File Server Performance Test

1
File Server Performance Test
June 7, 2004

• Windows Server 2003 Standard vs Samba 3.0 on Red
  Hat Enterprise Server 3

2
Veritest Windows File Server Performance
Comparison
Microsoft commissioned Veritest to conduct a
series of tests comparing the File serving
performance of Windows Server 2003 Standard to
Samba 3.0 on Red Hat Enterprise Linux ES 3.0 on
two hardware configurations 1) Small Business
Configuration single processor, single disk.
Dell PowerEdge 500SC 2) Medium
Business/Departmental dual processor, disk
array. HP DL380 G3
What is the purpose of this File Server
performance test?
Build Test Around Realistic File Server Scenarios
for Small, Medium, and Departmental/Remote Office
Use
Test Objectives
Demonstrate Flexibility to Configure Windows for
Optimal Performance
Compare Samba 3.0 on Red Hat Enterprise Linux and
Windows Server 2003 for File Serving Performance
Since 1987, VeriTest has provided outsourced
testing solutions that maximize revenue and
reduce costs for our clients. VeriTest helps
our clients identify and correct technology
problems in their products and in their line of
business applications. VeriTest is a division of
Lionbridge Technologies, Inc., a provider of
globalization solutions.
Ensure Independent 3rd Party Verification of
Results (Veritest)
3
Results Single Processor, Single Disk

• Windows Server 2003 outperformed Samba 3.0 on Red
  Hat Enterprise Server 3
• With similar SMB_COM_Flush configuration as
  Samba, Windows Server 2003 outperforms by as much
  as 54
• With default configuration, Windows Server 2003
  outperforms by as much as 20

Flexibility With the software update, Windows
Server 2003 allows administrators to change the
default behavior for the SMB_COM_Flush command
forcing the command to respond immediately
without writing to disk.
Windows Server 2003 vs. Samba 3.0 File Serving
Performance Tests conducted on Dell PowerEdge
500SC
Data security Unlike Sambas default
configuration, Windows Server 2003 ensures that
all data is written to disk by issuing the
SMB_COM_FLUSH command this causes the SMB
server to write all of the data and allocation
information for a particular file (or all files
open for a particular connection) to disk and
prevents data loss.
Throughput in Mbps
Windows Server 2003 Software Update
Windows Server 2003
Red Hat Enterprise Linux 3.0 -
Samba 3.0
(Strict sync no)
of Clients
The per client load on the filer server in this
test represents more than the per client load on
the file server in a real-world situation. So 24
clients on the chart below actually represents
more than 24 clients (potentially up to 4x). In
these real-world situations, this hardware
configuration is not realistic most dedicated
file servers would have multiple disks.
4
Results Dedicated File Serving (Departmental)

• Windows Server 2003 outperformed Samba 3.0 on Red
  Hat Enterprise Server 3
• With similar SMB_COM_Flush configuration as
  Samba, Windows Server 2003 outperforms by as much
  as 54
• With default configuration, Windows Server 2003
  outperforms by as much as 40

Flexibility With the software update, Windows
Server 2003 allows administrators to change the
default behavior for the SMB_COM_Flush command
forcing the command to respond immediately
without writing to disk. This configuration is
appropriate for certain scenarios where the
performance benefits may outweigh the risk of
data loss
Windows Server 2003 vs. Samba 3.0 File Serving
Performance Conducted on HP DL380
1000
900
800
700
Data security Unlike Sambas default
configuration, Windows Server 2003 ensures that
all data is written to disk by issuing the
SMB_COM_FLUSH command this causes the SMB
server to write all of the data and allocation
information for a particular file (or all files
open for a particular connection) to disk and
prevents data loss.
Throughput in Mbps
600
500
400
300
200
Windows Server 2003 Software Update
Windows Server 2003
100
Red Hat Enterprise Linux 3.0 -
Samba 3.0
(Strict sync no)
0
1
4
8
12
16
20
24
28
32
36
40
44
48
52
56
60
of Clients
The per client load on the filer server in this
test represents more than the per client load on
the file server in a real-world situation. So 24
clients on the chart below actually represents
more than 24 clients (potentially up to 4x). In
these real-world situations, this hardware
configuration is not realistic most dedicated
file servers would have multiple disks.
5
Value Driven File Sharing
How can File Serving be improved beyond
performance?
Simple File Serving
Increase End User Value Through Value Driven File
Sharing

• Value Driven File Sharing
• Data Availability
• Service Maintenance
• Performance
• Secure Data
• End User Value

• Simple File Serving
• Store files in a single place
• Share files
• Backup Files

6
Supporting Customer Data

• Cuna Mutual Group
• Windows Server 2003 enabled seamless directory
  support for its network and remote users,
  improved storage efficiency, and remote
  administration capabilities
• "After taking a thorough look at other solutions
  available, including open source alternatives, we
  decided to go with Windows Server 2003 for our
  file and print services. --- Dan Zadra, Senior
  Systems Manager, Technical Services, CUNA Mutual
  Group
• Computer Builders Warehouse
• Migrated from Linux to Windows Server 2003
  increased performance, scalability and security
• "We now have twice the number of stores and twice
  the number of internal users that we had on the
  Linux infrastructure, yet we're supporting them
  with half the hardware we used before, thanks to
  Windows Server 2003. If we'd stayed on Linux,
  instead of halving our hardware, we'd have had to
  double it." --- Nicholas Brookins, IT Director
• Aurora First Assembly Church
• Migrating from Linux and Novell to Windows Server
  2003 operating system provided reliable, scalable
  infrastructure services that the church could
  depend on now and as demands on the system grow
• Now, instead of chasing down information or
  waiting for the network to come back up, the 40
  end users on our staff can accomplish what must
  be done and start working on special projects we
  never could get to previously. Our new network
  enables each member of our office staff to
  recover one and a half hours per day previously
  lost to delays and downtime. That's real
  productivity savings." --- Myra Beals, Office
  Manager

7
Key Sound Bites

• Windows Server 2003 delivered approximately 46
  percent better peak file server throughput
  compared to Samba 3.0 running on the
  departmental-class HP DL 380 server configuration
  we tested.
• Windows Server 2003 delivered approximately 24
  percent better peak File server throughput
  compared to Samba 3.0 running on the
  workgroup-class Dell PowerEdge 500SC server
  configuration we tested.
• Adding the software update to the Windows Server
  2003 Standard Edition generated 59 percent better
  peak File server throughput compared to Samba 3.0
  running on the workgroup-class Dell PowerEdge
  500SC server configuration we tested with strict
  syncno.
• Additionally, we noticed a discernable dip in
  the throughput performance when running Samba 3.0
  in this configuration on the HP DL380 system at
  loads of 52 clients and above. We ran this test
  multiple times with the server and client tunings
  described in this report as well as with no
  server or client tunings and always saw the drop
  off in performance towards the end of the test
  suite

8
Appendix
9
File Serving Performance Test Summary
Test Overview

• Using Ziff Davis Medias NetBench 7.02, two
  hardware scenarios were used to compare the file
  serving throughput of Windows Server 2003 vs. Red
  Hat Enterprise Linux ES with Samba 3.0
• Small Business Scenario Dell PowerEdge 500SC
  (single Intell PIII _at_ 1GHz 1 GB RAM 20 GB IDE,
  Dual NIC)
• Medium Business/Remote/Departmental Scenario HP
  DL380 (Dual P4 _at_ 2.8 GHz 4 GB RAM, SCSI disk
  array)

Results Summary

• Windows Server 2003 outperforms Red Hat 3.0/Samba
  3.0 by as much as 54
• When matching Samba 3.0 defaults assuming stable
  storage, Windows Server 2003 delivers 54 better
  peak file server throughput (See Point A)
• By default, Windows Server 2003 balances high
  performance while insuring that data is securely
  stored on disk. In this case, Windows Server 2003
  had 40 higher peak file server throughput for
  the high end scenario (Points C D) and 20
  higher throughput (Point ?) in the low-end
  scenario

Small Business Scenario Single Processor, Single
Disk
Medium Business/Departmental Scenario
Dual-Processor, Disk Array
Windows Server 2003 vs. Samba 3.0 File Serving
Performance
Windows Server 2003 vs. Samba 3.0 File Serving
Performance
400
C
350
A
300
D
B
250
200
E
Throughput in Mbps
Throughput in Mbps
150
100
Windows Server 2003 - software update
Windows Server 2003 - software update
Windows Server 2003
50
Windows Server 2003
Red Hat Enterprise Linux 3.0 -
Samba 3.0
Red Hat Enterprise Linux 3.0 -
Samba 3.0
0
1
4
8
12
16
20
24
28
32
36
40
44
48
52
56
60
Number of Clients
Number of Clients
Highlights
Highlights
Windows Server 2003 peak throughput level is 20
greater than Red Hat Linux Enterprise Server 3.0
and Samba 3.0. While no accurate estimates of
the number of real production clients are
simulated by each NetBench client, this maybe
represented as a few users to more than 10 users
for each NetBench client tested. When file server
performance is a priority, both hardware and
software are tuned to meet this priority.
Windows Server 2003 default behavior ensures that
data in memory is flushed to disk whenever an
application makes this request. On single disk
systems, the disk will ultimately become a
bottle-neck. When Windows Server 2003 is
configured like Samba 3.0s default behavior, It
outperforms Samba 3.0 by 54 and maintains peak
throughput levels longer.
Windows Server 2003 delivers greater peak
performance than Samba 3.0 and Red Hat Enterprise
Server 3.0, along with the ability to configure
the file server for greater flexibility. When
server hardware is configured for multiple disks,
Windows Server 2003 delivers better performance
than Samba 3.0 and Red Hat Enterprise Server 3.0
no matter how you balance performance with data
security. And, this performance advantage is
sustained throughout the testing.
D
A
E
B
C
Source Veritest Windows Server 2003 vs. Red Hat
Linux 3.0 with Samba 3.0
10
Background on Performance Testing
Why Performance Test?
When is Performance a Priority?

• In general, performance is a priority in Medium
  and Enterprise businesses with File Print or
  Dedicated File Servers

• Performance Test for Planning
• Optimize Hardware, Server OS, and Client OS
• Evaluate Existing Hardware
• Understand File Server Limits
• Determine RAM CPU specs
• Verify improvements in OS performance
• Determine storage specification
• Performance Test as a Competitive Differentiator
• Platform Choice
• Workload Requirements
• Understanding OS Limits

Segment
Server Role
Server Role Details
Single Server Environment Priority on
Applications Lower priority on file server
performance
Multi-workload Server
Small Business
Departmental Branch Office Business Critical
Data End-user Productivity High Availability High
Priority on Performance
File Print Server
Med Business
Business Critical Data End-user Productivity High
Availability Ease of Management High Priority on
Performance
Dedicated File Server
Enterprise Business
What is the purpose for this latest File Server
performance test?
How can File Serving be improved beyond
performance?

• Value Driven File Sharing
• Integrated Directory Services (AD)
• Transparent data access (DFS)
• End-user data recovery (Shadow Copy)
• Integrated Collaboration (SharePoint)
• Off-line File Access (ReDirected MyDocs)

Simple File Serving
Build Test Around Realistic File Server Scenarios
for Small, Medium, and Departmental/Remote Office
Use
Test Objectives
Demonstrate Flexibility to Configure Windows for
Optimal Performance
Compare Samba 3.0 and Red Hat Linux and Windows
Server 2003 for File Serving Performance
Deliver End User Value Through Value Driven File
Sharing

• Value Driven File Sharing
• Data Availability
• Service Maintenance
• Performance
• Secure Data
• End User Value

• Simple File Serving
• Store files in a single place
• Share files
• Backup Files

Ensure Independent 3rd Party Verification of
Results (Veritest)
11
Considerations Default Data Settings for Samba
vs. Linux

• Windows Server 2003 and Samba 3.0 have different
  default configurations for the way that they
  handle client requests to flush data
• Samba 3.0 tries to balance Unix and Windows
  behavior by allowing the Linux file system to
  determine when to sync data to disk. By default,
  client application requests to flush data to disk
  are ignored.
• Windows Server 2003 allows the client application
  to communicate with the operating system when it
  expects the data to be on stable storage,
  insuring that all client application requests to
  flush data to disk are honored.

Issue
Default Perspective

• Samba View By default, do not honor client
  requests to flush data
• Many Windows applications (including the Windows
  98 explorer shell) seem to confuse flushing
  buffer contents to disk with doing a sync to
  disk. Under UNIX, a sync call forces the process
  to be suspended until the kernel has ensured that
  all outstanding data in kernel disk buffers has
  been safely stored onto stable storage. This is
  very slow and should only be done rarely.
  Setting this parameter to no (the default) means
  that smbd ignores the Windows applications
  requests for a sync call. There is only a
  possibility of losing data if the operating
  system itself that Samba is running on crashes,
  so there is little danger in this default
  setting. -- From Samba Documentation smb.conf
  file
• Windows Server 2003 By default, Applications can
  be assured that data in memory has been safely
  stored on disk.
• The current behavior of the SMB_COM_FLUSH command
  causes the SMB server to write all of the data
  and allocation information for a particular file
  (or all files open for a particular connection)
  to be written to disk.

Samba 3.0 is, on the simplest level, an SMB
redirector for the integration of UNIX and Linux
with Windows for file and print services,. As
such, Samba relies on the underlying UNIX
operating system to ensure that data is actually
stored securely on disk. Differences in the
Windows and UNIX worlds result in Samba needing
to strike a balance, choosing a default behavior
that favors performance over insuring that data
in memory is actually stored to disk. Samba
cannot know if a file server or disk subsystem is
adequately protected against potentially
destructive events such as a power outage, Samba
assumes that because there is only a
possibility of losing data that this is an
acceptable risk for the user. Windows Server
2003 honors a client applications interest in
knowing that its data has been securely store on
disk and assumes that no risk of data corruption
or data loss is acceptable to the user. In the
case where the user has rationalized both the
security of their hardware and the value of their
data, Windows Sever 2003 now provides the ability
for the user to choose to have the server ignore
client requests for a flush to disk.
Discussion
Configuration Option
Description
OS
Default Behavior
Value set to non-zero value The file server is
treated as stable storage and the SMB server
will immediately respond to SMB_COM_FLUSH
commands without writing data to disk. Value set
to zero The SMB server will write data to disk
in response to SMB_COM_FLUSH commands before
responding.
New SRV.SYS And Registry Change HostTreatedAsStabl
eStorage
Honors Client intent to flush data to stable
storage
Windows
Does not honor Client requests to flush to stable
storage
Strict Sync No The SMB server will ignore disk
sync requests. Strict SyncYes The SMB will write
to disk all requests for a disk sync
Samba Linux
Strict Sync
12
(No Transcript)