Alfresco benchmark report_bl100093

white paper
Benchmark Results: Scalability of Alfresco Content
Management System in a Unisys ES7000/one
Enterprise Server Environment

2
Table of Contents
Executive Summary 4
Overview 4
Requirements Definition 4
Performance 4
Scope 4
Assumptions 4
Goals 5
Metrics 5
Baseline Establishment 5
Key Metrics 5
Other Measurements 5
Hardware Required 5
System Software Required 5
Application Software Required 5
Application Overview 6
Test Methodology 7
Methods 7
Benchmark Business Process 1 – Load 10 Million Documents 7
Benchmark Business Process 2 - Load 100 Million Documents 7
Benchmark Business Process 3 – Demonstrate Adequate Response 7
Benchmark Business Process 4 - Test Reindex Performance (optional) 8
Benchmark Business Process 5 – Keep Loading 8
Results 8
Key Metrics 8
Other Measurements 8
Discussion 9
Achievement of Goals 9
Observations 9
Performance Analysis 9
System Statistics 9
JVM 9
Conclusions 10
Next Steps 10

3
Test System Configuration 10
Hardware Required 10
Database/Application Tier 10
Driver Systems 10
Storage 10
Software Required 11
Database Tier 11
Application Tier 11
Driver Systems 11
Network Required 11
Configuration Diagram 11
Storage Configuration 12
LUN Assignments 12
Meta-LUN Assignments 12
Storage Groups 13
Server Configuration 13
ES7000 Partition Assignments 13
Commodity Server Assignment 14
IP Address Assignments 14
Appendix - Test Documents used for load and retrieval 15

4
Executive Summary
In October 2007, Unisys provided Alfresco with a Unisys
ES7000/one Enterprise Server, EMC2 CX3-80 storage
Subsystem, and related equipment to benchmark the
scalability of Alfresco’s Content Management System (CMS)
on enterprise–class equipment.
The goals of this benchmark were to load at least 100 million
documents into a single repository, achieve a load rate of at
least 100 documents loaded per second, and demonstrate
acceptable response time for single-document retrieval with
a repository this large.
All three of the goals were met. A total of 107 million
documents were loaded into the repository at a rate of 140
documents per second. After some minor fixes to the Alfresco
code, all document retrieval operations were demonstrated
to take the same amount of time as in a clean system.
During the test, the ES7000/one Enterprise Server was
underutilized – processor idle time was approximately
80 percent. Since the goals were met “out of the box” and
due to the short testing timeframe, no attempts were made
to optimize the application for this scale hardware. It is
recommended that a follow-on test be scheduled where this
can be addressed. The goal of such a test would be to fully
utilize the ES7000/one Server and maximize the documents
loaded per second.
Overview
Alfresco is the leading Open Source alternative for enterprise
content management. It couples the innovation of Open Source
with the stability of a true enterprise-class platform.
The Open Source model allows Alfresco to use best-of-breed
Open Source technologies and contributions from the
Open Source community to get higher quality software
produced more quickly at much lower cost.
Requirements Definition
• Load 100 million documents into a single Alfresco repository
• Demonstrate adequate response times for various
document-retrieval operations
• Demonstrate scaling using clustered servers (if necessary
due to ceiling on the scale-up architecture)
Performance
Demonstrating load performance was a secondary
requirement to demonstrating the ability to support large-scale
storage. However, the latter is not possible without the former.
Read performance is critical only in that it must be
“acceptable” (acceptable response times are defined in the
section Test Methodology).
Scope
This test plan addresses the validation and performance
characterization using the existing application release only.
Only the tests described in the Requirements Definition were
covered in this test. Alfresco has shown interest in loading
one billion documents into the repository, however, at the
rate of 140 documents per second that was achieved during
the test, this would take a continuous run of approximately
three months. The EPC will be happy to work with Alfresco
to achieve this goal in a future engagement. We believe that
working to increase the document processing rate would bring
the total run time down to a more reasonable number.
Assumptions
• Successful installation of RHEL 4 U4 and associated
components
• Successful installation and performance of the Alfresco
application, which is a development version and is different
from the version used in the previous benchmark

Hardware Required
• ES7000/one Server as a database/application server
• EMC2 CX3-80 storage subsystem for database, indexes,
and image files
• 2.7 GHz commodity four-processor servers for driver
systems (quantity TBD)
System Software Required
• Red Hat Enterprise Linux, RHEL 4 Update 4
• MySQL 4.1.21 Enterprise (or Oracle 10g if needed)
• JDBC driver for MySQL 4.1.21 (or Oracle 10g if needed
due to performance issues with MySQL)
• Unisys JVM 5.0.10 (or Sun JVM 6.0 if needed due to
performance issues with Java SE 5.0- based JVM)
• Firefox web browser
• FTP client
• Apache Tomcat 5.5.20
Application Software Required
• Alfresco 2.2 (development version)
Goals
1. Load as many documents as possible in the time allocated,
theoretically 100 million documents
2. Achieve a load rate of 100 documents per second
3. Document single-document retrieval time using various
levels of concurrency
Metrics
Baseline Establishment
The baseline for this test was established during a previous
benchmark in 2006 using Alfresco version 1.4. This benchmark
was conducted using a dual-processor, dual-core 2.6 MHz
Opteron-based system with 12 GB memory. The following
results were obtained:
• Documents loaded: 10 million
• Documents loaded per second – 64
• Read content – 0.34961 seconds
• Read property – 0.41976 seconds
• Create content – 0.68788 seconds (not critical now)
• Create folder – 0.54419 seconds (not critical)
In so far as this is considered our “baseline,” the goal of
the current benchmark is to determine how much the previous
results can be exceeded using Alfresco version 2.2
(still under development) running on an enterprise-class
infrastructure. The primary goal of the current test is to
exceed this previous benchmark by a factor of 10.
The need to demonstrate scaling by way of using clustered
servers was not required.
Key Metrics
• Total number of documents stored
• Documents loaded per second
• Response time for document operations
• Processor utilization
Other Measurements
• Context switches
• Memory utilization
• Disk statistics (queue length, response time)
• Size of Lucene indexes
• Size of database
5

The Alfresco Server can either run an application server such as Tomcat, or as a standalone
application. Tomcat was used in order to simplify the installation and configuration of the server.
This also allowed the repository to be accessed via the user interface, and in particular via
the Node Browser web application.
The client loader program communicates with the Alfresco server via RMI.
The document structure and metadata is stored in the database while the actual content
binaries are stored on a file system that is locally accessible to the server VM. Indexes of
the document structure and documents’ text are maintained on the file system in the form
of Lucene indexes.
Application Overview
The following diagram shows the architectural overview of the application.
6
Client
Alfesco Server
File system
Content binaries
JVM
JVM
Database
Content metadata
File system
Lucene indexes

Test Methodology
Methods
Most clients of the Alfresco repository (the Repository) use
a common API for managing documents. The API will be
exported via RMI and made available to a remote benchmark
client (the Benchmark Application, BMLoader), written
specifically for these tests.
The document structure will be as follows:
• 10 stores (partitionable units)
• Each store has 10 million documents spread across 10
thousand folders
• Trimmed-down configuration settings to remove unused
functionality
• A list of the documents used for all of the tests is
provided in the Appendix
The Benchmark Loader or BMLoader is an application written
by Alfresco for use in benchmarks. The BMLoader simulates
a production environment and input to the Alfresco repository
without the necessity of installing the external third-party
applications that interface with the Alfresco. It will be
responsible for creating the document structure. For additional
information on this process, please refer to the Alfresco
website at: http://wiki.alfresco.com/wiki/Server_Benchmarks.
Performance metrics will be recorded by both the BMLoader
and by a separate process running within the Repository itself.
Client responsiveness can be measured using an FTP client
and performing common operations such as navigation and
uploading and downloading of files. Sub-second client response
times are adequate for these tests to be considered successful.
Benchmark Business Process 1 – Load 10 Million Documents
The first task will be to load 10 million documents into each
of the logical stores. Analysis of performance and hardware
utilization during the phase will take place. Hardware
configurations and their effects will be measured in order to:
• Demonstrate scaling as concurrency is increased
• Demonstrate scaling as system resources are added
• Demonstrate scaling as the server is clustered. This only
needs to be done if further performance improvements are
required or if there is enough time available in this phase.
Records of the resource usage must be kept. In particular,
the size of the Lucene indexes and databases.
From this, the best available configuration can be set up.
The predicted outcome of the next phase will be more certain,
which will be to get up to 100 million documents.
Benchmark Business Process 2 -
Load 100 Million Documents
With the available hardware and software optimized, the next
phase will be to ramp up to the first target milestone. 100
million documents in the repository is the target number of
documents that will be considered a success. The process
will be automated, with the BMLoader and Repository working
24 hours a day until the target is reached. Intervention will only
be required in the event of unexpected problems or a
significant reduction in the document load rate.
Records of the resource usage must be kept. In particular,
the size of the Lucene indexes and databases.
Benchmark Business Process 3 –
Demonstrate Adequate Response
The BMLoader application will perform a set of read and write
operations against the Repository. Since the BMLoader will
be using a fully functional API shared by the other Alfresco
clients, response times can be considered similar to those
that will be achieved by the clients themselves.
Tests using FTP and the Web Client will also be done.
Although results may be measured, they need only satisfy
the general user adequacy requirement.
7

Benchmark Business Process 4 -
Test Reindex Performance (optional)
This optional phase would be useful, but should only be
done if:
• the Lucene indexes can be safely moved away and
brought back later (backed up and restored)
• there is time available
The aim of this phase is to
• measure the re-index speed, which will be a measure of
the overall read performance
• see if the size of the Lucene indexes differs significantly
from the incrementally built indexes
Benchmark Business Process 5 – Keep Loading
This is just like phase 2, but continues until the last
possible moment. If further hardware is required and
available, this can be added in. Enough time must just be
available to repeat phase 3.
Results
Key Metrics
• Total number of documents stored
- 107 million
• Documents loaded per second
- 140 documents per second
• Response time for document operations
- Acceptable (see Discussion)
• Processor utilization
- average of approximately 20 percent on the application
server and 15 percent on the database server
Other Measurements
• Context switches
- Not recorded
• Memory utilization
- Application server – 30 GB (JVM)
• Disk statistics (queue length, response time)
- The only disk with a high queue depth was the
content store
• Size of Lucene indexes
- 4 GB each for each of the 10 stores
• Size of database
- ~800 GB
8

6. A full reindex was kicked off to see how long it would
take. The single process took approximately 24 hours to
reindex 10 percent of the complete set. Alfresco is
addressing this.
7. Full scalability for this test was achieved on a single
application server, therefore a clustered environment was
not necessary.
Performance Analysis
System Statistics
Processor utilization was low (~20 percent) on both the
database (four CPU) and application (eight CPU) servers.
The JVM was allowed to use all of the available memory but
probably did not need all of that to produce these levels of
performance. The only disk that showed significant queuing
was sdl on the application server which contained the
content store.
JVM
Testing was begun with the Sun JVM, v1.6 (aka 6.0). During
the test, the Unisys JVM was swapped for the Sun JVM and
a decline in performance was noted. It was determined that
this was due to the Unisys JVM being based on the
Sun JVM v1.5 (Java SE 5.0). The Alfresco team had noticed
a significant improvement in going from Sun v1.5 to Sun v1.6.
Unisys has a development version based on v1.6 which was
tried later in the test. It appeared to provide roughly the same
level of performance as was obtained with the Sun JVM.
9
Discussion
Achievement of Goals
All goals of this test plan were met. 107 million documents
were loaded (target was 100 million). The average load rate
was 140 documents per second, 40 percent higher than the
target of 100 documents per second and more than double
what was achieved during previous tests. The response times
for the read and write operations against the fully loaded store
were less than one second which is more than adequate.
Observations
1. A problem was detected during server startup. As part of
the consistency checks, a “select max(id)” query was
performed which was observed to take up to 10 minutes
to execute. The check was switched off and Alfresco has
corrected the issue.
2. The UI response times for most operations were not
noticeably different from the clean system. Browsing was
fast and creation and modification of folders showed no
sign of being impaired by the data volume or even by
having a benchmark run concurrently. Content upload
speeds were good, but slightly affected by running a
concurrent benchmark.
3. Deletion speeds were affected by the data volume.
A particular query that had to get executed N+1 times
(where N is the depth of the hierarchy being deleted)
was taking minutes to execute. This rendered delete
unusable. Alfresco has fixed the query.
4. The Alfresco Node Browser was used to perform
searches against both the standard Spaces Store and
the benchmark partitioned stores. The observed results
were normal. In other words, the usual large resultset
limitations that are currently present in the system were
not exacerbated by the data volumes. The Lucene indexes
showed no signs of degradation at volume.
5. The FTP client displayed the same good response as the
Web Client, with the exception of deletes as
mentioned above.

Test System Configuration
Hardware Required
Database/Application Tier
ES7000/one Server with eight cells, each cell containing:
• Four processors, Intel® Dual-Core Xeon® 7140M,
3.33 GHz with 16 MB L3-cache
• 32 GB memory
• Two LP11002-M4 4 Gb dual-port fibre-channel HBAs
• Two Intel® PRO/1000 MT dual-port gigabit NICs
• Two 146 GB internal disks, SCSI 320, mirrored
Driver Systems
Unisys ES3040L Servers, four required, each containing:
• Four Intel® Xeon® MP processors, 2.7GHz with 2MB cache
• Eight GB memory
• Two Broadcom NetExtreme gigabit NICs
• Two 36 GB internal disks, mirrored
Storage
One EMC2 CX3-80 Clariion Storage subsystem
100M documents, 50K average size, 5000 GB for
content storage
• 16 GB shared cache
• 8 4-Gb fibre-channel host ports
• 120 73 GB, 15K RPM, 4-Gb fibre-channel disks
• 60 146 GB, 15K RPM, 4-Gb fibre-channel disks
10
Conclusions
The test goals presented in this test plan were all met with
very little tuning. The goals were aimed at determining
whether the Alfresco system could provide acceptable
performance with a large content store (greater than 100
million documents) rather than optimizing either the
document load speed or retrieval times. The ES7000/one
Server and related test components were vastly underutilized
(with the exception of the content store disk). Because of the
test goals, the limited time available during this test, and
the fact that acceptable performance was achieved
essentially “out of the box,” minimal optimizations were
made and minimal performance statistics were captured.
The Unisys EPC team recommended that a future engagement
be scheduled which will focus on tuning the document load
process. One of the desired goals of Alfresco is to test with
one billion documents in the content store. This would only
be feasible with a significant increase to the document load
speed. At the observed rate of 140 documents per second,
the system would need to run continuously for 83 days in
order to achieve one billion documents.
Next Steps
First and foremost, all of the fixes discovered during this test
will be incorporated into the 2.2 release of the Alfresco
Repository. They will also be included in the community code
base as well.
Second, Alfresco and Unisys should seriously consider a
second set of tests to be run sometime in early 2008 to
address actual performance tuning on the ES7000/one Server.
From the metrics observed during this test, it is quite feasible
that the rate of document insertion into the repository can
be significantly increased. Along with faster processing times,
the increase in load speed would make it feasible to aim for
a higher total number of documents inserted into the
repository, making Alfresco and Unisys the leader in
document management.

Software Required
Database Tier
• Red Hat Enterprise Linux AS 4 Update 4 for AMD64/Intel64; or
• MySQL 4.1.21 Enterprise (or Oracle 10G,Enterprise Edition)
Application Tier
• Red Hat Enterprise Linux AS 4 Update 4 for AMD64/Intel64
• JDBC driver for MySQL 4.1.21 or Oracle 10g
• Unisys JVM 5.0.10 x64 (or Sun JVM 6.0 x64)
• Alfresco 2.2 (development version) (supplied by client)
Driver Systems
• Red Hat Enterprise Linux AS 4 Update 4
• Firefox
• FTP client
• Unisys JVM 5.0.10 i386
Network Required
• Public LAN for communications between servers and remote administration
Configuration Diagram
11
Alfresco Content Management System - System Landscape
Cisco 4506
Gigabit Switch
Unisys ES3040L Servers
EMC2 Clariion
CX3-80 Storage
Database Content Store
Lucene Incexes
Unisys ES7000/one
Enterprise Server
Database Server
Application Server
2

12
Storage Configuration
LUN Assignments
Meta-LUN Assignments
Array Name LUN ID RAID Type
Number
of Drives
Physical
Drive Size (GB)
LUN Size (GB)
Storage
Processor
Meta-LUN
Component
1 LUN 0 6138 5 8 133 936 A Yes
1 LUN 1 6143 5 8 133 936 A Yes
1 LUN 2 6142 5 8 133 936 A Yes
1 LUN 3 6141 5 8 133 936 A Yes
1 LUN 4 6140 5 8 133 936 A Yes
1 LUN 5 6139 5 8 133 936 A Yes
1 LUN 1 1 1+0 10 67 333 B No
1 LUN 2 2 1+0 10 67 333 A No
1 LUN 3 3 1+0 10 67 333 B No
1 LUN 4 4 1+0 10 67 333 A No
1 LUN 5 5 1+0 10 67 333 B No
1 LUN 6 6 1+0 10 67 333 A No
1 LUN 7 7 1+0 10 67 333 B No
1 LUN 8 8 1+0 10 67 333 A No
1 LUN 9 9 1+0 10 67 333 B No
1 LUN 10 10 1+0 10 67 333 A No
1 LUN 11 11 1+0 10 67 333 B No
Array Name LUN ID Type LUNs (ID) Meta-LUN Size (GB)
Storage
Processor
1 LUN 0 0 Stripe 6138, 6143, 6142, 6141, 6140, 6139 5.5TB A

13
Storage Groups
Array Storage Group LUN (ID) Host Host Drive Name/Function
1 Alfresco DB 1 usmv-epce2p0 sdb MySQL Database
1 Alfresco App 0 usmv-epce2p4 sdb Lucene Index
1 Alfresco App 2 usmv-epce2p4 Sdc Lucene Index
1 Alfresco App 3 usmv-epce2p4 Sdd Lucene Index
1 Alfresco App 4 usmv-epce2p4 Sde Lucene Index
1 Alfresco App 5 usmv-epce2p4 Sdf Lucene Index
1 Alfresco App 6 usmv-epce2p4 Sdg Lucene Index
1 Alfresco App 7 usmv-epce2p4 Sdh Lucene Index
1 Alfresco App 8 usmv-epce2p4 Sdi Lucene Index
1 Alfresco App 9 usmv-epce2p4 Sdj Lucene Index
1 Alfresco App 10 usmv-epce2p4 sdk Lucene Index
1 Alfresco App 11 usmv-epce2p4 sdl Content Store
Server Configuration
ES7000 Partition Assignments
Name CPUs CPU Speed Cache (MB) Memory (GB) NICs FC Adapters Function
usmv-epce2p0 4 3.4 16 32 4 4 Database server
usmv-epce2p4 4 3.4 16 32 4 4 Application server

14
Commodity Server Assignments
IP Address Assignments
Name CPUs CPU Speed Cache (MB) Memory (GB) NICs
Storage Size
(GB)
Storage Type Function
usmv-epcc2-1 4 2.7 2 8 2 36 RAID Client
usmv-epcc2-2 4 2.7 2 8 2 36 RAID Not used
Name
IP Address
(Primary)
IP Address
(Secondary)
Subnet Mask DNS Server (Primary)
DNS Server
(Secondary)
usmv-epce2p0 172.26.102.0 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epce2p4 172.26.102.4 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-1 172.26.142.1 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-2 172.26.142.2 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-3 172.26.142.3 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-4 172.26.142.4 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-5 172.26.142.5 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-6 172.26.142.6 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-7 172.26.142.7 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-8 172.26.142.8 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-9 172.26.142.9 N/A 255.255.0.0 172.26.252.1 172.26.252.2
usmv-epcc2-10 172.26.142.10 N/A 255.255.0.0 172.26.252.1 172.26.252.2

15
Appendix – Test Documents used for load and retrieval
File Name/Type Size (Bytes)
1768.jpg 25,817
1784.jpg 23,926
2634.jpg 25,638
2897.jpg 27,608
3517.jpg 26,232
3531.jpg 26,236
5099.jpg 25,902
5102.jpg 27,562
512.jpg 24,371
5132.jpg 24,178
5135.jpg 27,321
6016.jpg 24,044
653.jpg 25,231
656.jpg 26,736
659.jpg 25,532
6979.jpg 27,011
6985.jpg 23,467
8227.jpg 24,367
8230.jpg 24,999
8329.jpg 26,990
8350.jpg 27,511
alfresco.log 3,238
Break Even Analysis.xls 36,352
Cashflow Forecast.xls 23,552
crud.txt 7,583
Employee Interview Questionnaire.doc 44,544
Marketing Activity Analysis Sheet.doc 29,696
ml-permissions.txt 271
New Client Needs Analysis Form.doc 62,976
Project Briefing Paper For Contractors.doc 49,664
QA component list Final.ods 53,957
Site Development Process.doc 52,224
sitepoint_logo.jpg 69,448
srs_template.doc 55,296
Update.pdf 49,733
User-Authentication-HOWTO.pdf 51,350
Valgrind-HOWTO.pdf 56,827
VCR-HOWTO.pdf 58,144
VMailMgr-HOWTO.pdf 44,336
Web Development Seminar.doc 43,520
Web development tips- Ten (or a few more) files every Web site needs.pdf 83,446
Windows-Newsreaders-under-Linux-HOWTO.pdf 50,427
WordPerfect.pdf 51,693
XFree86-R200.pdf 48,470

For more information, contact your Unisys representative.
In a hurry to learn more, Visit:
http://www.unisys.com
Specifications are subject to change without notice.
© Unisys Corporation. All rights reserved.
Unisys is a registered trademark of Unisys Corporation. Intel and Xeon are registered trademarks of Intel Corporation. Microsoft and
Windows are registered trademarks of Microsoft Corporation. Red Hat is a registered trademark of Red Hat, Inc. Linux is a registered
trademark of Linus Torvalds. All other brands and products referenced herein are acknowledged to be trademarks or registered
trademarks of their respective holders.
Printed in U S America 12/07
*BL100093-200*
BL100093-200

Alfresco benchmark report_bl100093

More Related Content

What's hot (20)

Viewers also liked (14)

Similar to Alfresco benchmark report_bl100093 (20)

Alfresco benchmark report_bl100093