DELL SCALABLE ENTERPRISE TECHNOLOGY CENTER SERIES
Performance Scaling
with Dell PowerEdge 2950 Servers and VMware Virtual Infrastructure 3
To assess virtualization scalability and performance, Dell engineers tested two Dell™
PowerEdge™ 2950 servers with dual-core Intel® Xeon® 5160 and quad-core Intel Xeon
X5355 processors alongside a previous-generation PowerEdge 2850 server with dualcore Intel Xeon processors in a virtualized environment based on VMware® Virtual
Infrastructure 3. The results show that the PowerEdge 2950 servers can provide
enhanced performance and host more virtual machines than the PowerEdge 2850.
BY TODD MUIRHEAD
Related Categories:
Dell PowerEdge servers
W
hen deciding how to implement a virtualized envi-
performance increases possible when migrating to a new
ronment with industry-standard components such
enterprise IT environment.
as Dell PowerEdge servers and VMware Virtual InfrastrucDell Scalable Enterprise
Technology Center
Intel
ture 3 software, two key factors are how many virtual
Hardware configuration for test environment
machines (VMs) a single server can run and what level of
The PowerEdge 2950 is a dual-socket server that sup-
performance these VMs can provide. Although the answer
ports Intel Xeon 5000, 5100, and 5300 series processors.
Scalable enterprise
depends on many different factors, testing with common
The Dell test team configured one PowerEdge 2950 with
Virtualization
applications and workloads can provide a starting point
two dual-core Intel Xeon 5160 processors at 3.0 GHz, and
for enterprises planning such an environment.
another with two quad-core Intel Xeon X5355 processors at
VMware
Visit www.dell.com/powersolutions
for the complete category index.
In September and October 2006, Dell engineers tested
2.66 GHz. The Intel Xeon 5160 has a 4 MB processor cache
virtualized environments based on VMware Virtual Infra-
that is shared between the two cores and a 1,333 MHz
structure 3 on two differently configured PowerEdge 2950
frontside bus. The Intel Xeon X5355 is essentially two dual-
servers and one previous-generation PowerEdge 2850
core Intel Xeon 5160 processors combined, and so has an
server. Using three different workloads, the test team
8 MB level 2 (L2) cache (4 MB shared by each set of two
assessed the number of VMs that each server could host
cores) and a 1,333 MHz frontside bus.
and the overall system performance, and found that the
An additional benefit of Intel Xeon 5160 processors is
PowerEdge 2950 servers offered significant gains in both
lower power consumption compared to previous-generation
regards compared to the PowerEdge 2850. These tests
Intel Xeon processors. A previous study found that ninth-
are intended to provide an example of the scalability and
generation PowerEdge servers with Intel Xeon 5100 series
www.dell.com/powersolutions
Preprinted from Dell Power Solutions, February 2007. Copyright © 2006 Dell Inc. All rights reserved.
DELL POWER SOLUTIONS
1
DELL SCALABLE ENTERPRISE TECHNOLOGY CENTER SERIES
PowerEdge 2850
server
PowerEdge 2950 servers
Two dual-core Intel Xeon
processors DP at 2.8 GHz
with two 2 MB L2 caches
(one per core)
Two dual-core Intel Xeon
5160 processors at
3.0 GHz with 4 MB
cache (shared)
Frontside bus
800 MHz
1,333 MHz
Memory
8 GB (400 MHz DDR2
DIMMs)
16 GB (667 MHz fully buffered DIMMs)
Internal disks
Two Ultra320 SCSI
73 GB, 10,000 rpm drives
Two Serial Attached SCSI (SAS) 146 GB,
15,000 rpm drives
Network interface
cards (NICs)
Two 10/100/1,000 Mbps
internal NICs and one
Intel PRO/1000 XT
Gigabit* Ethernet NIC
Two 10/100/1,000 Mbps internal NICs
PowerEdge Expandable
RAID Controller (PERC)
4/ei
PERC 5/i
Two QLogic 2340 PCI
Extended HBAs
One QLogic 2462 PCI Express HBA
VMware ESX Server 2.5.2
VMware ESX Server 3
processors use up to 25 percent less power than
eighth-generation PowerEdge servers with dual-
Processors
Disk controller
Two quad-core Intel
Xeon X5355 processors at 2.66 GHz with
8 MB cache (shared)
core Intel Xeon processors.1 The combination
of lower power consumption and higher performance over previous-generation processors can
result in increased performance per watt.
PowerEdge 2950 servers support up to
32 GB of RAM when using 4 GB dual in-line
memory modules (DIMMs), but in the Dell tests
both servers were configured with 16 GB of RAM
using 2 GB, 667 MHz fully buffered DIMMs.
The PowerEdge 2850 server was configured
with two dual-core Intel Xeon processors at
2.8 GHz—the fastest supported processors for
this server—with a dedicated 2 MB cache for
each core, 800 MHz frontside bus, and 8 GB of
RAM using 400 MHz double data rate 2 (DDR2)
Fibre Channel HBA
Virtualization
software
DIMMs. Figure 1 summarizes the server configuration in the test environment.
The PowerEdge 2950 servers were connected to a storage area network (SAN) with
* This term does not connote an actual operating speed of 1 Gbps. For high-speed transmission, connection to a Gigabit Ethernet server and
network infrastructure is required.
a dual-port QLogic 2462 PCI Express host
bus adapter (HBA) and utilized storage on a
Figure 1. Configurations for the PowerEdge 2850 server and two PowerEdge 2950 servers used in the test
environment
Dell/EMC CX3-80 array with twenty 146 GB,
15,000 rpm disks. The PowerEdge 2850 was
connected to the SAN with two QLogic 2340 PCI Extended HBAs
Dell/EMC CX700
Dell/EMC CX3-80
Attached
servers
One PowerEdge 2850
server
Two PowerEdge 2950
servers
10,000 rpm disks. The three types of VMs—each running a different
Controller
cache
3,614 MB (2,818 MB write,
796 MB read)
10,384 MB (3,072 MB
write, 7,312 MB read)
2005, SUSE LAMP, and NetBench” section in this article—were
Fibre Channel
speed
Fibre Channel 2 (FC2)
Fibre Channel 4 (FC4)
divided into four 5-disk (4+1) RAID-5 logical units (LUNs). The
Disk
enclosures
Two DAE2 disk array
enclosures
Four DAE3P disk array
enclosures
storage configuration used in the test environment.
Disks
Twenty 73 GB,
10,000 rpm disks
Twenty 146 GB,
15,000 rpm disks
Virtualization platform for test environment
LUNs
Four 5-disk RAID-5 LUNs
Four 5-disk RAID-5 LUNs
The Dell tests used VMware Virtual Infrastructure 3 as the virtualiza-
Software
EMC® Navisphere®
Manager and Access
Logix™ software
EMC Navisphere Manager
and Access Logix software
as well as features such as load balancing and VMware High
and utilized storage on a Dell/EMC CX700 array with twenty 73 GB,
workload, as described in the “Test workloads: Microsoft SQL Server
spread across the 20 disks on each storage array. These disks were
three types of VMs were evenly divided across the LUNs so that a
quarter of each type were on each LUN. Figure 2 summarizes the
tion platform; this package includes ESX Server 3 and VirtualCenter 2
Availability (VMware HA). ESX Server allows multiple VMs to run
simultaneously on a single physical server. Each VM runs its own
Figure 2. Configurations for the Dell/EMC CX700 and CX3-80 storage used in the test
environment
OS, which in turn has its own set of applications and services.
Because ESX Server isolates each VM from other VMs on the
1 Based on testing performed by Dell Labs in May 2006 using the SPECjbb2005 benchmark on a PowerEdge 2950 server with two dual-core Intel Xeon 5160 processors at 3.0 GHz (Woodcrest) and then with two dual-core Intel
Xeon 5080 processors at 3.73 GHz (Dempsey); 4 GB of 667 MHz and 533 MHz fully buffered DIMM memory; two SAS 73 GB, 15,000 rpm hard disk drives; and the Microsoft Windows Server 2003 Enterprise x64 Edition OS, as
compared to a PowerEdge 2850 server with two dual-core Intel Xeon processors at 2.8 GHz (Paxville); 4 GB of 400 MHz DDR2 memory; two SCSI 36 GB, 15,000 rpm hard disk drives; and Windows Server 2003 Enterprise x64
Edition OS with Service Pack 1 (SP1). Actual performance and power consumption will vary based on configuration, usage, and manufacturing variability. For more information about ninth-generation PowerEdge server performance, see “Scaling Business Applications with New Servers and Storage,” by Mark Nickerson, Joe Pollock, Stori Waugh, and Stacy Hower, Dell Power Solutions, August 2006, www.dell.com/downloads/global/power/
ps3q06-20060283-CoverStory.pdf; and “Improved Virtualization Performance with 9th Generation Servers,” by David J. Morse, Dell Inc., August 2006, www.dell.com/downloads/global/solutions/9g%20_servers.pdf.
2
DELL POWER SOLUTIONS
Preprinted from Dell Power Solutions, February 2007. Copyright © 2006 Dell Inc. All rights reserved.
February 2007
DELL SCALABLE ENTERPRISE TECHNOLOGY CENTER SERIES
level of usage that still allows for workload spikes. The test team
calculated utilization levels by averaging the values from the esxtop
utility run on the ESX Server service console during each test, and
also monitored the utilization from the VirtualCenter console, as
shown in Figure 3.
Each workload ran simultaneously on multiple VMs under the
same load. By keeping all settings on the VM and driver systems
identical and then observing how many VMs could be run simultaneously, the test team was able to measure how many VMs each
physical server could support. Figure 4 shows the configuration for
each type of VM in the test environment.
Microsoft SQL Server 2005. On the SQL Server 2005 VMs, the
test team installed 32-bit versions of Microsoft Windows Server
2003 Release 2 (R2) Enterprise Edition and SQL Server 2005 with
Service Pack 1 (SP1).3 The SQL Server version of the Dell DVD Store
Figure 3. Processor utilization graph in VMware VirtualCenter 2 during a SUSE LAMP test
database was loaded into SQL Server 2005 using the scripts provided
with the DVD Store download to create the medium-size database.
same physical server just as physical systems are isolated from
The complete DVD Store application code, including SQL Server and
one another, administrators have flexibility in using ESX Server
LAMP versions, is freely available for public use under the GNU
to run different types of applications and operating systems at the
General Public License (GPL) at linux.dell.com/dvdstore. The DVD
same time. VirtualCenter 2 enables administrators to consolidate
Store database simulates the database back end of a simple Web-
control and configuration of ESX Server systems and VMs, which
based storefront. The database size is small (approximately 1 GB),
can improve management efficiency in large environments.
and representative of a database used for development or testing.
The PowerEdge 2950 servers used ESX Server 3, and the
To simulate a load on the VMs, the test team used the DVD
PowerEdge 2850 server used ESX Server 2.5.2; both were managed
Store driver program, which is included in the DVD Store download.
by an existing VirtualCenter 2 console. All VMs were first tested
Each SQL Server 2005 VM was driven by four threads of the driver
on the PowerEdge 2850, then migrated to each PowerEdge 2950 in
application with a 20-millisecond delay.
turn. The test team used VirtualCenter 2 to shut down the VMs,
SUSE LAMP. For the LAMP workload, the test team installed
migrate them to a PowerEdge 2950, upgrade the virtual hardware
32-bit versions of Novell SUSE Linux Enterprise Server 9, Apache 2,
and VMtools, and then restart the VMs.
and MySQL 5 on a VM. The MySQL version of the DVD Store
application was loaded into MySQL 5, and the PHP version of the
Test workloads: Microsoft SQL Server 2005, SUSE LAMP,
and NetBench
To compare the relative performance of the PowerEdge 2850 and
DVD Store application was set up on Apache. In this setup, the
Workload
Memory
Disk
Virtual NIC
type
Number
of virtual
processors
Microsoft SQL
Server 2005
512 MB
10 GB
Vmxnet
1
SUSE® Linux® Enterprise Server OS with a LAMP (Linux, Apache,
MySQL, PHP) stack, and the Microsoft Windows Server® 2003 OS
SUSE LAMP
1,024 MB
10 GB
Vlance/Flexible*
1
NetBench
512 MB
10 GB
Vmxnet
1
PowerEdge 2950 servers, the test team ran three workloads on each
server: the Microsoft® SQL Server™ 2005 database platform with
an online transaction processing (OLTP) workload, the Novell®
with NetBench 7.03.2 To simulate how enterprises typically run
applications on VMs using ESX Server in a production environment, the test team increased the number of VMs until proces-
* The SUSE LAMP VMs used a Vlance NIC under ESX Server 2.5.2, but this was upgraded to Flexible as part
of the migration to ESX Server 3.
sor utilization for the entire physical server exceeded 85 percent,
with the highest utilization reaching 88 percent—a reasonably high
Figure 4. Configurations for the virtual machines used in the test environment
2 The SUSE LAMP and NetBench workloads used for these tests were also used in previous studies; for more information, see “VMware ESX Server Multiple Workload Performance on Dell PowerEdge 2850 and PowerEdge 6850
Servers,” by Todd Muirhead, Dave Jaffe, and Scott Stanford, Dell Enterprise Product Group, July 2005, www.dell.com/downloads/global/solutions/vmware_mixed_workload.pdf; and “VMware ESX Server Performance Gains on
Dell PowerEdge 2850 Dual Core Servers,” by Todd Muirhead and Dave Jaffe, Dell Enterprise Product Group, November 2005, www.dell.com/downloads/global/solutions/esx_2850_dualcore.pdf. These two studies also used a
SQL Server 2000 workload that was updated to use SQL Server 2005 for the tests described in this article.
3 The use of Microsoft SQL Server 2005 in these tests does not indicate that Dell or Microsoft has tested or certified SQL Server with VMware virtualization software. As described at support.microsoft.com/?kbid=897615, Microsoft
typically does not support problems with Microsoft operating systems or applications that run on VMs using non-Microsoft virtualization software unless the same problem can be reproduced outside the VM environment.
www.dell.com/powersolutions
Preprinted from Dell Power Solutions, February 2007. Copyright © 2006 Dell Inc. All rights reserved.
DELL POWER SOLUTIONS
3
DELL SCALABLE ENTERPRISE TECHNOLOGY CENTER SERIES
PowerEdge 2850 with
dual-core Intel Xeon
processors
PowerEdge 2950 with dual-core
Intel Xeon 5160 processors
PowerEdge 2950 with quad-core
Intel Xeon X5355 processors
Workload
Number
of VMs
Performance Number
of VMs
Performance Performance
gain compared
to PowerEdge
2850
Number
of VMs
Performance
Performance
gain compared
to PowerEdge
2850
Performance
gain compared
to PowerEdge
2950 with Intel
Xeon 5160
processors
Microsoft SQL
Server 2005
14
11,782 OPM
20
20,026 OPM
70%
32
29,346 OPM
149%
47%
SUSE LAMP
16
2,650 OPM
25
6,057 OPM
129%
44
9,852 OPM
272%
63%
NetBench
18
411 MB/sec
23
593 MB/sec
44%
42
1,001 MB/sec
144%
69%
Figure 5. Workload performance results for each server in the test environment
Web tier and the database tier ran on the same VM to create a
complete LAMP stack.4
Test results measuring scalability and performance
The test team first ran the VMs on the PowerEdge 2850 server in
The driver for the LAMP stack differs from the driver used in
successive tests, adding VMs in each round as described in the
the SQL Server testing in that it sends HTTP requests and receives
“Test workloads: Microsoft SQL Server 2005, SUSE LAMP, and
HTML code returned from the Apache/PHP layer, whereas the SQL
NetBench” section in this article. Next, they migrated the VMs
Server driver communicates directly with the database. However,
to the PowerEdge 2950 server with Intel Xeon 5160 processors
the LAMP workload measures the same parameters: total orders per
and repeated the tests. Finally, they migrated the VMs to the
minute (OPM) handled by the application, and average response
PowerEdge 2950 server with Intel Xeon X5355 processors and
time experienced by the simulated customers. Each SUSE LAMP
repeated the tests once again.
VM was driven by a single thread of the driver program with a
20-millisecond delay.
The difference in the number of VMs and the associated performance metric—OPM for SQL Server 2005 and SUSE LAMP and
NetBench. NetBench 7.03, developed by PC Magazine, is a
megabytes per second for NetBench—indicated the relative differ-
benchmark tool designed to simulate a file server workload. The
ence in performance. The test team calculated the performance
program creates and accesses a set of files according to predefined
results for the SQL Server 2005 and SUSE LAMP VMs by totaling the
scripts. NetBench is typically run with an increasing number of
OPM from all the VMs running in the test environment; NetBench
client engines running against a single server to measure how much
provides the megabytes-per-second metric as part of the results
throughput (in megabytes per second) can be achieved with a given
displayed at the end of a test. Figure 5 summarizes the results for
number of connections.
the three workloads on each server.
The NetBench VMs were installed with the 32-bit version of
For all three workloads, the number of VMs that each PowerEdge
Microsoft Windows Server 2003 R2 Enterprise Edition. To determine
2950 server could host was consistently higher than the number the
how many VMs could run on an ESX Server host, the test team
PowerEdge 2850 could host, with the largest number of VMs achieved
increased the number of VMs and the number of client engines at the
by the PowerEdge 2950 server with quad-core Intel Xeon X5355
same rate until the processor utilization on the ESX Server host reached
processors running the SUSE LAMP workload. This PowerEdge 2950
85 percent. NetBench 7.03, with the included standard DiskMix script,
server, running with eight cores and 16 GB of RAM, was able to host
was used with a 0.6-second think time to connect two client engines
more than 40 VMs for two of the three workloads.
to each
VM.5
This simulates multiple file servers on the same ESX
The PowerEdge 2950 servers also showed significant gains in
Server host, similar to a file server consolidation scenario. The driver
workload performance when compared with the PowerEdge 2850.
systems on which the client engines ran had mapped drives to all
The PowerEdge 2950 server with Intel Xeon X5355 processors showed
the test VMs. In NetBench the test directories path file was modified
much larger performance gains than the PowerEdge 2950 with Intel
so that as successive client engines were added, they would use the
Xeon 5160 processors, with the largest gain—272 percent—generated
next drive letter, which corresponded to the next VM.
by the SUSE LAMP workload. Figure 6 illustrates the relative
4 The LAMP stack has been fully documented in “MySQL Network and the Dell PowerEdge 2800: Capacity Sizing and Performance Tuning Guide for Transactional Applications,” by Todd Muirhead, Dave Jaffe, and Nicolas Pujol,
Dell Enterprise Product Group, April 2005, www.dell.com/downloads/global/solutions/mysql_network_2800.pdf.
5 The NetBench client driver systems were two Dell PowerEdge 6650 servers with four Intel Xeon processors at 2.8 GHz, 8 GB of RAM, Intel Gigabit Ethernet adapters, and Windows Server 2003. The NetBench client driver
systems and ESX Server hosts were connected to a Dell PowerConnectt™ 5224 Gigabit Ethernet switch. The NetBench controller ran Windows Server 2003 Enterprise Edition.
4
DELL POWER SOLUTIONS
Preprinted from Dell Power Solutions, February 2007. Copyright © 2006 Dell Inc. All rights reserved.
February 2007
DELL SCALABLE ENTERPRISE TECHNOLOGY CENTER SERIES
PowerEdge 2950 with dual-core
Intel Xeon 5160 processors
SQL Server 2005
PowerEdge 2950 with quad-core
Intel Xeon X5355 processors
SUSE LAMP
NetBench
0
50
100
150
200
250
300
Performance increase over PowerEdge 2850 (percent)
Figure 6. Relative performance increases for the PowerEdge 2950 servers over a
PowerEdge 2850 server with dual-core Intel Xeon processors
performance increases for the PowerEdge 2950 servers over the
PowerEdge 2850 server.
Enhanced virtualization performance
Dual-core Intel Xeon 5160 and quad-core Intel Xeon X5355 processors are designed to provide a significant boost in performance over
previous-generation dual-core Intel Xeon processors. Performance
increases may vary depending on the workload and other factors,
but in the Dell tests described in this article, the PowerEdge 2950
performance gains were significant enough to allow more VMs
to be hosted, with greater overall system throughput, than was
possible on the PowerEdge 2850—including performance gains
of up to 272 percent on a PowerEdge 2950 server with quad-core
Intel Xeon X5355 processors.
The performance gains of the quad-core Intel Xeon X5355 over
the dual-core Intel Xeon 5160 demonstrate that in virtualized environments, additional cores enable the hosting of additional VMs
and help increase overall performance. These results suggest the
type of enhanced scalability and performance possible with ninthgeneration Dell PowerEdge servers and VMware virtualization software, which enterprises can take into account when planning to
implement or upgrade a virtualized environment.
Todd Muirhead is a senior engineering consultant on the Dell Scalable
Enterprise Technology Center team. Todd has a B.A. in Computer Science
from the University of North Texas and is Microsoft Certified Systems
Engineer + Internet (MCSE+I) certified.
TALK BACK
Tell us how the Dell Scalable Enterprise Technology Center
can help your organization better simplify, utilize, and scale
enterprise solutions and platforms. Send your feedback and
ideas to setc@dell.com.
www.dell.com/powersolutions
Preprinted from Dell Power Solutions, February 2007. Copyright © 2006 Dell Inc. All rights reserved.
DELL POWER SOLUTIONS
5