. References.....................................................................................................................................35

3
 1. Introduction
A virtual hard disk (VHD) is a file that encapsulates a hard disk image. VHDs can be used in new and
interesting ways. VHDs first were created to be the storage media for virtual machines (VMs). Today,
VHDs are used to ship trial versions of software, used in backup solutions, used for bug triage (e.g.
customers can convert a physical disk to virtual and share it), and even used to store multiple boot
environments. VHDs are a very flexible storage container and are not tied to any single file system
format. Since June 2005, Microsoft has made the VHD Image Format Specification available to third
parties under the Microsoft Open Specification Promise (OSP).

Microsoft began using VHD technology in Microsoft Virtual PC around 2003, and then continued its use
in Microsoft Virtual Server release in 2005. The next major release happened as part of Hyper-V in
Windows Server 2008. Currently VHD support is made as part of Windows Server 2008 R2 and high-end
client SKUs of Windows 7. VHDs were limited to use by virtual machines running in Virtual PC/Virtual
Server/Hyper-V and loopback mounting of VHDs in the parent partition sometimes referred to as the
management operating system. The integration of VHD support into the operating system was
drastically improved in Windows Server 2008 R2 which added native support. Native support means the
technology is integrated into the operating system and no longer requires a virtualization solution such
as Hyper-V to be available. Native support added the following features:

 boot from VHD

 integrated support for attaching and detaching VHDs via inbox APIs (virtdisk.dll) and the “Disk
Management” control panel (diskmgmt.msc or command line tool DiskPart (“attaching” is the
term used to describe the action of mounting a VHD so it can be used by Windows directly, or
as a disk in the guest operating system of a virtual machine).
 attaching VHDs from inside VHDs, and many performance improvements.

There are three VHD formats each with different performance characteristics. The three formats of
VHD are fixed, dynamic and differencing.

Fixed Sized Virtual Hard Disk

On disk storage space needed VHD Size

4
A fixed sized VHD uses a file in which the space to store the file is allocated on the physical storage when the
virtual hard disk is created. The file size is the same as the size specified for the virtual hard disk. As their name
implies, fixed sized VHDs occupy the same space on the underlying physical storage device as their specified size.
However, once a fixed sized VHD is created, the size can be increased when the disk is offline by editing the disk to
expand it. Reducing the size is not supported. Because the physical storage required for a fixed size VHD is
allocated when the VHD is created, there is a better chance at optimal placement and organization on-disk which
yields the best performance. The disadvantage is the space is committed even if it is not used.

Dynamically Expanding Virtual Hard Disk

On disk storage space needed

VHD Size

A dynamically expanding VHD is a file that at any given time is as large as the actual data written to it
plus the size of on-disk meta-data. Dynamically expanding disks are useful because they do not require
all the storage needed to contain the maximum size of the disk to be reserved up front. The VHD file
starts quite small (e.g. 42KB is a typical physical size of an empty 20GB disk) and grows as new blocks in
the disk are used. There are a number of optimizations around dynamically expanding disks that
improve performance; however, in general their read/write performance is slower than fixed disks. One
optimization is the selection of data block size which can be either 512KB or 2MB; another is skipping
allocation of all-zero blocks.

Differencing Virtual Hard Disk

On disk storage space needed

  …  

  VHD Size

root Diff 0 Diff N

A differencing VHD is a file representing the current state of the virtual hard disk as a set of modified
blocks in comparison to a parent virtual hard disk. Differencing VHDs can be associated with either a
fixed sized or dynamically expanding VHD. Differencing VHDs can also be associated with another

5
differencing VHD but they cannot be associated with a physical disk. Differencing VHDs are used to
prevent changes from being made in their parent VHD to which they are applied and are used to
implement a number of additional features. In Hyper-V, differencing VHDs are also created
automatically whenever snapshots are taken of a virtual machine. Note differencing VHDs used for
snapshot purpose are named with an AVHD file extension to help users easily distinguish them from
regular differencing VHDs. Differencing VHDs may also be used to deploy a “golden” or “master” image,
because you can associate multiple differencing VHDs to one parent VHD. Some disadvantages of
differencing VHDs are increased caching needs and the inability to grow or shrink the VHD size. You can
however compact differencing VHDs to reclaim physical space usage.

There are several important limitations for VHDs:

 VHDs can be mounted only on NTFS volumes (although you can still save a VHD file on
FAT/FAT32 assuming the maximum file size limit is not violated). For example, if you have a
differencing VHD chain, then every VHD along the chain must sit on an NTFS volume to make
VHD attaching work.
 VHDs cannot be mounted within a compressed folder in Windows Server 2008 R2. This was
possible in Hyper-V role in Windows Server 2008, but this capability was explicitly blocked in
the Hyper-V role in Windows Server 2008 R2 since the compressed file size limit is relatively
small. A dynamically expandable VHD can easily outgrow that limit and get corrupted.
 In addition to the maximum file size of NTFS, dynamic or difference VHDs cannot exceed
2040GB. The reason for the 2040G limit is the length of each Block Allocation Table entry is set
to 4 Bytes and the maximum valid value is 0xFFFFFFFE (0xFFFFFFFF means an unused entry). If
you multiply that value by 512B sector size and then subtract the overhead of on disk meta-data
structures, 2040G will be the maximum size of dynamically or differencing VHDs.

The remainder of this paper is focused on giving in-depth performance data on Virtual Hard Disks.
Comparisons are generally made between native, Window Server 2008 and Windows Server 2008 R2.
When comparing Windows Server 2008 and Window Server 2008 R2, no virtual machines are used
unless explicitly stated. The goal of this paper is to look at virtual hard disk performance and not Hyper-
V performance.

2. Performance Measurement Methodologies

A. Test Details
Storage performance testing is the science used to determine the characteristics of storage I/O
subsystem. The performance results are always related to certain software and hardware platform
settings within a controlled experimental environment. We take into consideration the following factors
and assumptions to make meaningful VHD performance measurements and comparisons which are
crucial to obtain valid conclusions.

Choice of Targets: Raw Disk vs. Regular File vs. VHD

We use raw disk (aka physical disk or bare-metal disk) and raw file (i.e. a normal non-VHD file) to
compare with native VHD performance results. The benefits of adding these two baselines can help us

6
quickly determine if a performance issue or performance improvement is solely VHD-related or comes
from other Windows components.

 A raw disk provides direct access to the storage media instead of through its file system and is
considered the thinnest layer in terms of sending I/Os down to the disk. This can give us a good
estimation of what the maximum throughput and minimum CPU cycles a VHD can achieve.
Occasionally a raw disk does not give the best performance due to additional optimization (e.g.
I/O batching) done by other layers, but so far it has been proved a sound guard for sanity
validation purposes.
 A raw file shares the common I/O paths with a VHD file except those added by the VHD driver
stack. By comparing a VHD with a raw file, we can find out how much overhead the VHD adds on
top of a raw file. Ideally we want a VHD performance to be on-par with a raw file for the major
performance metrics.

Windows hypervisor impact on performance measurement

In Windows Server 2008 R2, VHD support is native to the operating system and does not depend on the
presence of the Windows hypervisor. From a throughput perspective, the impact of the Windows
hypervisor is quite small based on past experimental results. This is mainly due to the fact that
performance critical workloads are re-directed to synthetic VMBus channels instead of using the longer
emulation path. To get the most accurate CPU utilization and to focus on native performance, the
Windows hypervisor is turned on only during VHD performance measurement in Windows Server 2008
which is required to mount VHDs on a Windows Server 2008 machine while it remains off for all other
performance testing scenarios.

Choice of Workloads
We used two types of performance workloads. The advantage of these workloads is that they are
neutral to operating systems and underlying storage platforms. This makes it easier to compare VHD
with other virtual disk formats.

 Monolithic I/O-Based Workloads

The workloads here consist of single type of I/O. We have been using these monolithic types of
I/O workloads since they are relatively simple and quick to configure and run, and give a good
baseline figure to easily identify the impact of changes during the development cycle.

Sequential Reads & Writes 4K 64K 256K 1M

Random Reads & Writes 4K 8K 32K 64K

The two overlapping I/O sizes (4K/64K) here also give us another opportunity to compare the
performance of sequential I/O with random ones.

 Application-Based Workloads
We picked some of most commonly used server workloads to simulate enterprise customer
scenarios. They are well-recognized by the storage community and adopted by the industry.

7
 Workload Category I/O Size Percentage of READ Percentage of RANDOM vs.
vs. WRITE SEQUENTIAL

Web File Server 4KB 95% RD vs. 5% WR 75% RAND vs. 25%  SEQ

Web File Server 8KB 95% RD vs. 5% WR 75% RAND vs. 25%  SEQ

Web File Server 64KB 95% RD vs. 5% WR 75% RAND vs. 25%  SEQ

Decision Support 1MB READ RANDOM

System DB

Media Streaming 64KB 98% RD vs. 2% WR SEQUENTIAL

SQL Server Log 64KB WRITE SEQUENTIAL

OS Paging 64KB 90% RD vs. 10% WR SEQUENTIAL

Web Server Log 8KB WRITE SEQUENTIAL

OLTP DB 8KB 70% RD vs. 30% WR RANDOM

Exchange Server 4KB 67% RD vs. 33% WR RANDOM

Workstation 8KB 80% RD vs. 20% WR 80% RAND vs. 20% SEQ

Video on Demand 512KB READ RANDOM

               

 I/O Queue Depth: 1 to 256

Measuring I/O performance with various queue depths is another key to scrutinizing how native
VHD performs with underlying platforms, e.g. how quickly we can hit the saturation point. It can
also help determine how efficiently the native VHD driver stack handles outstanding I/Os as the
queue depth increases. In addition, it can help us determine if there are any hidden
performance issues when a certain pattern is not observed in the results. For example, before
the target is saturated, generally we should see the throughput increase as more outstanding
I/Os are issued.

Choice of Performance Metrics

Two major performance metrics are considered in this report: throughput and latency.

Throughput is measured either in terms of I/Os per second (IOPS) or megabyte per second (MBPs).
Typically under a monolithic I/O workload, for small sized I/Os, we intend to use IOPS while reserving
MBPS for larger sized I/Os. For Application based workloads, we use IOPS exclusively for performance
comparisons. Higher throughput means better performance.

8
Latency (average I/O response time) is measured in terms of milliseconds. Combined with throughput,
we can use latency to easily identify the storage saturation point when throughput gets flat and latency
keeps going up. Lower latency means shorter response time and better performance.

Choice of Target File Size

Since we’re comparing raw disk vs. raw file vs. VHD, it becomes necessary to set preferred size to be
equal or very close to each other to make such type of comparison meaningful.

 Random performance is largely impacted by the target size.

 Minimizing the performance impact of the on-disk layout for files is another concern (i.e.
there can be a performance gap between an outer track of a physical disk drive and an inner
track). Creating a large file to occupy the entire disk guarantees a consistent disk layout.
 Minimizing the impact of hardware caches (depending on the RAM size) is also important so
the entire file cannot be easily fit or pre-fetched into the cache to get false performance
results.

Choice of Data Block Size

The default data block size for dynamically expanding and differencing VHDs changed from 512KB in
Windows Server 2008 to 2MB in Windows Server 2008 R2. We made the change in default block size
based on extensive testing across numerous scenarios. Based on our experimental results we found
2MB to be a better default. This does not mean it is better for all scenarios and in particular there are
some simple primitive tests where 512KB is a better choice. However the primitive tests are generally
not representative of what real workloads do.

This leads to the question when comparing Windows Server 2008 and Windows Server 2008 R2 if we
should use the default setting or use the same block size in all tests. We chose to use the default
because this is the finely tuned parameter and we want to compare the out-of-box performance
because this is what most users will experience.

Choice of VHD Attach Mode

The goal of this report is to focus on VHD driver stack performance and treat VHDs just like regular disk
drives. Measuring VHD performance directly on parent partition has the following benefits:

 Only VHD related components are involved. No Hyper-V storage stack is needed.
 The performance results are comparable to a raw file sitting on top of physical file system.
 The performance impact of a guest operating system is eliminated.

Unless indicated otherwise, the VHD being tested is attached as a loopback drive on the parent partition,
and then the loopback drive is used as a raw disk during the performance measurement process (just
like how we measure the performance of a physical drive). We use virtual machine mode when there is
a need to compare performance between the parent partition and a virtual machine.

9
Reducing variation in measurements
In reality, an enterprise server may be engaged in many computing activities simultaneously when
dealing with the disk I/O subsystem. To minimize any interference from other applications and get
accurate performance results, we take explicit action to reduce the overhead of other applications and
services. For example, if we leave Hyper-V Manager open during I/O performance measurement within
VM, Hyper-V Manager will periodically communicate with the VM to get the screen contents update.
This can produce skewed results varying the accuracy by as much as 5 to 10%.

B. Test tools
IOMeter is a popular I/O benchmarking tool in storage community and it is also the tool used to
measure native VHD performance. Here are the major IOMeter settings being used in this report:

 Version: 2008-06-22RC2(x64)
 Ramp-up Time: 30 sec
 Test Run Time: 60 sec
 Number of IOMeter Worker: 1
 I/O Alignment Size: Sector(512B)

C. Server Platform and Operating System

We used an Intel Nehalem-EP server with dual quad-core (2128MHZ) processors, 6GB RAM with NUMA
enabled and a 1TB SATA RAID0 based system drive. Windows Server Enterprise 2008 (x64) and
Windows Server Enterprise 2008 R2 (x64) are operating systems were used for all performance
experiments.

D. Storage Platform Settings

To simplify the performance measurement environment, we used Direct–Attached Storage (DAS) from
Dell as our target. DAS is not as common as SAN under enterprise scenarios. However, DAS typically has
fewer performance factors (e.g. block size, thin provisioning, de-duplication etc.) to get involved so we
can purely focus on software stack performance and eliminate platform dependent factors as far as
possible. The performance results collected on DAS are generally more sustainable due to the similar
reasons. Certain features existing in SAN environment may impact actual VHD performance results. For
example, you may see performance penalties caused by misalignment of dynamic VHDs on certain type
of SANs which may not show up on a DAS platform

Here are the general settings:

Storage Platform Hardware Settings

 Enclosure: Dell PowerVault MD1000
 Capable of holding up to 15 3.5” drives without expansion
 Single path without built-in RAID controller cards

 Hard Disk Drives: Seagate Cheetah SAS Drive (Quantity:15)

 Model: ST3146356SS
 15000 RPM, 3.5”, 146GB, 16M Disk Cache

10
  Sustained data transfer rate for single drive is approximately 160MBps.
 Native physical sector size: 512Bytes

 RAID Controller Card: LSI 8880EM2

 LSI1078 Storage Processor, PowerPC 500MHZ
 Supports 3Gbps SAS/SATA drives
 Supports eight external SAS/SATA ports
 512MB DDR2 Cache (667MHZ),
 PCIe x8
 Based on performance measurements, the maximum bandwidth the LSI 8880EM2
(mainly determined by storage processor) can sustain is approximately 1600MBps for
reads 700MBps for writes.

Storage Platform Software Settings

 Seagate SAS drive firmware: HS0C
This was the latest firmware available for these drives at the time of publication.

 RAID Controller firmware: Mega RAID(3.6) 11.0.1-0013

This was the latest firmware published by LSI we could use to flash this RAID controller card and
keep it up-to-date.

 LSI Mega SAS Driver

Making sure both Windows Server 2008 R2 and Windows Server 2008 use the same third party
mini-port driver is crucial for our performance comparison purpose.

 Windows Server 2008 R2 ships with the latest driver version available inbox at the
time of writing, 4.5.0
 Windows Server 2008 ships with version 2.3.0. We were able to upgrade to 4.4.0
and there is no performance difference between 4.4.0 and 4.5.0.

 RAID Array Settings

We used a 15-spindle RAID0 configuration. RAID array properties are detailed below. A 256K
stripe unit size was chosen based on internal performance investigations for LSI RAID controller
cards. We’re attempting to bypass cache effects along the I/O path by using write-through for
disk controller cache and disabling disk drive cache.

 RAID Level: RAID 0

 Size:2190GB (146GB x 15)
 Stripe Size: 256 KB per disk
 Number Of Drives:15
 Cache Policy: Write Through, Read Ahead, Direct I/O
 Disk Cache Policy: Disabled
Note: enabling disk cache does bring up additional performance benefits as long as file system
flush and disk controller flush frequency are well set. In production environment, we do
recommend customers to enable controller cache when a battery backup unit is available.

11
 We chose RAID0 for two reasons:

 RAID0 gives the best performance amongst all RAID levels

 No parity calculation is required, and thus no additional I/Os are generated for each
write. This can help simplify performance overhead calculations.

E. VHD Details
The introduction gave an overview description of the different types of VHD. This section gives more
detail on the format and impact on overall performance and usage. The full VHD specification is
available online at http://technet.microsoft.com/en-us/virtualserver/bb676673.aspx .

 Fixed sized VHDs

The file structure of fixed sized VHDs is relatively straightforward. Fixed sized VHDs do not contain
any data blocks and they are essentially a flat file plus footer. We recommend using fixed size VHD
when performance is the top concern for customers. From an implementation perspective, the VHD
driver itself does not impose any file size limitation (except the one from NTFS, i.e. 16TB). In this
report, we created a fixed VHD which is about 2040GB, the maximum size supported by Hyper-V
Manager UI.

 Dynamically expanding VHDs

Dynamically expanding VHDs can grow dynamically without requiring space allocation up-front. The
following table shows the actual physical space occupied by a fully populated 2040GB dynamically
expanding VHD. In Windows Server 2008 R2, we changed the default block size from 512KB to 2MB
for performance reasons. However, the block size change also has an impact on the actual physical
file size. Smaller blocks size means more data blocks are needed for a pre-defined virtual capacity.
Thus more space has to be reserved for the Block Allocation Table (4 bytes per entry) and Sector Bit
Map (512B regardless of block size for alignment purpose). For example, a 2040GB dynamically
expanding VHD in Windows Server 2008 has to reserve 16MB for BAT and 2GB for sector bit maps.
From the dynamically expanding VHD perspective, these sector bit maps are really not as useful as
they are in a differencing VHD and could be optimized.

2040GB Dynamically 2040GB Dynamically

Expanding VHD in Windows Expanding VHD in Windows
Server 2008 R2 Server 2008
VHD Capacity 2040GB 2040GB
Data Blocks within VHD 1044480 4177920

BAT (Block Allocation Table) Approximately 4M(default: Approximately 16M(default:

2M/Block) 512K/Block)
Sector Bit Maps Approximately 510M Approximately 2G
(512B/Block)
Footer/Header 2K 2K

12
 Differencing VHD
A Differencing VHD is very much like the dynamic VHD as they are both dynamically expandable.
Their internal structures remain largely the same although runtime interpretation (e.g. sector
bitmap) may be different. You can convert a differencing VHD to a dynamic VHD by choosing to
merge to a new file. But there are also notable differences between them. For example, you can
expand a dynamic VHD (in terms of virtual size) but that is not allowed for differencing VHDs due to
the inherent size limitation inherited from the parent. A differencing VHD also requires the presence
of its parent to perform I/O activity.

Differencing VHDs are the foundation for Hyper-V VM snapshots and many other customer
scenarios (e.g. multiple client images share the same golden parent image). In the previous section,
we laid out details about physical space usage for fully populated dynamically expanding VHDs. The
Differencing VHD format has added a little bit more overhead in terms of actual physical file size
because it has to use built-in parent locators to save the parent path information. The absolute
parent locator path within a differencing VHD is set to 512B by default and the relative parent
locator path is set to 64KB (this one is indeed optional if there the differencing VHD does not share
any common path with the parent [for example when they are located on different volumes]). When
the native VHD stack tries to open a differencing VHD, both the differencing and the parent VHD are
opened. The parent locator information may get updated when the parent location information is
changed. For example, in case the differencing chain is broken, a reconnect operation lets you
specify the location of parent and re-establish the link between parent and child.

When accessing differencing VHDs, parent VHDs are opened in read only mode by the driver and the
child VHDs are opened in read/write mode so all the changes will stay within the child. However, the
VHD format and VHD driver assume that parent VHD will not change and have no inherent
protection against that assumption. In other words, it is possible to modify a parent VHD which will
corrupt the relationship between it and any child VHDs. A differencing VHD merge into a parent also
corrupts all other sibling VHDs not on the merging path regardless of merging success or not. The
child and the other siblings may never be made aware of if there are changes made in their parent
and results may become unpredictable.

A differencing chain can consist of one or more VHDs. However, the VHD specification and native
VHD driver does not requires the data block size for the entire chain be the same. For best practices,
we recommend using 2MB for both the parent and child VHDs to improve performance as the reads
to the parent may be impacted due to a smaller block size. For certain scenarios like booting a
differencing VHD on a physical machine, the child and parent VHDs must use the same data block
size (note that this rule does not apply to booting a Virtual Machine in Hyper-V). In Windows Server
2008 R2, differencing VHDs created via DiskPart or Hyper-V Manager always uses the same data
block size as the parent. That means you will get a 512KB differencing VHD instead of the default
2MB even in Windows Server 2008 R2 if that is the data block size of the parent VHD. If parent is a
fixed sized VHD, then the default data block size will be always used for the differencing VHD.

13
 During performance experiments, the differencing chain length is set to one, i.e. one parent and one
differencing VHD (child). A 2040G dynamically expanding VHD is chosen as the parent and remains
unpopulated. We then fully populate the differencing VHD to make it as big as possible on the
physical volume.

F. File System and Volume Settings

VHD only works with NTFS and is generally considered as a large file in terms of average file size. In this
report, we always chose 64K as the NTFS cluster (Allocation Unit) size for VHD and Non-VHD file
performance measurement. That also aligns with SQL Server performance best practices when a large
database file is used. All the physical volumes remain clean: a fresh format is performed before creating
VHD files to minimize the performance impact of fragmentation.

14
 3. Comparison of Windows Server 2008 and Windows Server 2008 R2
VHD Performance
The following section compares Windows Server 2008 and Windows Server 2008 R2 overall VHD
performance. We look at creation speed, various workloads, impact of queue depth, block size, chain
depth and other parameters that impact overall VHD performance.

A. Fixed Size VHD creation speed

Creating new VHDs generally happens when a virtual machine is provisioned or a new boot environment
is needed. The time it takes to create dynamically expanding VHDs and differencing VHDs is on the
order of a couple of seconds. However, fixed sized VHDs require the entire size of the virtual disk to be
written during creation. This means the fixed sized VHD creation speed is determined by the time to
zero out all the space allocated up front.

Below is a 2040GB fixed VHD creation speed comparison based on a directly-attached storage between
Windows Server 2008 and Windows Server 2008 R2. Note: the speed improvement may vary on
different storage platforms and VHD file size.

Fixed VHD Creation Speed Comparison(VHD Size: 2040GB)

450

400

350

300
WS08
250 WS08 R2
MBps

200

150

100

50

15
 G. Fixed Sized VHD Performance Comparison
The fixed sized VHD performance has been on-par with the physical disk since Windows Server
2008/Hyper-V release to manufacturing In Windows Server R2 fixed VHD performance remains intact,
i.e. it as good as raw disk or raw file.

SQL Server Log Workload

 64KB I/O size, 0% Read, 100% Write, 0% Random, 100% Sequential

SQL Server Log 64KB 100%Sequential 100%Write Throughput

12000

10000
WS08_RawDisk
8000 WS08_RawFile
WS08_FixedVHD
IOPS

6000
WS08R2_RawDisk
4000 WS08R2_RawFile
WS08R2_FixedVHD
2000

0
1 2 4 8 16 32 64 128 256
I/O Queue Depth

16
 SQL Server Log 64KB 100%Sequential 100%Write Latency
30.0000

25.0000
WS08_RawDisk
20.0000 WS08_RawFile
Latency(ms)

WS08_FixedVHD
15.0000
WS08R2_RawDisk
10.0000 WS08R2_RawFile
WS08R2_FixedVHD
5.0000

0.0000
1 2 4 8 16 32 64 128 256
I/O Queue Depth

Exchange Server Workload

 4KB I/O size, 67% Read, 33% Write, 100% Random

Exchange Server 4KB 100%Random 67%Read 33%Write Throughput

6000

5000
WS08_RawDisk
4000 WS08_RawFile
WS08_FixedVHD
IOPS

3000
WS08R2_RawDisk
2000
WS08R2_RawFile

1000 WS08R2_FixedVHD

0
1 2 4 8 16 32 64 128 256
I/O Queue Depth

17
 Exchange Server 4KB 100%Random 67%Read 33%Write Latency
60.0000

50.0000
WS08_RawDisk
40.0000 WS08_RawFile
Latency(ms)

WS08_FixedVHD
30.0000
WS08R2_RawDisk

20.0000 WS08R2_RawFile
WS08R2_FixedVHD
10.0000

0.0000
1 2 4 8 16 32 64 128 256
I/O Queue Depth

H. Dynamically Expanding VHD Performance Comparison

The following table shows a summary of peak IOPS comparison for a fully populated 2040GB
dynamically expanding VHD using a Dell MD1000 DAS. Note: the queue depth corresponding to the peak
IOPS may differ when comparing Windows Server 2008 R2 with Windows Server 2008.

Peak IOPS Windows Windows Peak IOPS Windows Windows Server

Server 2008 Server 2008 R2 Server 2008 2008 R2
4K Sequential 52070 92067 4K Random 1667 5165
Reads Reads
4K Sequential 7528 43892 4K Random 157 2611
Writes Writes
64K Sequential 24599 24631 8K Random 1734 5064
Reads Reads
64K Sequential 1083 10302 8K Random 155 3117
Writes Writes
256K 6262 6288 32K 1642 4405
Sequential Random
Reads Reads
256K 371 2500 32K 146 3079
Sequential Random
Writes Writes
1M Sequential 1565 1563 64K 1506 3761
Reads Random
Reads
1M Sequential 85 616 64K 137 2888
Writes Random

18
 Writes

19
Media Streaming Workload
 64KB I/O size, 98% Read, 2% Write, 0% Random, 100% Sequential

Media Streaming 64KB 100%Sequential 98%Read 2%Write Throughput

14000

12000
WS08_RawDisk
10000
WS08_RawFile
8000
WS08_DynamicVHD
IOPS

6000 WS08R2_RawDisk

4000 WS08R2_RawFile

WS08R2_DynamicVHD
2000

0
1 2 4 8 16 32 64 128 256
I/O Queue Depth

Media Streaming 64KB 100%Sequential 98%Read 2%Write Latency

35.0000

30.0000
WS08_RawDisk
25.0000 WS08_RawFile
Latency(ms)

20.0000 WS08_DynamicVHD
WS08R2_RawDisk
15.0000
WS08R2_RawFile
10.0000 WS08R2_DynamicVH
D
5.0000

0.0000
1 2 4 8 16 32 64 128 256
I/O Queue Depth

20
Online Transaction Processing Workload
 8KB I/O size, 70% Read, 30% Write, 100% Random, 0% Sequential

OLTP 8KB 100%Random 70%Read 30%Write Throughput

6000

5000
WS08_RawDisk
4000 WS08_RawFile
WS08_DynamicVHD
IOPS

3000
WS08R2_RawDisk
2000
WS08R2_RawFile

1000 WS08R2_DynamicVHD

0
1 2 4 8 16 32 64 128 256
I/O Queue Depth

OLTP 8KB 100%Random 70%Read 30%Write Latency

700.0000

600.0000
WS08_RawDisk
500.0000 WS08_RawFile
Latency(ms)

400.0000 WS08_DynamicVHD
WS08R2_RawDisk
300.0000
WS08R2_RawFile
200.0000 WS08R2_DynamicVHD

100.0000

0.0000
1 2 4 8 16 32 64 128 256
I/O Queue Depth

21
 I. Differencing VHD Performance Comparison
The following table shows a summary of peak IOPS comparison for a fully populated 2040G differencing
VHD (depth: 1) using a Dell MD1000 DAS. Note: the queue depth corresponding to the peak IOPS may
differ when comparing Windows Server 2008 with Windows Server 2008 R2.

Peak IOPS Windows Server Windows Peak IOPS Windows Windows Server
2008 Server 2008 R2 Server 2008 R2
2008
4K Sequential 39908 93352 4K 159 5167
Reads Random
Reads
4K Sequential 7320 44933 4K 156 2673
Writes Random
Writes
64K 6782 24651 8K 157 5033
Sequential Random
Reads Reads
64K 1076 10321 8K 155 3129
Sequential Random
Writes Writes
256K 1414 6294 32K 150 4410
Sequential Random
Reads Reads
256K 364 2511 32K 146 3077
Sequential Random
Writes Writes
1M 355 1565 64K 142 3760
Sequential Random
Reads Reads
1M 84 629 64K 136 2904
Sequential Random
Writes Writes

22
Web Server Log Workload
 8KB I/O size, 0% Read, 100% Write, 0% Random, 100% Sequential

Web Server Log 8KB 100% Sequential 100%Write Throughput

60000

50000
WS08_RawDisk
40000 WS08_RawFile
WS08_DiffVHD
IOPS

30000
WS08R2_RawDisk
20000
WS08R2_RawFile

10000 WS08R2_DiffVHD

0
1 2 4 8 16 32 64 128 256
I/O Queue Depth

Web Server Log 8KB 100% Sequential 100%Write Latency

60.0000

50.0000
WS08_RawDisk
40.0000 WS08_RawFile
Latency(ms)

WS08_DiffVHD
30.0000
WS08R2_RawDisk
20.0000 WS08R2_RawFile
WS08R2_DiffVHD
10.0000

0.0000
1 2 4 8 16 32 64 128 256
I/O Queue Depth

23
Decision Support System Database Workload
 1MB I/O size, 100% Read, 0% Write, 100% Random, 0% Sequential

DSS-DB 1MB 100%Random 100%Read Throughput

600

500
WS08_RawDisk
400 WS08_RawFile
WS08_DiffVHD
IOPS

300
WS08R2_RawDisk
200
WS08R2_RawFile

100 WS08R2_DiffVHD

0
1 2 4 8 16 32 64 128 256
I/O Queue Depth

DSS-DB 1MB 100%Random 100%Read Latency

7000.0000

6000.0000
WS08_RawDisk
5000.0000
WS08_RawFile
Latency(ms)

4000.0000 WS08_DiffVHD
WS08R2_RawDisk
3000.0000
WS08R2_RawFile
2000.0000
WS08R2_DiffVHD
1000.0000

0.0000
1 2 4 8 16 32 64 128 256
I/O Queue Depth

24
Performance Impact of Differencing Chain Length
In Windows Server 2008, experiments uncovered a critical performance issue in differencing VHD:
performance of accessing VHDs that belong to a chain of differencing disks degraded very quickly as
differencing chain length increased. This issue has been addressed in Windows Server 2008 R2. The
following performance results show the read performance improvements (all writes are directed only to
the last VHD in the chain) to the top level VHD as the differencing chain length increases.

Top Level VHD 64KB Sequential Reads Throughput

800.00

700.00

600.00

500.00 WS08
WS08
MBps

400.00
R2
300.00

200.00

100.00

0.00
1 2 4 8 16 32 64
Diff VHD Chain Length

Top Level VHD 4KB Random Reads Throughput

12.00

10.00

8.00
WS08
WS08
MBps

6.00 R2

4.00

2.00

0.00
1 2 4 8 16 32 64
Diff VHD Chain Length

25
 4. Comparison of Native and Virtual Machine VHD Performance
We have been focusing on the performance comparison between physical disk, raw file and a native
VHD all residing on the parent partition. However, VHD was originally designed and developed as the
storage container format for virtualization environments. Virtual Machines remain the most common
use for VHD deployment. In this section, we compare the I/O performance between native storage
(physical disk and native VHD) and virtualized storage (pass-through disk and Virtual Machine VHDs).

J. Hyper-V Virtual Machine Performance Testing Settings

 Guest Operating System: Windows Server Enterprise 2008 R2 x64
 RAM: 1024MB
 Number of Virtual Processors: 1
 Guest Storage Controller
o IDE: system drive for booting guest OS
o SCSI: data drive for performance measurement

26
 WS08 R2 Sequential I/Os Throughput (Native vs. Virtual Machine)
1800.00
1600.00
1400.00
1200.00
1000.00 PhysicalDisk(Native)
800.00
PassthruDisk(VM)
600.00
400.00 FixedVHD(Native)
MBps

200.00 FixedVHD(VM)
0.00

(Queue Depth: 64)

WS08 R2 Random I/Os Throughput (Native vs. Virtual Machine)

4000
3500
3000
2500
PhysicalDisk(Native)
2000
PassthruDisk(VM)
1500
1000 FixedVHD(Native)
IOPS

500 FixedVHD(VM)
0

(Queue Depth :64)

27
 WS08 R2 Real Apps Workload Throughput (Native vs. Virtual Machine)
18000
16000
14000
12000 PhysicalDisk(Native)
PassthruDisk(VM)
10000
FixedVHD(Native)
8000
FixedVHD(VM)
6000
IOPS

4000
2000
0

(Queue Depth: 64)

The graphs here show the performance gap between a physical disk, pass-thru disk and a fixed VHD
(either attached to native or a VM) is small. The potential performance difference not shown here, may
come from the path length (i.e. CPU cycles per I/O) as a pass-through disk generally requires less CPU
resources due to the shorter path compared with fixed sized VHD.

28
 5. Comparison of VHD type performance [fixed sized vs. dynamically
expanding vs. differencing]
Dynamically expanding VHDs are not recommended for virtual machines that run server workloads in a
production environment. Fixed VHD is preferred considering it has better performance with committed
space allocation so it will not run out of physical backing storage.

The following graphs show, in Windows Server 2008 R2, the gap between a fixed VHD and a fully
populated dynamic or differencing VHD is fairly small. The major performance gap comes from small I/O
writes. However, that gap may be widened if the VHD has to expand itself first before taking any user
data.

WS08 R2 Sequential I/Os Throughput (Native Mode)

1800.00
1600.00
1400.00
1200.00
1000.00 PhysicalDisk
800.00 FixedVHD
600.00
DynamicVHD
400.00
MBps

200.00 DifferencingVHD
0.00

(Queue Depth: 64)

WS08 R2 Random I/Os Throughput (Native Mode)

4000
3500
3000
2500
PhysicalDisk
2000
1500 FixedVHD
1000 DynamicVHD
IO PS

500 DifferencingVHD
0

(Queue Depth: 64)

29
 WS08 R2 Real Apps Workload Throughput (Native Mode)
18000
16000
14000
12000
PhysicalDisk
10000
8000 FixedVHD
6000 DynamicVHD
4000 DifferencingVHD
IOPS

2000
0

(Queue Depth:64)

30
 6. How to choose your Hyper-V and VHD Storage Container Format
Customers always have to make a choice when they need to decide what the appropriate storage
container format is for deploying virtual machines using Hyper-V. The following summary table is
intended to make the decision-making process easier:

Storage Container Pros Cons

Pass-through Disk  Fastest performance  VM snapshot cannot be taken

 Simplest storage path  Disk is being used exclusively and
because file system on directly by a single virtual
host is not involved. machine.
 Better alignment under  Pass-through disks cannot be
SAN. backed up by the Hyper-V VSS
 For shared storage based writer and any backup program
pass-through, no need to that uses the Hyper-V VSS writer.
mount the file system on
host and that may speed
up VM live migration.
 Lower CPU utilization
 Support very large disks
Fixed sized VHD  Highest performance of  Up front space allocation may
all VHD types. increase the storage cost when
 Simplest VHD file format large of number fixed VHD are
to give the best I/O deployed.
alignment.  Large fixed VHD Creation is time-
 More robust than consuming.
dynamic or differencing  Shrinking the virtual capacity (i.e.
VHD due to the lack of reducing the virtual size) is not
block allocation tables possible.
(i.e. redirection layer).
 File-based storage
container has more
management advantages
than pass-through disk.
 Expanding is available to
increase the capacity of
VHD.
 No risk of underlying
volume running out of
space during VM
operations
Dynamically expanding  Good performance  Interleaving of meta-data and
or  Quicker to create than data blocks may cause I/O
fixed sized VHD alignment issues.
 Grow dynamically to save  Write performance may suffer

31
Differencing VHD disk space and provide during VHD expanding.
efficient storage usage.  Dynamically expanding and
 Smaller VHD file size differencing VHDs cannot exceed
makes it more nimble in 2040GB
terms of transporting  May get VM paused or VHD
across the network. yanked out if disk space is
 Blocks of full zeros will running out due to the dynamic
not get allocated and growth.
thus save the space  Shrinking the virtual capacity is
under certain not supported.
circumstances.  Expanding is not available for
 Compact operation is differencing VHDs due to the
available to reduce the inherent size limitation of parent
actual physical file size. disk.
 Defrag is not recommended due
to inherent re-directional layer.

Note: there are types of compaction for dynamic or differencing VHDs. One is done when file system is
presented (i.e. you mount the VHD and bring the disk online and the other is not (VHD is offline). We
recommend mounting a VHD first before doing compaction to improve the efficiency. This is also the
default behavior when a compaction is initiated in Hyper-V Manager. However, DiskPart allows you to
do an offline compaction without prompting you to mount it first. When that is the case, only data
blocks full of zero will get released since no file system is involved. This may limit the effectiveness of the
compact operation. Note that online compaction will fail with a file system limitation error if any volume
snapshots exist.

32
 7. Summary of supported and practical limits
Virtual Hard Disks have a lot of flexibility in formats, block size and usage. The following table describes some
architectural limits around VHD in native use and in use by virtual machines followed by their definitions. One
thing worth pointing out here is Windows Server 2008 R2 provides two different user interfaces to deal with VHDs:
Hyper-V Manager and virtual disk service (VDS) based tools DiskMgmt/DiskPart. Customers may note inconsistent
behavior, for example, VHD compaction as discussed previously the maximum size of fixed size VHDs.

Windows Server 2008 Windows Server 2008 R2

Virtualization Feature
Hyper-V(RTM)

Nested Depth (VHD in VHD) 0 2

Default VHD Block Size 512KB 2MB

Minimum VHD Size 3MB 3MB

Limited by the number of disks

Number of Attached VHDs Limited by the virtual bus (64)
Windows allows

Max Differencing Chain Effective = 8 Effective =1024

Length Supported = 50 Supported = 1024

Max Fixed Sized VHD Size 2048GB(2040G via Hyper-V Manager) 16TB(2040G via Hyper-V Manager)

Max Dynamic VHD Size 2040GB 2040GB

Max Differencing VHD Size 2040GB 2040GB

 Nested Depth – refers to the number of VHDs contained in VHDs that can be attached. For
example nested depth 2 means a VHD is contained in a VHD that is attached.
 Default VHD Block size – This is the default size of data blocks in Dynamic and Differencing VHDs.
In R2 the size was increased to improve overall throughput.
 Number of Attached VHDs – This is the number of VHDs that can be attached to the system at
any one time. Keep in mind that after all drive letters are used you will have to use volume
naming or junction points to references drives.
 Max Differencing VHD chain length – This defines how many virtual hard disks can be associated
in a hierarchical chain of ‘descendents’ – starting with the original parent disk and ending with a
child disk that has no children below it.

33
  Max fixed sized/dynamically expanding/differencing VHD Size – This is the maximum size in
bytes that the format can present as a single disk.

In addition to the limits above there are some limits specific to VHD use in virtual machines that impacts
overall performance. The following table lists those limits followed by their description.

Windows Server
Virtualization Feature Windows Server 2008 R2
2008

Largest Virtual SCSI IO Size 128KB 8MBytes

Largest Virtual IDE IO Size 128KB 128KB

 Largest Virtual SCSI IO Size – This is the maximum size of and IO passed from a VM to the
management partition for processing over the Virtual SCSI path. Larger I/O size may improve
CPU efficiency by reducing I/O split.
 Largest Virtual IDE IO Size – This is the maximum size of IO passed from a VM to the
management partition for processing over the Virtual IDE path.

If in addition to the limits above you can also see a feature comparison (including storage) between
Windows Server 2008 and Windows Server 2008 R2 at:

http://blogs.msdn.com/tvoellm/archive/2009/08/05/what-s-new-in-windows-server-2008-r2-hyper-v-
performance-and-scale.aspx

34
 8. Closing
Virtual Hard Disk technology has continued to improve release-over–release, both in terms of the
scenarios supported (like boot from VHD) as well as overall performance. This paper details the
significant improvements in dynamically expanding and differencing VHD random performance, the
substantially faster fixed sized VHD creation speed, and almost zero percent drop in throughput to large
differencing VHD chains due to better caching.

When choosing the right VHD for your environment you should consider both the access performance
and storage needs. With the improvements demonstrated in Windows Server 2008 R2 the choice has
less to do with the access speed and more to do with the amount of memory used due to advanced
caching.

35
 9. Acknowledgements
We want to thank the following people and the entire VHD team for the contribution and hard work
that were incorporated into the VHD performance improvement in Windows Server 2008 R2:

Dustin Green, Todd Harris, Mike Ebersol, Karthik Thirumalai, John Howard, Chris Eck, Taylor Brown

10. References

 Loopback mounting of VHDs -

http://blogs.msdn.com/virtual_pc_guy/archive/2008/02/01/mounting-a-virtual-hard-disk-
with-hyper-v.aspx
 Inbox APIs (virtdisk.dll) - http://msdn.microsoft.com/en-us/magazine/dd569754.aspx
 VHD Specification - http://technet.microsoft.com/en-us/virtualserver/bb676673.aspx
 VHD FAQ - http://technet.microsoft.com/en-us/library/dd440865(WS.10).aspx
 What’s new in Hyper-V in Windows Server 2008 R2 -
http://www.microsoft.com/windowsserver2008/en/us/hyperv-r2.aspx

36