White Paper - Technical & Business Validation · Edison: IBM . Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 2 Clients using POWER7 systems and PowerVM
Post on 28-May-2020
7 Views
Preview:
Transcript
89 Fifth Avenue, 7th Floor
New York, NY 10003
www.TheEdison.com
212.367.7400
White Paper
IBM PowerVM Virtualization
Technology on IBM POWER7 Systems
A Comparison of PowerVM and
VMware vSphere 4.1 update 1
Virtualization Performance
Printed in the United States of America
Copyright 2011 Edison Group, Inc. New York. Edison Group offers no warranty either
expressed or implied on the information contained herein and shall be held harmless for errors
resulting from its use.
All products are trademarks of their respective owners.
First Publication: September 2011
Produced by: Craig Norris, Sr Analyst; Barry Cohen, Editor-in-Chief; Manny Frishburn, Editor
Table of Contents
Executive Summary ..................................................................................................................... 1
Introduction .................................................................................................................................. 3
Objective .................................................................................................................................. 3
Audience .................................................................................................................................. 3
Contents of this Report .......................................................................................................... 3
The Business Value of Virtualization ...................................................................................... 4
Benchmark Comparison Study: PowerVM vs. VMware ...................................................... 6
AIM7 Benchmark .................................................................................................................... 6
TPoX Benchmark .................................................................................................................. 10
Summary of Results ............................................................................................................. 21
IBM’s Virtualization Infrastructure: POWER7 Systems .................................................... 22
PowerVM ............................................................................................................................... 23
Processor Virtualization ...................................................................................................... 23
Memory Virtualization ........................................................................................................ 24
I/O Virtualization .................................................................................................................. 24
Partition Mobility ................................................................................................................. 24
Partition Hibernation ........................................................................................................... 24
Workload Partitioning ......................................................................................................... 25
Systems Management .......................................................................................................... 25
PowerVM Advantages ......................................................................................................... 26
Conclusions ................................................................................................................................. 29
Appendices .................................................................................................................................. 32
Appendix 1 — Benchmark Configuration Information ........................................................ 32
Appendix 2 — General Benchmark Descriptions .................................................................. 34
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 1
Executive Summary
Today’s business organizations need to rein in IT costs, but without sacrificing
performance, security, reliability, and flexibility. A new era has emerged in which it is
now possible, through intelligent and strategic use of new and/or advanced technology,
to achieve breakthrough economics, considerably reducing the cost of delivering the
workloads central to a business’s operation.
IBM has been aggressively making strides in pioneering IT infrastructure designed to
harness trends and innovation to deliver top-notch functionality with the greatest
efficiency and with the greatest data center savings. The vendor’s initiative — defined
by IBM as Smarter Computing — has helped many forward-thinking organizations to
design, tune, and manage their IT infrastructures to make them designed for data, tuned
to the task, and managed in the cloud.
A cornerstone of this initiative is a move toward architectures optimized for specific
purposes and built around deep domain knowledge. The goals here comprise reducing
deployment times for systems from months to days, improving performance with
utilization rates of up to 90 percent, and reducing floor space, power consumption,
labor, and total cost per workload. The key operative technology advancement
harnessed in achieving these goals is server consolidation through virtualization.
Using virtualization technology to consolidate data center servers has become an
integral component of how successful companies approach the design of their IT
systems. However, the majority of businesses fall far short of realizing the full potential
of server consolidation. On average, consolidation ratios are only around six virtual
machines (VMs) per physical server. Even world-class organizations are only
consolidating at a ratio of about 18 to 1 at best. Much higher VM densities are possible
without degrading system performance, reducing data center consolidation expenses
significantly and yielding a considerable economic advantage to organizations.
Under the banner of “Power is performance redefined,” IBM has introduced an
impressive profile of servers with the 2010 launch and 2011 release of Power Systems
servers and blades. These products are based on the IBM POWER7 processor
architecture, ranging from 2-socket to 32-socket with up to 256 cores.
To evaluate the value IBM’s virtualization technology can offer clients, IBM engaged
Edison Group to help provide a clear understanding of the benefits that can be seen
when organizations implement virtualization technology as part of their IT
environment. IBM virtualization technologies support a server virtualization ratio of
1,000 to 1, outdoing competitors and providing for massive data center consolidation.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 2
Clients using POWER7 systems and PowerVM virtualization technology achieve higher
operational savings by using greater VM density. Many of the advantages stem from the
fact that PowerVM technology is built directly into the firmware of all Power Systems
servers. The widely-deployed VMware vSphere 4.1 update 1 and other x86-based
virtualization products are typically third-party software add-ons, sold and installed
separately.
This technical white paper presents benchmark results showing greater VM
consolidation ratios than demonstrated in previous benchmarks and demonstrating the
extent of the performance lead that PowerVM virtualization technologies deliver over
x86-based add-on virtualization products. The tests, running two workload benchmarks
of different consolidation ratios on POWER7 processor-based and comparable Intel-
based systems, demonstrate the exceptional performance and scalability of PowerVM
virtualization technologies compared to VMware vSphere 4.1 update 1 on an x86-based
platform. Key findings include the following:
PowerVM technology on an IBM Power 750 system performs up to 115 percent
better than VMware vSphere 4.1 update 1 in whole core configuration with a
consolidation ratio of 32 to 1.
PowerVM technology on an IBM Power 750 system performs up to 200 percent
better than VMware vSphere 4.1 update 1 in a processor resource over-commit
configuration (five virtual machines share a core) with a consolidation ratio of 40 to
1.
PowerVM technology on a 4-socket IBM Power 750 system demonstrated linear
scaling with 50 percent more absolute throughput performance compared to
VMware vSphere 4.1 update 1 on a 4-socket Intel Xeon 7560 system.
The benchmark results clearly reveal that PowerVM virtualization technology on
POWER7 processor-based platforms offers greater performance than that offered by
VMware vSphere 4.1 update 1 on Intel x86 platforms. They enable high consolidation
ratios, broader scalability, and increased flexibility for a far superior virtualization
solution. PowerVM virtualization technology on POWER7 processor-based platforms
not only uses system resources in shared processor mode more efficiently, but also
delivers superior performance when resources are over-committed with a higher
consolidation ratio. Together they establish PowerVM virtualization technology as the
consolidation system of choice for organizations wishing to realize the full advantages of
greater VM density.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 3
Introduction
Objective
The objective of this white paper is to compare the performance of PowerVM
virtualization technologies on POWER7 processor-based server platforms against
VMware vSphere 4.1 update 11 on comparable Intel x86 platforms. It describes tests
using industry-standard benchmarks to compare virtualization technologies. The results
were reviewed, analyzed, and presented by Edison Group.
Audience
This paper is intended for anyone interested in the advantages of server consolidation
through virtualization. IT managers, CIOs, system architects, and others will find
valuable information that will help them further enhance and adopt virtualization
technology within their IT environments.
Contents of this Report
This white paper contains the following major sections:
The Business Value of Virtualization — This section discusses the business value
propositions underlying the benchmark evaluations presented in this paper.
Benchmark Comparison Study: PowerVM Virtualization Technology versus
VMware vSphere 4.1 update 1 — This section presents the comparative testing
itself, describing the test bed setup, the benchmarks, the actual tests, and the results
of the tests.
IBM’s Virtualization Infrastructure: POWER7 Processor-Based Systems — This
section describes the Power Systems virtualization infrastructure, its components,
and its advantages.
Appendices — The appendices contain configuration information and general
descriptions for the benchmarks used in the tests discussed in this paper.
1 The benchmarks and analysis were done with an n-1 Intel Xeon-based server and VMware vSphere 4.1
rather than the most recent versions of Intel Xeon servers and VMware vSphere because , the Intel Xeon
processor used in the HP DL 580 G7 server for these benchmarks is the most widely-used x86 processor in
the field, OEM vendors are continuing to ship servers based on this technology, the VMware vSphere
version benchmarked is the most widely-used x86 version in the field., and the test runs of benchmarks
typically take several months to plan, approve, configure, execute, analyze, and report.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 4
The Business Value of Virtualization
Inefficiencies have cropped up in data center operations as applications, workloads, and
data have multiplied. These include: underutilization of server processor capacity,
memory bottlenecks that restrict performance, server sprawl and its related difficulties
in deployment and management, as well as higher energy bills from excessive power
demands. Such inefficiencies increase costs, both through expenditures for equipment
purchases and licensing, as well as through greater demands on administrative staff
resources, etc.
Virtualization technologies allow IT organizations to consolidate workloads running on
multiple operating systems and software stacks, and to dynamically allocate platform
resources to meet specific business and application requirements. Server virtualization,
the foundation platform for today’s data center, is quickly reaching maturity. More than
half of business server workloads are now deployed on virtual machines. According to
IDC, 2 virtualization has become the default build for new server installations, driving
down costs and establishing the foundation for more efficient and flexible configurations
and technology platforms. The average size of virtualized workloads increased threefold
between 2006 and 2009. The performance of virtualization is a critical factor to realize
success of server pools and cloud computing (and is also a key component in IBM’s
roadmap in its Smarter Computing initiative).
Well-implemented virtualization solutions may be employed to:
Reduce hardware expenditures by consolidating multiple environments, including
underutilized servers, and systems with varied and dynamic resource requirements.
Reduce costs for power and cooling, floor space, hardware maintenance, and
software licensing.
Grow and shrink resources dynamically according to business needs.
Deploy new workloads through provisioning VMs or new systems rapidly to meet
changing business demands.
Develop and test applications in secure, independent domains while allocating
production to its own domain on the same system.
Transfer live workloads to support server migrations, balance system load, or avoid
planned downtime that can otherwise adversely impact productivity.
Control server sprawl, reducing system management costs.
2 The Value of Memory-Dense Servers: IBM’s System x MAX5 for its eX5 Server Family, March 2010, IDC
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 5
Despite this, the majority of businesses fall far short of seizing upon the full potential of
server consolidation. Their average consolidation ratio hovers around six VMs per
server,3 yet economic advantages from data center consolidation increase significantly at
much higher VM densities. By increasing the consolidation ratio per system, businesses
can reduce capital expenditures and operational costs by reducing the number of
systems in their data center or IT organization.
IBM’s Smarter Computing systems, which allow for greater VM density without
degrading system performance, can deliver considerable economic advantages to
organizations using them. This study examines the performance and scaling aspects of
PowerVM and VMware vSphere 4.1 update 1 virtualization at high consolidation ratios
(32:1 and 40:1) across two different commonly employed industry benchmarks (AIM7
and TpoX). The case of 40:1 consolidation ratio — “five virtual machines per core”—
was mapped to achieve a higher amount of compression than the client deployment
consolidation ratio surveyed in 2010.4
3 According to a recent Aberdeen Group report, Best-in-Class Practices for Virtualizing Microsoft Applications,
August 2010, even the best-in-class organizations in the study consolidate at only an 18:1 ratio. 4 http://www.networkworld.com/news/2010/121510-vmware-server.html
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 6
Benchmark Comparison Study:
PowerVM vs. VMware vSphere 4.1 update 1
AIM7 Benchmark
AIM7 is a well-known open source benchmark. It is widely used by UNIX computer
system vendors to compare system performance. It comprises three pre-defined tests
suites (compute, multi-user, and database). Each suite is a mix of compute-, memory-
and I/O-intensive atomic tests covering a wide range of operations. AIM7 also stresses
the guest operating system’s kernel performance within virtualized environments. The
testing described in this paper used the compute server test suite.
Methodology
For AIM7 scaling tests, all 32 available cores were used to scale from one to 32 virtual
machines on both platforms. The Power Linux version used on PowerVM virtualization
technology was SuSE 11 SP1, while SuSE 11 SP1 x86_64 version was used as guest OS on
VMware vSphere 4.1 update 1. (Configuration details of the tests are in the appendices)
Results
AIM7 was scaled in one, two, four, eight, 16, and 32 virtual machines (each virtual
machine having one virtual processor). Scaling was close to linear on both the POWER7
processor-/PowerVM technology-based systems and the Intel/VMware vSphere 4.1
update 1 platforms. The tests were run at close to 100 percent utilization to measure the
absolute performance of AIM7 in each VM configuration.
POWER7 processor-/PowerVM technology-based systems demonstrated more than two
times (110 percent) better performance than Intel/VMware vSphere 4.1 update 1 at one,
two, four, eight, and 16 VM configurations, while at 32 VM, PowerVM technology
demonstrated a 115 percent advantage (Figure 1).
NOTE: The VM configuration and the test results can be found in the tables
following the graphs for each test in this paper, starting with Figure 1.
Table 1 shows the details on throughput and CPU utilization for each configuration. In
this test, the VMs on both platforms were configured as close to identically as possible.
In the case of PowerVM, each logical partitioning (LPAR) was given one core
entitlement, one vCPU (virtual CPU), and 3 GB RAM; in the case of VMware vSphere 4.1
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 7
update 1, each VM was given one vCPU and 3 GB RAM, with the remainder left at
default options.
Figure 1. AIM7 Benchmark Virtual Machine Scaling Performance
System Configuration for AIM7
Benchmark (1 to 32 VM Scaling)
# of
VMs
Total
Virtual
CPUs
% CPU
Utilization Jobs / min
IBM Power 750 3.5 GHz DPSM mode, 4
sockets, 512 GB RAM, SMT4 enabled,
PowerVM and SLES11 SP1 (Power Linux)
1
1
94.6
19048.5
2 2 94.3 38120.5
4 4 97.8 76189.5
8 8 94.6 152249.8
16 16 98 303983.8
32 32 96.9 603085.1
HP ProLiant DL580 G7, 2.26 GHz, 8 cores /
24 MB cache (4 sockets Intel Xeon 7560
Processors, 512 GB system RAM, (HT and
Turbo enabled in BIOS Intel VTx with EPT
HW virtualization assist) VMware
vSphere 4.1, SLES11 SP1 (GA x86_64)
1
1
100
9068.6
2 2 99.89 18137.2
4 4 94.15 36180.1
8 8 100 72398.3
16 16 92.5 144365.4
32 32 95.2 280726.8
Table 1. AIM7 Benchmark Multiple Virtual Machine Scaling Results
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 8
PowerVM and VMware vSphere 4.1 update 1 technologies differ in the way they map a
physical processor to a virtual processor. PowerVM virtualization technology maps all
four threads of a core (SMT4, introduced with POWER7 processor-based systems) to a
virtual processor. So, PowerVM technology leveraged POWER7 SMT4 technology with
one vCPU configuration. VMware vSphere 4.1 update 1 maps one of the two threads of
a core (Intel’s HT technology) to a virtual processor. Therefore, VMware vSphere 4.1
update 1 was not able to leverage Intel’s HT technology with one vCPU configuration
per VM.
The tests on VMware vSphere 4.1 update 1 were thus repeated with two vCPU per VM
configuration in order to observe performance with two threads running on a core.
Because the VMware vSphere 4.1 update 1 VM was reconfigured to have two virtual
processors, the test team wished to ensure that each VM was assigned a core to match
with PowerVM technology. So, CPU affinity was used to assign two threads (the
primary and secondary thread of a core) to two virtual processors of each VM.5 This set
of tests was a fair comparison with PowerVM test results, since it allowed the workload
to consume all the capacity of the system in a manner similar to POWER/PowerVM
technology. The results of the second test are shown below (Figure 2).
The second test results with two vCPU reveal that results for the Intel Xeon processor
running VMware vSphere 4.1 update 1 had improved, but still lagged behind
POWER/PowerVM results. PowerVM technology still demonstrated up to 59 percent
higher throughput performance than Intel 7560 /VMware vSphere 4.1 update 1 at close
to 100 percent utilization in each of the tests.
Power 750/PowerVM technology demonstrated higher AIM7 throughput performance
than the HP system with Intel 7560 processor using VMware technologies. Many factors
contributed to this superior performance, including: PowerVM technology efficiency,
IBM POWER7 SMT4 technology, and IBM POWER7 processor core frequency
(specifically, the fact that IBM POWER7 technology supports higher frequency with the
same processor capacity than does Intel Xeon technology).
5 That is, one vCPU of a VM was assigned to an even number logical processor, and a second vCPU of a VM
was assigned to an odd number logical processor. For example, the first vCPU of the first VM was assigned
to logical cpu0, and the second vCPU of the first VM was assigned to logical cpu1, so that all the primary
and secondary threads of cores were consumed by the workload running on that VM.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 9
Figure 2. AIM7 Benchmark Multiple Virtual Machine Scaling with two vCPU for VMware
vSphere 4.1 update 1
System Configuration for AIM7
Benchmark (1 to 32 VM Scaling)
# of
VMs
Total
Virtual
CPUs
% CPU
Utilization
Jobs /
min
IBM Power 750 3.5 GHz DPSM mode, 4
sockets, 512 GB RAM, SMT4 enabled,
PowerVM and SLES11SP1 (Power Linux)
1 1 94.6 19048.5
2 2 94.3 38120.5
4 4 97.8 76189.5
8 8 94.6 152249.8
16 16 98 303983.8
32 32 96.9 603085.1
HP ProLiant DL580 G7, 2.26 GHz, 8 cores /
24 MB cache (4 sockets) Intel Xeon 7560
Processors, 512 GB system RAM, (HT and
Turbo enabled in BIOS Intel VTx with EPT
HW virtualization assist),VMware
vSphere 4.1, SLES11 SP1 (GA x86_64)
1 2 95.19 12274.8
2 4 99.67 24351.7
4 8 95.75 48671.4
8 16 95.32 97531.6
16 32 99.8 190598.1
32 64 92.09 379976.1
Table 2. AIM7 Benchmark Multiple Virtual Machine Scaling with two vCPU for VMware vSphere
4.1 update 1
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 10
TPoX Benchmark
TPoX (Transaction Processing over XML) is an application-level “XML database”
benchmark based on a financial application scenario. It simulates an actual application
that performs queries, inserts, updates, and deletes in a concurrent multi-user workload.
It is an XML OLTP benchmark using data-oriented XML structures, very large numbers
of relatively small XML documents (1 kb to 20 kb), short read/write transactions, and a
high degree of concurrency. It models a security-trading scenario that uses a real-world
XML Schema (FIXML). TPoX is an open-source benchmark developed by IBM in
collaboration with Intel and others. It is available at:
http://tpox.sourceforge.net/tpoxresults.htm 6
A database application, TPoX stresses CPU, memory, and storage I/O; however, in a
multi-VM environment, this benchmark also stresses the virtualization infrastructure
supporting these resources on both platforms.
Methodology
The next set of tests were conducted using the TPoX benchmark. These tests involve a
higher degree of processor contention, using a VM-to-core ratio of 5 to 1. Because of this
increased ratio, the shared pool configuration was reduced in these tests to eight cores
on both platforms in order to limit the maximum VMs to 40 on each platform.
The TPoX benchmark is I/O-intensive and its performance has a dependency on storage
performance. Identical storage subsystems were used on both of the VM platforms. A
logical array (12 spindles) with RAID5 was used to host four VMs on each in order to
avoid I/O blender 7. Both the data and logs for the database are configured on the same
set of disks in order to simplify the configuration for hosting 40 VMs.
Each VM used a 1 GB database in order to match up with each VM’s CPU (0.2 core) and
memory capacity (3 GB). A single-tier TPoX configuration was chosen for each VM
where the client and the database reside in the same VM.
The VM configuration has multiple options on both PowerVM and VMware vSphere 4.1
update 1 technologies. 8
6 Reference: http://nativexmldatabase.com/2011/03/04/new-tpox-benchmark-results-available/ 7 http://www.networkworld.com/news/2010/102510-burning-questions-virtualization-storage.html 8 On PowerVM, each VM was configured with 0.2 core/one vCPU/uncapped mode/3 GB RAM with shared
processor pool allocated with one, two, four and eight cores (up to one socket) for five-VM, 10-VM, 20-
VM, and 40-VM, respectively. There were three dedicated LPARs configured to consume the other three
sockets on Power 750 system. On VMware, two sets of configurations were used; the first set includes a
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 11
Results
The database for each VM on each of the platforms was populated with the same
configuration set. The transaction rate for populating the database is shown in Table 3,
below.
Power 750/PowerVM
HP DL580G7/ VMware
vSphere 4.1 update 1
1vCPU
Order (inserts per second) 1,591 746
Custacc (inserts per second) 684 271
Table 3. TPoX Database Populated Rate for First Configuration Set
As these results indicate, the performance rate for populating the database is two to two-
and-a-half times better for POWER/PowerVM technology than with Intel Xeon 7560
/VMware vSphere 4.1 update 1. Figure 3, below, presents results demonstrating that the
transactions throughput performance on POWER/PowerVM technology is as much as
three times better than Intel Xeon 7560 /VMware vSphere 4.1 update 1. 9
configuration where each VM was given one vCPU/20 percent of a core — 452 MHz limit/4 GB RAM, and
advanced shared panel settings that included 1) hyperthread core sharing and, 2) Scheduling affinity set to
0-15 (logical processors). The idea was to run five VMs on a single core; with five vCPUs the entire core
should be utilized in hyperthreading mode. 9 Because processor utilization in the first VMware configuration set made it harder to report total
percentage, in this case pool utilization was used. With a single virtual processor per VM, it would not be
realistic to map to either a primary or secondary thread per VM. For example, in the five-VM test, where the
goal was to use 20 percent of a core, binding a VM could be done either to a primary or to a secondary
thread, in which case some VMs would be running on primary and others would running on secondary.
Thus, the decision was made to use the pool to assign cpu0 to cpu15 for all the tests. At a lower number of
VMs, VMware used around 20 percent from each of the cores in the pool; at 40 VMs the pool utilization
matched with PowerVM as it is shown in Figure 3.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 12
Figure 3. TPoX Benchmark Results in 40:1 Consolidation Ratio
Table 4, on the following page, presents detailed information on the total number of
TPoX users used in each test, pool utilization, throughput, and VM configuration for
each tests.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 13
System Configuration for TPoX
Benchmark (1 to 40 VM Scaling)
# of
VMs
Total
Virtual
CPUs
Total # of
TPoX
Users
% Pool
Utilization
Transactions
per second
IBM Power 750 3.5 GHz DPSM mode, 4
sockets, 512 GB RAM, SMT4 enabled,
PowerVM, AIX 7.1 is the host OS for
each VM. VIOS is configured with 0.2
core/1 vCPU/ uncapped mode/ 4 GB
RAM. Each LPAR is configured with 1
vCPU/ uncapped/3 GB RAM 3 LPARs
have 0.2 cores and 2 LPARs have 0.1
core Shared pool has 1 core
5 5 50 12.5 612.2
8 LPARs are configured each with 0.2/1
vCPU/ uncapped/ 3 GB memory, 2
LPARs are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios
has 0.2/1 vCPU/uncapped/4 GB
memory. Shared pool has 2 cores
10 10 100 24.5 1155
18 LPARs are configured each with
0.2/1 vCPU/ uncapped/ 3 GB memory,
2 LPARs are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios
has 0.2/1 vCPU/uncapped/4 GB
memory. Shared pool has 4 cores
20 20 200 49 2137
38 LPARs configured with 0.2core/
1vCPU/uncapped and 2 LPARs
configured with 0.1core/1vCPU/
uncapped. Shared pool has 8 cores
40 40 400 98 4169.8
HP ProLiant DL580 G7, 2.26 GHz, 8
cores / 24 MB cache (4 sockets) Intel
Xeon 7560 Processors, 512 GB system
RAM (HT and Turbo enabled in BIOS
Intel VTx with EPT HW virtualization
assist) VMware vSphere 4.1 update1.
Each VM has guest OS RHEL6 GA.
Each VM is given 0.2 of a core/1
vCPU/3 GB memory. DB2 buffer pool
for data is configured in each VM.
Schedule affinity is set to cpu0 and
cpu1.
5 5 50 21.8 203.18
Schedule affinity is set to cpu0 to cpu3 10 10 100 33.89 397.15
Schedule affinity is set to cpu0 to cpu7 20 20 200 56.62 760.52
Schedule affinity is set to cpu0 to cpu15 40 40 400 97.26 1571.27
Table 4. TPoX Benchmark Results
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 14
Figures 4A and 4B depict the response time for each transaction type — query, update,
delete, and insert — for each test on both platforms.
Figure 4A. TPoX Query and Update Response Time
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 15
Figure 4B. TPoX Delete and Insert Response Time
As shown in Figure 4A and 4B, the response time on VMware vSphere 4.1 update 1 was
two to six times higher compared to PowerVM virtualization technology as the number
of VMs scaled from five to 40 VMs. The pool utilization was higher as well, while
throughput was lower on VMware vSphere 4.1 update 1 than on PowerVM technology.
Even though hyper threading (HT) technology was leveraged in this test, VMware
vSphere 4.1 update 1 performance remained one-third of that demonstrated by
PowerVM technology.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 16
# of
VMs
Total
Virtual
CPUs
Avg
query
rsp
(sec)
Avg
update
rsp
(sec)
Avg
delete
rsp
(sec)
Avg
insert
rsp
(sec)
IBM Power 750 3.5 GHz DPSM mode, 4
sockets, 512 GB RAM, SMT4 enabled,
PowerVM, AIX 7.1 is the host OS for
each VM. VIOS is configured with 0.2
core/1 vCPU/ uncapped mode/ 4 GB
RAM. Each LPAR is configured with 1
vCPU/ uncapped/3 GB RAM 3 LPARs
have 0.2 cores and 2 LPARs have 0.1
core. Shared pool has 1 core
5
5
0.09
0.08
0.03
0.05
8 LPARs are configured each with 0.2/1
vCPU/ uncapped/ 3 GB memory, 2
LPARs are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios has
0.2/1 vCPU/uncapped/4 GB memory.
Shared pool has 2 cores
10 10 0.09 0.08 0.04 0.06
18 LPARs are configured each with 0.2/1
vCPU/ uncapped/ 3 GB memory, 2
LPARs are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios has
0.2/1 vCPU/uncapped/4 GB memory.
Shared pool has 4 cores
20 20 0.1 0.1 0.06 0.07
38 LPARs configured with 0.2core/ 1
vCPU/uncapped and 2 LPARs
configured with 0.1core/1 vCPU/
uncapped. Shared pool has 8 cores
40 40 0.1 0.15 0.1 0.12
HP ProLiant DL580 G7, 2.26 GHz, 8
cores / 24 MB cache (4 sockets) Intel
Xeon 7560 Processors, 512 GB system
RAM (HT and Turbo enabled in BIOS
Intel VTx with EPT HW virtualization
assist) VMware vSphere 4.1 update1.
Each VM has guest OS RHEL6 GA. Each
VM is given 0.2 of a core/1 vCPU/3 GB
memory. DB2 buffer pool for data is
configured in each VM. Schedule
affinity is set to cpu0 and cpu1.
5
5
0.22
0.33
0.205
0.26
Schedule affinity is set to cpu0 to cpu3 10 10 0.22 0.376 0.26 0.3
Schedule affinity is set to cpu0 to cpu7 20 20 0.22 0.39 0.275 0.315
Schedule affinity is set to cpu0 to cpu15 40 40 0.24 0.348 0.18 0.265
Table 5. TPoX Response Time for Each Transaction Type
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 17
Power 750/PowerVM
HP DL580G7/ VMware
vSphere 2vCPU
Order (inserts per second) 1591 1176
Custacc (inserts per second) 684 333
Table 5A. TPoX Database Populated Rate for Second Set (2 vCPU) of Configuration
A second configuration set on VMware vSphere 4.1 update 1 was added in order to
restrict the VMs to run within the core similar to the PowerVM virtualization
configuration.10 Again, the database of each VM in this new configuration on VMware
vSphere 4.1 update 1 was populated. The transaction rate for populating the database is
shown in the table A. The results of this set of tests were compared with results for
PowerVM technology, as shown in Figures 5, A and B.
POWER/PowerVM still retained 2.3 times better performance than HP Intel/VMware
vSphere 4.1 update 1 technologies, even with reconfiguration using CPU affinity
(VMware Scheduling Affinity group) on VMware vSphere 4.1 update 1. CPU utilization
on both platforms remained close to identical.
Figure 5. TPoX Performances with VMware vSphere 4.1 update 1 Virtual Machine
Reconfiguration
10 In this set each virtual machine was configured to have two vCPUs, using the CPU affinity feature in
VMware; the first vCPU was bound to the primary thread while the second vCPU of a VM was bound to
secondary thread of a core. For the five-VM test, all 10 vCPUs were bound to one core (both primary and the
secondary thread); for the 10-VM test, all 20 vCPUs were bound to two cores; for the 20-VM test, all 40
vCPUs were bound to four cores; and for 40-VM test, all 80 vCPUs were bound to eight cores. In each case
each VM was able to leverage both primary and secondary threads concurrently.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 18
System Configuration for TPoX
Benchmark (5 to 40 VM Scaling)
# of
VMs
Total
Virtual
CPUs
% CPU
Utilization
Transactions
per second
IBM Power 750 3.5 GHz DPSM mode, 4
sockets, 512 GB RAM, SMT4 enabled,
PowerVM, AIX 7.1 is the host OS for each
VM. VIOS is configured with 0.2 core/1
vCPU/ uncapped mode/ 4 GB RAM. Each
LPAR is configured with 1 vCPU/
uncapped/3 GB RAM 3 LPARs have 0.2
cores and 2 LPARs have 0.1 core. Shared
pool has 1 core
5
5
100
612.2
8 LPARs are configured each with 0.2/1
vCPU/ uncapped/ 3 GB memory, 2 LPARs
are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios has
0.2/1 vCPU/uncapped/4 GB memory.
Shared pool has 2 cores
10 10 98 1155
18 LPARs are configured each with 0.2/1
vCPU/ uncapped/ 3 GB memory, 2 LPARs
are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios has
0.2/1 vCPU/uncapped/4 GB memory.
Shared pool has 4 cores
20 20 98 2137
38 LPARs configured with 0.2core/ 1
vCPU/uncapped and 2 LPARs configured
with 0.1 core/1 vCPU/ uncapped. Shared
pool has 8 cores
40 40 98 4169.8
HP ProLiant DL580 G7, 2.26 GHz, 8 cores /
24 MB cache (4 sockets) Intel Xeon 7560
processors, 512 GB system RAM (HT and
Turbo enabled in BIOS Intel VTx with EPT
HW virtualization assist) VMware
vSphere 4.1 update1. Each VM has guest
OS RHEL6 GA. Each VM is given 2 vCPUs
unlimited/3 GB memory. DB2 buffer pool
for data is configured in each VM.
Schedule affinity is set to cpu0 and cpu1.
5
10
100
259
Schedule affinity is set to cpu0 to cpu3 10 20 100 490.5
Schedule affinity is set to cpu0 to cpu7 20 40 100 997.9
Schedule affinity is set to cpu0 to cpu15 40 80 100 1906.4
Table 6 TPoX Performance with VMware vSphere 4.1 update 1 Virtual Machine Reconfiguration
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 19
Consider how the response time improved with CPU scheduling affinity on VMware
vSphere 4.1 update 1. The query response time was reduced by approximately 30
percent. However, the impact on other transactions’ response time was negligible.
The question arose as to how these results would compare to previously published
TPoX benchmark results. No published results using virtualization technologies existed,
so Edison Group compared these results with those of published results for testing non-
virtualized systems on a comparable Intel Xeon 7560 system.11
Figure 6A. TPoX Query and Update Response Time
Figure 6B. TPoX Delete and Insert Response Time
11 In March 2010, Intel had published TPoX benchmark results on an Intel Xeon 7560 system with 32
cores/256 GB RAM using a 1 TB database in a non-virtualized environment. Further results can be found at
http://tpox.sourceforge.net/TPoX_Results_X7560.pdf
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 20
System Configuration for TPoX
Benchmark (5 to 40 VM scaling)
# of
VMs
Total
Virtual
CPUs
Avg
query
rsp
(sec)
Avg
update
rsp
(sec)
Avg
delete
rsp
(sec)
Avg
insert
rsp
(sec)
IBM Power 750 3.5 GHz DPSM mode, 4
sockets, 512 GB RAM, SMT4 enabled,
PowerVM, AIX 7.1 is the host OS for
each VM. VIOS is configured with 0.2
core/1 vCPU/ uncapped mode/ 4 GB
RAM. Each LPAR is configured with 1
vCPU/ uncapped/3 GB RAM 3 LPARs
have 0.2 cores and 2 LPARs have 0.1
core Shared pool has 1 core
5
5
0.09
0.08
0.03
0.05
8 LPARs are configured each with 0.2/1
vCPU/ uncapped/ 3 GB memory, 2
LPARs are configured with 0.1/ 1
vCPU/uncapped/3 GB memory, vios
has 0.2/1 vCPU/uncapped/4 GB
memory. Shared pool has 2 cores
10 10 0.09 0.08 0.04 0.06
18 LPARs are configured each with
0.2/1 vCPU/ uncapped/ 3 GB memory, 2
LPARs are configured with 0.1/1
vCPU/uncapped/3 GB memory, vios
has 0.2/1 vCPU/uncapped/4 GB
memory. Shared pool has 4 cores
20 20 0.1 0.1 0.06 0.07
38 LPARs are configured with 0.2core/ 1
vCPU/uncapped and 2 LPARs
configured with 0.1core/1 vCPU/
uncapped. Shared pool has 8 cores
40 40 0.1 0.15 0.1 0.12
HP ProLiant DL580 G7, 2.26 GHz, 8
cores / 24 MB cache (4 sockets) Intel
Xeon 7560 Processors, 512 GB system
RAM (HT and Turbo enabled in BIOS
Intel VTx with EPT HW virtualization
assist) VMware vSphere 4.1 update1.
Each VM has guest OS RHEL6 GA.
Each VM is given 2 vCPU unlimited/3
GB memory. DB2 buffer pool for data is
configured in each VM. Schedule
affinity is set to cpu0 and cpu1.
5
10
0.145
0.3
0.245
0.26
Schedule affinity is set to cpu0 to cpu3 10 20 0.17 0.36 0.285 0.33
Schedule affinity is set to cpu0 to cpu7 20 40 0.16 0.39 0.25 0.275
Schedule affinity is set to cpu0 to cpu15 40 80 0.17 0.345 0.25 0.28
Table 7 TPoX Response Time with VMware vSphere 4.1 update 1 Reconfiguration
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 21
These previously-published results were better than what was achieved in tests using
the HP Intel Xeon 7560 system described here. The difference in these results could be
attributed to differences in storage subsystem, database size, execution of a large
number of software images such as guest OS, database middleware, etc. Most
significant, however, is that the tests described here were conducted in a virtualized
environment using VMware vSphere 4.1 update 1, which adds overhead in comparison
to a non-virtualized environment.
Summary of Results
Overall, PowerVM virtualization technology demonstrated superior performance over
VMware vSphere 4.1 update 1 in two different configurations, each configuration
covering two different virtual machine densities featuring high resource contention. As
demonstrated using the AIM7 and TPoX benchmarks, the difference in throughput
performance was quite considerable throughout, ranging from 50 percent better to as
much as 200 percent better on PowerVM technology.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 22
IBM’s Virtualization Infrastructure:
POWER7 Processor-Based Systems
The currently available POWER7 processor-based systems combine excellent
performance, scalability, and modularity. IBM’s clients realize a high return on their
investments with flexible, responsive infrastructures that easily adapt and grow based
on business needs. A virtualization hypervisor is built into Power Systems to provide
superior performance over competitive systems which rely on third-party virtualization
software such as the widely-deployed VMware vSphere 4.1 update 1.
POWER7 processor-based systems offer balanced systems designs that automatically
optimize workload performance and capacity at either a system or a virtual machine
level. Features include:
TurboCore workload-optimizing mode for maximum per-core performance for
databases.
MaxCore for parallelization and maximum capacity throughput.
Intelligent threading technology to utilize more threads when it benefits workloads.
Intelligent Cache technology to optimize cache utilization, flowing from core to core.
Intelligent Energy that maximizes performance dynamically when thermal
conditions allow.
Active Memory Expansion 12 that dynamically provides more memory on an as-
needed basis.
Active Memory Sharing that allows for logical over-commitment of physical
memory and de-duplication.
IBM PowerVM technology — the virtualization software built into the POWER7
processor-based systems — offers an unprecedented level of platform support,
scalability, efficient resource utilization, flexibility, and heterogeneous server
management. IBM PowerVM virtualization offers autonomic resource affinity, resulting
in higher workload performance in a virtualized environment. IBM POWER7 Systems,
and PowerVM technology with its efficient virtualization, are an excellent foundation
for cloud computing environments.
12 Supported on AIX operating systems only.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 23
PowerVM Virtualization Technology
With IBM POWER processor-based systems and IBM PowerVM virtualization
technologies, an organization can consolidate applications and servers using
partitioning and virtualized system resources to achieve a more flexible and dynamic IT
infrastructure. PowerVM delivers robust virtualization for IBM i, IBM i, and Linux
environments on IBM POWER processor-based systems. The POWER Hypervisor is
integrated as part of the system firmware and supports multiple operating
environments on a single system. PowerVM virtualization technology offers the
flexibility of combining dedicated and shared resources in the same partition. IBM
Power Systems servers and PowerVM technology are designed to deliver a dynamic
infrastructure that can help reduce costs, manage risk, and improve service levels.
Processor Virtualization
PowerVM technology’s advanced dynamic logical partitioning (LPAR) capabilities
allow a single partition to act as a completely separate AIX, IBM i, or Linux operating
environment. Partitions can be assigned either dedicated or shared processor resources.
With shared resources, PowerVM virtualization technology can automatically adjust
pooled processor resources across multiple operating systems, borrowing processing
power from idle partitions to handle high transaction volumes in other partitions.
PowerVM technology’s Micro-Partitioning supports up to 10 dynamic logical partitions
per processor core. Depending upon the Power server, up to 1,000 independent
virtualized servers can be run on a single physical Power server — each virtualized
server with its own fractional processor share, memory, and I/O resources. These
partitions can be assigned at a granularity of 1/100th of a core. Consolidating systems
with PowerVM technology can reduce operational costs, improve availability, ease
management, and improve service levels, while allowing businesses to deploy
applications quickly.
Shared processor pools increase throughput by allowing for the automatic non-
disruptive balancing of processing power between partitions assigned to shared pools. It
also provides for the ability to reduce processor-based software licensing costs by
capping the processor core resources used by a group of partitions.
Shared dedicated capacity allows for the “donation” of spare CPU cycles from dedicated
processor partitions to a shared processor pool. The dedicated partition maintains
absolute priority for dedicated CPU cycles. Enabling this feature can help to increase
system utilization without compromising the computing power for critical workloads in
a dedicated processor.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 24
Memory Virtualization
PowerVM technology features Active Memory Sharing, a technology that intelligently
and dynamically reallocates memory from one partition to another for increased
utilization, flexibility, and performance. Active Memory Sharing enables the sharing of a
pool of physical memory among logical partitions on a single server. This helps reduce
the need for reserve memory resource capacity in a consolidated environment by
increasing the efficiency of memory utilization, driving down system costs. The memory
is dynamically allocated among the partitions as needed, to optimize the usage of
physical memory in the pool. Along with shared memory, PowerVM technology also
supports dedicated memory allocation, which enables partitions having shared memory
to coexist in the same system as partitions having dedicated memory.
I/O Virtualization
The Virtual I/O Server (VIOS) is an integral part of PowerVM technology. A special-
purpose partition, VIOS eliminates the need for dedicated network adapters, disk
adapters and disk drives, and tape adapters and tape drives in the guest partitions
running as VMs. It can reduce costs by virtualizing I/O resources to those partitions.
VIOS owns the resources that are shared with clients; a physical adapter assigned to the
VIOS partition can be shared by one or more other partitions. With VIOS, guest
partitions can easily be created for test, development, or production purposes. PowerVM
technology also supports dedicated I/O along with VIOS on the same system. Therefore,
a single system can have I/O hosted by VIOS for some partitions and other partitions
with dedicated I/O devices. An organization can thus reserve a dedicated VM of a given
capacity that can be relied upon for high-priority and/or mission-critical workloads,
while assigning other VMs to a general resource pool.
Partition Mobility
Live Partition Mobility facilitates the migration of a running AIX or Linux partition from
one physical server to another without requiring application downtime for planned
system maintenance, migrations, provisioning, and workload management.
Partition Hibernation
IBM POWER7 systems support Partition Hibernation, where a partition can be
suspended and resumed at a later time. In a suspended state, a partition’s resources can
be used by other partitions while the suspended partition’s state is stored in a paging
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 25
space on a persistent storage device. Partition Hibernation can be used for resource
balancing and for planned CEC outages for maintenance or upgrades.
Workload Partitioning
PowerVM technology also supports a software partitioning technology provided by the
AIX operating system, a mode of virtualization capability called Workload Partitions
(WPARs). Introduced with AIX Version 6, WPAR is independent of hardware features.
It enables consolidation of workloads on a single AIX operating system by providing
isolation between workloads running in different WPARs. From an application
perspective, each workload is running in its own operating system environment. A key
feature of WPAR is mobility; a running WPAR can be relocated from one VM to another
on the same operating system platform. This enables applications to be migrated to
another system during planned maintenance operations, to balance workloads, to
provision rapidly to meet growth dynamically, and to improve energy efficiency by
further consolidating on the fly during low load periods.
Systems Management
IBM Systems Director (Express, Standard, and Enterprise Editions) for Power servers
supports the PowerVM environment. It is IBM’s tool for heterogeneous platform
management of Power Systems, IBM System x, IBM System z, and IBM System Storage
systems. IBM Systems Director Editions support advanced management functions such
as system discovery, workload lifecycle management, health monitoring, system
updates, and topology mappings. It also provides the ability to take action on defined
event thresholds of monitored system components.
IBM Systems Director VMControl transforms Systems Director from managing
virtualization to using virtualization in order to better manage an entire IT
infrastructure. It is offered as a plug-in option included with the Systems Director
Standard and Enterprise Editions. Together, IBM Systems Director and VMControl help
reduce the total cost of ownership in a virtual environment by increasing asset
utilization and reducing the time and effort required to deploy workloads. Using them,
administrators can maintain high levels of availability through proactive monitoring
and collaborative troubleshooting, reducing costs further.
VMControl is available in three editions, to suit the varying levels of virtualization
deployment at client sites:
VMControl Express Edition provides basic VM lifecycle management.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 26
VMControl Standard Edition adds virtual appliance lifecycle management to capture
information from active systems and store it in a repository as reusable system
images (called virtual appliances).
VMControl Enterprise Edition adds system pool lifecycle management. It allows
users to create and manage system pools – or groups of virtual appliances deployed
across multiple physical servers – as easily as managing a single entity. The
advanced virtualization management capabilities of VMControl provide a pathway
for organizations to build sophisticated cloud computing environments.
PowerVM Virtualization Technology Advantages
PowerVM virtualization technology offers a secure virtualization environment built on
the advanced RAS features and excellent performance of the Power Systems platform.
PowerVM technology delivers numerous advantages, including:
High resource utilization — PowerVM technology makes the most efficient
utilization of IT investments by virtualizing resources that include processors,
memory, and I/O across multiple virtual machines.
Flexibility — PowerVM technology runs on all Power Systems servers, from blades
to high-end servers. It provides the greatest flexibility by supporting both dedicated
and shared resource models. Unlike VMware vSphere on the x86 platform,
PowerVM virtualization technology allows virtual machines to have all dedicated
resources (CPU, memory and I/O) , or all shared resources (virtual processors,
virtual memory, virtual I/O), or a mix of dedicated and shared resources in the same
LPAR.
Quality of Service — PowerVM technology ensures that workloads achieve high
quality of service even when LPARs share processors from a shared pool.
Scalability — PowerVM technology can reduce server purchases by supporting
partitions as small as 1/10 of a processor. POWER7 processor-based high-end
systems support up to 256 physical processors in a single LPAR and up to 1,000
partitions in a system.
Availability — Live Partition Mobility (LPM) helps eliminate planned downtime by
allowing partitions to be moved to another server while running, freeing hardware
for upgrades or maintenance without interrupting productive operations. In a
system pool, LPM enables autonomic load balancing across multiple systems.
Resource pools — IBM PowerVM technology has enhanced CPU and memory
affinity to improve performance of resource-intensive workloads, such as database
workloads, across multiple virtual machines sharing resources in a system. IBM
VMControl enhancements make it easier to deploy and manage large numbers of
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 27
these virtual machines in a shared resource pool spanning one or more physical
systems.
Integrated Virtualization
Because of its level of sophistication and maturity, PowerVM technology is commonly
employed with enterprise-class applications and workloads. Power Systems servers
implement virtualization architecture with components embedded in the hardware,
firmware, and operating system software, all while running with significantly less
overhead. The capabilities of this integrated virtualization architecture are significantly
different and, in many areas, more advanced than VMware vSphere 4.1 update 1 and
other third-party software, which must be installed on x86 hardware that leverages
hardware-assist virtualization optimizations.
Power Systems servers and PowerVM virtualization technology capabilities are more
granular and more closely integrated than are those of VMware vSphere 4.1 update 1 or
Microsoft Hyper-V (or equivalent x86-based virtualization tools), as well as Oracle VM
for SPARC. The Power Systems platform also benefits from numerous industry-leading
availability optimization features. These distinctive capabilities have led to widespread
adoption of Power Systems servers to support the significantly more demanding
performance and uptime requirements of transaction- and database-intensive systems.
Greater Partition Isolation
By enabling “firmware-based” partitions, PowerVM technology provides greater
partition isolation than software-based virtualization technologies. Firmware-based
logical partitions (or virtual machines) reduce the potential for performance bottlenecks
and contribute to higher levels of availability and security than may be realized with
software-based virtualization. They also contribute to increased linear scalability.
Partitioning and Workload Management Integration
The importance of workload management cannot be overstated. Partitioning creates the
potential to utilize capacity very efficiently. The extent that this potential is realized in
practice depends on the mechanisms that allocate system resources, monitor, and
control workload execution across partitions. If these mechanisms are ineffective, a high
proportion of system capacity may remain idle at any given time.
Close integration of partitioning and workload management capabilities help prevent
surges in workloads running in individual partitions from impacting performance and
availability. POWER7 processor-based systems have a large number of cores per socket,
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 28
abundant memory, and a great deal of I/O bandwidth per core. They also support a high
number of threads per core with simultaneous multithreading (SMT). Different
workloads can benefit from different processor core thread settings; processor-intensive
workloads might benefit from using one thread (SMT1) while workloads that are I/O-
intensive can benefit from using several. POWER7 processor-based systems support up
to an SMT4 setting.
Thus, POWER7 processor-based systems consolidate an unprecedented number of
partitions and can handle workload surges more effectively, for demonstrably higher
performance.
Accommodating Greater Consolidation Density
PowerVM technology is optimized to handle business-critical systems and complex
multi-partition production environments. IBM Power Systems and PowerVM
technologies allow a high consolidation ratio and thus greater levels of efficiency in
utilization.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 29
Conclusions
Virtualization has become a pervasive means of consolidating workloads on fewer
systems, controlling server sprawl and minimizing costs. With IBM Power Systems and
PowerVM virtualization technologies, organizations can achieve virtualization with
outstanding performance. For every benchmark and every scenario covered in this
paper, IBM Power Systems with PowerVM technology demonstrated superior
performance and greater efficiency in using system capacity at higher utilization, as well
as at higher resource contention (over-commit levels), and superior scaling with higher
throughput performance.
In summary, this study has shown that IBM POWER7 systems and PowerVM
technology have demonstrated:
Higher throughput performance for both AIM7 and TPoX benchmarks, ranging from
50 percent better to as much as 200 percent better.
Higher efficiency in resource over-commit mode (higher consolidation ratio), with
the response time on PowerVM virtualization technology two to six times shorter
compared to response time for VMware vSphere 4.1 update 1 as the number of VMs
scaled from five to 40 VMs.
Higher processor affinity by default (40 VMs sharing eight cores), retaining 2.3 times
better performance than HP Intel/VMware vSphere 4.1 update 1 technologies, even
with reconfiguration using CPU affinity (VMware Scheduling Affinity group) on
VMware vSphere 4.1 update 1.
Efficient leveraging of maximum configured processor capacity.
Accurate accounting of resource usage within a VM.
Tighter integration across system, hypervisor, and guest OS.
Better performance overall than Intel Xeon 7560 /VMware vSphere 4.1 update 1.
The charts that follow summarize the results of the tests described in this paper.
The first — the AIM7 performance benchmark 32-core VM scaling (scale-up) results—
shows that PowerVM on POWER7 delivers superior scale-up efficiency that
outperforms VMware vSphere 4.1 update 1 by up to 115 percent while running the same
Linux workloads and virtualized resources.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 30
0
100000
200000
300000
400000
500000
600000
700000
Jo
bs
/min
1VM 2VM 4VM 8VM 16VM 32VM
Number of Virtual Machines
AIM7 Performance Benchmark
32 VM Scale-out on 32 cores
VMware vSphere 4.1 on HP DL580(1vcpu) VMware vSphere 4.1 on HP DL580(2vcpu)
PowerVM on Power 750(1vcpu)
In fact, PowerVM on POWER7 retains its superiority even configured with an additional
virtual CPU per VM.
The second chart — the TPOX performance benchmark 5 VM per core (scale-out) —
shows that PowerVM on POWER7 delivers robust scale-out efficiency that outperforms
VMware vSphere 4.1 update 1 by up to 201 percent while running the same workloads
and virtualized resources.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 31
0
1000
2000
3000
4000
5000
Jo
bs
/min
5VM 10VM 20VM 40VM
Number of Virtual Machines
TPOX Performance Benchmark
40 VM Scale-out on 8 cores
5 VMs per core
VMware vSphere 4.1 on HP DL580(1vcpu) VMware vSphere 4.1 on HP DL580(2vcpu)
PowerVM on Power 750(1vcpu)
PowerVM maximizes workload performance and system resources while running
multiple virtual machines on a core better than does VMware vSphere 4.1 update 1.
IBM Power Systems — with the superior performance of PowerVM virtualization
technology and with features such as reliability, security, high availability, and
resiliency — are well positioned for cloud computing and smarter planet solutions today
and in the future.
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 32
Appendices
Appendix 1 — Benchmark Configuration Information
IBM started competitive research on PowerVM virtualization in 2009 and published two
papers 13 comparing IBM POWER processor-based systems and PowerVM virtualization
technologies to Microsoft Hyper-V and VMware vSphere 4.0 update 1 running on an HP
ProLiant DL 370 G6/ Intel Xeon 5570 processors. Both these studies, which show the
superior performance of POWER processors and PowerVM technology, took a
simplified approach to answering the two most commonly expressed considerations in
deploying virtualization technologies:
1. How efficient is the technology?
2. How well does the technology scale?
The current study builds upon those simple premises to include two additional
considerations:
1. How efficient is the technology when resources are in high contention?
2. How well does it scale as virtual machine density increases?
Test Bed Setup
The servers employed for this study were chosen for their equivalencies from the
standpoint of core and socket count.
IBM POWER7 Processor-Based Server
The IBM POWER7 processor-based IBM Power 750 Express system was used in this
study to demonstrate the capabilities of IBM’s PowerVM virtualization technology. The
system was configured with four sockets, 3.5 GHz, 32 cores (eight per socket) supporting
up to four threads (SMT4) per core, and 512 GB of RAM.
13 A Comparison of PowerVM and x86-Based Virtualization Performance, Oct 2009
http://www-03.IBM.com/support/techdocs/atsmastr.nsf/WebIndex/WP101574
A Comparison of PowerVM and VMware Virtualization Performance, April 2010
http://www.spectrumconsulting.co.nz/aix/wp-content/uploads/PowerVM_VMware.pdf
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 33
HP ProLiant DL580 G7 Intel Xeon X7560 (2.26GHz/8-core/24MB/130W)
Processor
HP ProLiant DL580 G7 is a rack-mounted, high-performance Intel Xeon 7560-based
server; this system was selected to demonstrate the capabilities of VMware vSphere 4.1
update 1 virtualization technologies. The system was configured with four sockets, with
eight cores each, supporting up to two threads per core (HT mode). The system was also
enabled for Turbo Mode, Intel VTx with EPT HW Virtualization assist.
Infrastructure Configuration
System Configuration Storage Configuration
IBM Power 750, 3.5 GHz, 8 cores per
socket
POWER7 Processors, 128 GB RAM per
socket.
IBM DS4800 (4 GB cache), one 4 Gb Fiber
Channel adapter.
Each array has 12 (32 GB) disks using
RAID5.
Each array is shared by 4 virtual
machines, each getting 40 GB virtual
disk space.
HP ProLiant DL 580 G7, 2.26 GHz, 8
cores / 24 MB(4 sockets) Intel Xeon 7560
Processors, 512 GB system RAM.
IBM DS4800 (4 GB cache), one 4 Gb Fiber
Channel adapter.
Each array has 12 (32 GB) disks using
RAID5.
Each array is shared by 4 virtual
machines, each getting 40 GB virtual
disk space.
Software Used
Category PowerVM Technology VMware vSphere
Hypervisor Power Hypervisor (IBM Power
750 in-built hypervisor)
VMware vSphere 4.1 Update 1
Guest OS SuSE 11, SP1
AIX 7.1
SuSE 11, SP1 GA x86_64
RHEL6 GAx86_64
Middleware IBM DB2 v9.7 IBM DB2 v9.7
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 34
VMware vSphere 4.1 update 1 Virtual Machine Technical Configuration Details
1. VMware Virtual Machine was created using Virtual Machine version 7, which is
compatible with vSphere 4.0 hosts and greater, and provided greater virtual machine
functionality than earlier versions.
2. A Virtual Disk LSI Logic Parallel adapter was used. It was noted (in vSphere Help)
that the LSI Logic Parallel adapter and the LSI Logic SAS adapter offer equivalent
performance.
3. The VMware vSphere 4.1 update 1 system was updated to the latest VMware Tools.
4. Scheduling affinity group was used to bind cores to virtual machines.
5. Memory affinity was enabled.
6. vSpheretop –ab and vmstat were collected from the virtual machine.
Appendix 2 — General Benchmark Descriptions
The performance tests described here characterized hypervisor efficiency and scalability.
Both benchmarks stress the entire stack of application, middleware, OS, and
hypervisors. Neither benchmark requires external clients to drive the load.
The following tests were conducted:
1. Demonstrate the effect that adding virtual processors incrementally has on
throughput performance in a single VM. Where direct performance comparisons
were to be made, the testing team limited the number of virtual processors to the
lesser of the maximum supported across the two virtualization platforms.
Note: While consolidation deployments by definition entail multiple VMs,
understanding how each technology deals with processor scaling in the simplest
possible configuration within a single VM provides insights into hypervisor
efficiency.
2. Demonstrate the effect that adding VMs has on throughput performance.
Throughput is monitored as the number of VMs is scaled from 1 to n. Throughput in
each VM was also evaluated using varying numbers of virtual processors and load.
Note: This will show the effect of multiple VMs running on a system in a non-
over-commit as well as an over-commit resource environment.
Each of these tests (1 and 2) was run on different workloads. The tests included running
the same workloads (homogeneous) or a mix of workloads (heterogeneous) across
Edison: IBM – Virtualization Performance White Paper: A Comparison of PowerVM and VMware Page 35
multiple VMs concurrently. This revealed how each class of workload is affected by the
respective types of resource scaling in each test.
To ensure fair comparison across platforms and to remove variability across each set of
tests, the following actions were taken:
Similar VM configurations were deployed in terms of virtual processors and
memory allocated per VM.
The same set of “benchmark parameters” was used across platforms.
Tuning was performed based on best practices of respective platforms (VMware
vSphere 4.1 update 1, RHEL 6.0, AIX 7.1, DB2 tuning).
top related