EANTC/Spirent Cloud Computing Testing Whitepaper
With Local Area Network (LAN) and Internet bandwidth becoming less and less of a limiting factor, computer and network architectures are experiencing a paradigm shift. Data centers have been reassessed regarding both their architecture and the services they are able to provide.
Enterprises with giant data centers have thus realized that with certain technology advances they can monetize on their investment in storage and computing resources in their data center by offering them up to customers as a new type of service: public cloud computing. Enterprises building their own private data centers, or private clouds, are also interested in the benefits of these advances.
Download the white paper to learn more about Data Center testing
Cloud Computing Testing
April 2010
The New Data Center
With Local Area Network (LAN) and Internet band-
width becoming less and less of a limiting factor,
computer and network architectures are experiencing
a paradigm shift. Data centers have been reassessed
regarding both their architecture and the services they
are able to provide. Enterprises with giant data
centers have thus realized that with certain technology
advances they can monetize on their investment in
storage and computing resources in their data center
by offering them up to customers as a new type of
service: public cloud computing. Enterprises building
their own private data centers, or private clouds, are
also interested in the benefits of these advances.
Such advances have transformed the data center to a
converged and flexible environment. One of the
building blocks of the transformation has been Fibre
Channel over Ethernet (FCoE), which allowed the
convergence of LAN links (typically Ethernet) with
Storage Area Networks (SAN) links, most of which use
Fibre Channel. SAN has abstracted storage from
servers for years, however, a more recent trend has
yielded the power and flexibility of abstracting the
physical server resource from its users - virtualization.
This report details the methodology and results for a
series of tests conducted by EANTC together with
Spirent Communications. Each test focuses on a
certain piece of the data center infrastructure. With
one exception of a test conducted on public cloud
enterprise services, all tests were performed within a
private mock data center network built at Spirent
premises in Calabasas, CA. During the testing EANTC
and Spirent engineers were addressing the following
common questions:
Data Center Benchmarking - How can I measure
the performance of my new hybrid Data Center
interconnection devices?
Firewall Performance and Scalability - How does
my firewall affect the performance of my data
center when the number of user sessions increase?
Virtual Security - How do I verify that my Virtual
Machines (VMs) are secure and block connectivity
to other VMs?
Availability - How can I see my Service Level
Agreement (SLA) is honored by my provider? Do
advanced features such as VM migration affect
availability to the customers to whom I provide?
WAN Optimization - How do I prove that such
systems really increase my application
performance and provide return on investment
(ROI)?
Public Cloud Computing - Do my cloud computing
services provide me the resources and reliability on
par with my own data center?
This paper focuses on methodologies for testing poten-
tial issues arising as companies shift to a hybrid model
of both public and private cloud computing.
WAN
Aggregation
Core
AVv
STCv
Access
SANTOR
TOR - Top Of Rack (FCoE/FC/ETH capable)
Performance
Availability
Security
Scalability
Private
Public
WAN Optimizer
Firewall
WAN Optimizer
End of Row
STCv - Spirent TestCenter Virtual
AVv - Avalanche Virtual
SAN - Storage Area Network
WAN - Wide Area Network
Provier D
Provider A
Provider E
Provider B
Provider F
Provider C
Cloud Computing Testing – Page 2 of 10
Spirent TestCenter
Figure 1: Generic Data Center Context
Private Cloud Data Center Test
Scenario
The tests were grouped into two sections - testing a
private in-house data center, and testing a cloud
service from a customer perspective. For the private
data center, we assembled enough devices to build a
small emulated data center and conduct tests on the
equipment we wanted to hone-in on such as the Fire-
wall, Data Center Bridging, Virtualization, and WAN
optimization. In each specific test section below we
describe the test setup, however Figure 1 shows how
they fit together to build the data center puzzle. Since
this is a public report with the purpose of discussing
test methodology, vendor and model names tested will
remain anonymous.
Data Center Benchmarking:
Queueput
Performance of packet switching devices continues to
be a crucial and staple test for any enterprise or
service provider installing a new switch or router. The
IETF published RFCs 2544 and 2889 some 10 years
ago to define a benchmarking methodology for IP and
Ethernet (respectively) forwarding devices. These
days, as Data Center Bridging (DCB) forwards data in
a significantly differently way than LAN switches or
routers, Spirent has jointly proposed an update to the
benchmarking RFCs focusing on methodology for
benchmarking data center bridges in an individual
draft
1
.
When one speaks of a Storage Area Network (SAN),
most likely one refers to either iSCSI or Fibre Channel.
While these are seen by some as competing technolo-
gies, it is not debated that Fibre Channel has an
extremely large install base - the IDC (International
Data Corporation) tells us that $9.2 billion was spent
on Fibre Channel, FCoE, and Infiniband equipment in
2009. In the past years, as network protocols continue
to reap the benefits of convergence, the T11
committee from the InterNational Committee for Infor-
mation Technology Standards (INCITS) has defined a
way to converge Fibre Channel and Ethernet networks
- Fibre Channel over Ethernet (FCoE). The standard
facilitates reduction of cables and ports in the data
center. There was one major issue with putting Fibre
Channel over Ethernet - Ethernet is a lossy protocol.
Fibre Channel is not.
The solution came from the IEEE working group for
Data Center Bridging (DCB) that defined control proto-
cols such as DCBX (Data Center Bridging Capabilities
Exchange Protocol) for automatic discovery of
features, but mostly defined enhancements to enable
lossless Ethernet. This is done by using Priority Flow
Control (PFC, defined in 802.1Qbb) where a device
uses a special frame - a PAUSE frame - to tell the
upstream device that traffic rates are approaching a
point where loss could occur. This allows the upstream
device to stop transmitting traffic and thus avoids loss
altogether.
Since RFCs 2544 and 2889 define throughput as the
highest data rate achievable withoutframe loss,the
same terminology could not apply to DCBs. The data
center benchmarking draft proposed by Spirent now
defines queueput as the highest data rate achievable
on a queue before observing any PAUSE frames on it.
Otherwise an algorithm similar to the original RFCs is
used. Since this is done for each priority queue, it has
been dubbed not as throughput, but queueput. Addi-
tional test methodologies are defined in the IETF draft,
however, queueput measurements were performed
here.
Since most disk arrays and SANs still have native
Figure 2: Data Center Migration
LAN SAN
SANLAN
Ethernet links
Fibre Channel links
FCoE capable
Ethernet links
Cloud Computing Testing – Page 3 of 10
Fibre Channel interfaces only, hardware vendors have
started to produce switches capable of converging1. http://tools.ietf.org/html/draft-player-dcb-benchmarking-01
Ethernet and Fibre Channel by means of FCoE. These
are typically called “Top of Rack” switches (placed at
the top of a data center rack) so they may now
connect to servers, the LAN, and the SAN, leaving
much more flexibility in the data center rack design.
Such a device, was our DUT for this test.
While the queueput test can be used on any queue
configured on the device, our test focused on the FCoE
queue. Typically only one queue is configured for this
use, and is considered the utmost highest priority (it is
also usually priority code point = 3 in the VLAN
header. The Spirent TestCenter connected two 10
Gigabit Ethernet ports to our DUT and transmitted two
flows - a “low priority” Ethernet queue statically set to
5 Gbit/s, and the FCoE queue which was configured
to increase its bandwidth steadily until PAUSE frames
were received.
The standard recommends testing 128, 265, 512,
1024, 1280, 1520, and 2176 Byte frames. Traffic of
256 Byte frames reached 4.8 Gbit/s without sending
PAUSE frames, and only 4.1 Gbit/s for 128 Byte
frames. The rest of the frame sizes achieved a
queueput of 4.9 Gbit/s, seeing flow control kick in at
5.0 Gbit/s. In fact, early tests conducted before
adding the “low priority” traffic showed this queueput
result, and was the reason for statically sending 5
Gbit/s of “low priority”. This is not a major cause for
concern as Fibre Channel interface typically send
large frames. This tells the user of such a device that
they should not expect to make use of the full 8 Gbit/s
on the native Fibre Channel interfaces which would
normally translate to 6.8 Gbit/s of good data (after
encoding).
As a follow up “sanity check” test, we also tested
sending storage traffic from FCoE interfaces to Fibre
Channel interfaces. We connected 6 ports of each on
the DUT to 6 ports of each on the Spirent TestCenter
paired port configuration (Ethernet port 1 to Fibre
Channel port 1, Ethernet port 2 to Fibre Channel port
2, etc). At this low rate we did not observe any loss or
PFC (PAUSE) frames.
In addition to conducting these hybrid tests (Ethernet/
IP and FCoE on a single Ethernet port in the queueput
test, and FCoE to Fibre Channel in the above test) with
the Spirent TestCenter, we were also able to observe
the latency values. We did not set specific expecta-
tions for latency, but observed them nevertheless and
saw none that would be a cause for alarm.
WAN Optimization
Dubbed by some as WAN accelerators (which of
course do not literally change the speed of your WAN
links - unfortunately) Wide Area Network (WAN) opti-
mization hardware aims to increase efficiency of end-
to-end traffic in a few ways. Several vendors produce
such equipment, each working in slightly different
ways. Some analyze data to transmit only changed
pieces (why re-upload an entire file when you only
had to edit the authors name?) and minimizing unnec-
essary control data, some simply cache. WAN opti-
mizers’ selling point is not necessarily technical - install
one and the time you spend waiting for an upload will
be reduced.
Our setup allowed us to test point-to-point connectivity.
WAN optimizers need to store a lot of data and there-
fore consist of pretty heavy duty servers, and a few
network interfaces. Two optimizers were used in the
test, each positioned on one end of the emulated site.
Between each of our two DUTs we inserted the Spirent
GEM which adds delay and loss to traffic in an
attempt to emulate a WAN. The Spirent GEM can be
statically or dynamically configured - the latter
allowing the user to create complex configuration by
emulating specific real world-like scenarios. These
scenarios, and their impairment, have been defined
Figure 3: Queueput Results
Frame Size (Bytes)FCoE
Queueput
(Gbit/s)
Server
SAN
Figure 4: FCoE to Fibre Channel Topology
8 Gigabit
Fibre Channel
10 Gigabit
Ethernet
/x6 /x6
Farm
Cloud Computing Testing – Page 4 of 10
and transmitted 10% of line rate bidirectionally in a by the ITU-T (G.1050) and TIA (TIA-921), “Network
Model for Evaluating Multimedia Transmission Perfor-
mance over Internet Protocol”. We chose a relatively
tough, but realistic profile from the catalogue
presented by G.1050 - rate combination 131. The
standard builds a series of complex tables in order to
define the impairments for over a hundred realistic
scenarios. The 131 scenario means that each site has
a 20 Mbit/s LAN with 768 and 128 Kbit/s access
links to a 13 Mbit/s WAN. We ran tests with and
without the WAN optimizer for both 131A (low
severity conditions) and 131F (more sever conditions).
Two Spirent TestCenter ports running Avalanche soft-
ware were then connected to the two DUTs, one
emulating the data center side, and one user side of
the enterprise. A variety of applications were
emulated to reflect a realistic user profile, and to see
how each application was affected by the delay, and
optimization:
Data - File transfers on a series of real files were
executed using both HTTP and CIFS (Microsoft’s
“Common Internet File System”)
Video - Users requested an H.264 encoded MPEG
video stream with RTSP from the server.
Voice - Emulated users speaking to each other from
each location.
We collected the average download time after
running the tests for five minutes for each of the two
Spirent GEM profiles without the WAN optimizers.
After adding the optimizers to the topology we ran the
test again to allow them to cache, eliminating cache
ramp-up time as a test factor. We then repeated the
same test with the WAN optimizer active.
The results were interesting. As expected, the opti-
mizer reduced response times quite a bit for file based
did not show a noticeable difference. However, with
the less severe impairment profile the optimizer actu-
ally decreased CIFS performance. This repeated
throughout our iterations. Without speculating too
much into the algorithms of the WAN optimizer, we
can still conclude that WAN optimizers will increase
performance in most scenarios.
Going Virtual
Virtualization. The term is everywhere in the tech
world these days, and for good reason. The ability to
use the resources in a physical server for different
virtual machines gives an administrator the power to
create different user domains on a machine level,
where every user has the experience of using their
own server to themselves, and allows him to be more
flexible with the resources - Virtual Machines (VMs)
can be migrated from one server to another with
minimal interruption.
Figure 5: WAN Optimizers Test Topology
WAN OptimizerWAN Optimizer
Spirent TestCenter
Spirent GEM
Figure 6: WAN Optimization Results
CIFSHTTP
131A Impairment
Average
Download
Time
(s)
CIFSHTTP
131F Impairment
Virtual Switch
AVv
Virtual Switch
Windows
STCv
Linux
ManagementLAN
Physical Server
Eth0 Eth1
VM
VM
Cloud Computing Testing – Page 5 of 10
data such as HTTP and CIFS. Voice and video data
Figure 7: Basic Virtual Network
As CPU speeds have increased over the years, and
disks became cheaper and networks’ capacity grew,
the idea to efficiently use said resources cemented
itself under the term virtualization. The concept is fairly
simple: a server (physical machine with memory, CPU,
disk and network interface) spends a large portion of
its time idling. Why not use these free resources to
offer more services with the same hardware? To facili-
tate this virtual machines (VMs) vendors offer virtual-
ization. Many VMs can in effect run on the same hard-
ware, allowing efficient hardware resource utilization.
What does this mean for networking? As shown in
Figure 7 there is a concept of virtual switches and
virtual links, but there are also different kinds of
switches. How do you architect a virtual network?
How do you make it secure? For all of the above - how
can this be tested? Spirent has developed virtual
versions of their tools to answer the latter question. Just
as a VM is a virtual PC or server (running Windows,
Linux, etc), Spirent TestCenter virtual (STCv) and
Avalanche virtual (AVv) are virtual pieces of Spirent’s
test equipment. Their idea is to use them to address
new testing challenges in the virtual space.
Virtualization High Availability
One big advantage of virtualization is the ability to
move a virtual machine from one physical server to
another - while the VM continues to run. In order to do
this, with minimal loss, a distributed virtual switch is
required. The difference with the distributed virtual
switch compared to the standard one is its ability to
synchronize across physical platforms. Since this
feature allows public clouds to offer highly available
services both customers and the cloud service provider
have a vested interest in knowing the impact such
virtual machine move has on the service.
To measure the effect a virtual machine move has on
the service we transmitted traffic between two10
Gigabit Ethernet ports- one virtual residing in a virtual
machine and one physical within Spirent TestCenter.
We transmitted 1280-Byte frames at 1,000 frames per
second. We quickly realized then, that the virtual
world is not frame-loss free, even before conducting
the migration. We constantly observed roughly one
lost frame per 5 seconds. This meant that defining out
of service time as a summation of the number of lost
frames, would lead to inaccurate results - frames were
lost after all even when VMs were not being moved
around. The solution was to monitor the receive rate
on the physical port using Spirent TestCenter’s “High
1.5 Seconds Out of Service Time
10
Receive
Rate
on
Physical
6
4
0
2
8
Port
(
Mbit/s)
Figure 8: Availability Test Setup
STCv
Distributed Virtual Switch
Migration
STCv
Spirent
TestCenter
AfterBefore
Migration Migration
10 Gigabit Ethernet Switch
Cloud Computing Testing – Page 6 of 10
Time (min:sec:millisec)
Figure 9: Availability Results
Resolution Sampling”. After starting the sampling, and
starting traffic, we performed the migration. The
system typically took about 10-15 seconds to copy the
VM over to the new physical server, but this duration is
very much hardware dependant. We conducted the
test three times in order to gain confidence in the
results. The graph for all three runs looked similar to
Figure 9, showing about 1.5 seconds of out of service
time on the Ethernet layer. This matched the results
collected based on the frame loss observed on that
port. The results show that accepting VM High Avail-
ability in a service level agreement (SLA) should be
taken with a grain of salt. For a web server, loosing
1.5 seconds of traffic during migration could have
significant consequences. Both HTTP and TCP are
robust enough to compensate for the loss. In compar-
ison to video however, which uses UDP as transport,
the results will be devastating. This amount of frame
loss will lead to the viewer watching a blank or pixi-
lated screen and is likely to result in a very busy call
center.
Security
When it comes to installing and testing security
systems and firewalls to protect data centers from
attacks and users who should not have access, there
are generally two main questions to answer: can my
system block what it should, no more no less? and
how does the introduction of this system affect the
performance of the network? Physical hardware fire-
walls have tried their best to give impressive answers
to these questions for years but this does not help
within the virtual world. To take a peek at the status of
data center security we ran a test on both a physical
firewall and a virtual one.
The traffic profile used was the same as for our WAN
optimizer test above consisting of Voice, Video, CIFS
data and HTTPS data which filled approximately 90%
of the Gigabit Ethernet interface (the protocols do not
create a steady traffic rate, hence “approximately”).
The physical firewall connected to the Spirent Test-
Center with two Gigabit Ethernet interfaces running
Avalanche software. We ran a test both with, and
without, rules enabled (Flow 1 in Figure 10). The rules
blocked a set of 5 IP addresses (of the 250 emulated)
from all applications. The good news was that not only
did the firewall block what it should have (no more, no
less) but the CIFS and HTTPS response times, and the
voice and video MOS scores, were not negatively
To test the virtual firewall we used two different
scenarios: one where the Spirent TestCenter port was
connected directly to the server (through a switch)
running Spirent Avalanche Virtual (Flow 2 in Figure
10), and a second where only Spirent Avalanche
Virtual was used to emulate user traffic (Flow 3 in
Figure 10). The results were similar - each were a bit
less consistent since virtual performance is still unpre-
dictable, however we were impressed to find that the
virtual firewall correctly blocked only what we config-
ured it to block.
Public Cloud Computing
As explained earlier in this report, enterprises with
giant data centers are realizing that they can sell the
unused resources as a service, often called Infrastruc-
ture as a Service (IaaS), by modulating them with virtu-
alization. This is an attractive proposition for those
customers or small and medium businesses that cannot
afford to build their own data center. The price of
building a data center is only one argument against
rolling out one’s own. A second prominent argument is
expertise. Companies might rely on large data sets,
but are in effect not in the business of managing data
centers. Therefore, “outsourcing” data centers and
computing power to specialists is gaining more
ground. The increase in companies offering such
services begs the question - how do they compare?
More importantly - how can a user verify that these
Figure 10: Security Test Setup
Spirent
TestCenter
Ethernet Switch
Server
Firewall
AVv1
Virtual Switch Virtual Firewall
AVv2
Flow 1
Flow 2
Flow 3
Cloud Computing Testing – Page 7 of 10
impacted when rules were enabled. offers match the needs?
Test Setup. We contracted and established services
with six different public cloud providers. Due to the
terms and conditions of these services, we have left the
public clouds nameless. It should be noted that these
services are all priced differently, and offer slightly
different services (NetworkWorld has published an
article specifically on this aspect
1
). Still, they all
provided us with the ability to start a VM and load
applications as desired.
On each cloud provided VM we established the
typical three tier architecture: front end web server
accepting HTTPS connections, the application server,
and an SQL database. We setup one web page that
was accessible to anyone and returned a simple
string, “It Works!”. Another page was established with
a search field, which queried the SQL database. Our
SQL database was loaded with entries for equipment
type objects and fields for their price, so users could
search for an equipment type and the price would be
returned - simple, but realistic. Finally a third URL was
setup to trigger a CPU intensive task. For this we were
recommended to use the graphic software POV-Ray to
calculate the light reflections on a particular 3D
graphic. The point was the use up CPU resources, and
see how it affected our web users.
Scalability. Many enterprises already rely on public
cloud services to host their public web sites expecting
some average number of users to access their sites. In
our first test we measured each cloud’s HTTPS scal-
ability. For each cloud we simulated 410 transactions
per section (a user searching through our small under-
10-entry SQL database) using a Spirent TestCenter
running Avalanche software connected to the Internet.
We staggered the ramp up of emulated user sessions
for one minute, then attempted to hold the 410 total
sessions for another minute.
The results differed, as expected. Again, each cloud
provider is offering different hardware resources, at a
different cost. Still, we imagined ourselves in the place
of the enterprise user, without detailed knowledge of
the cloud resources, looking to host a web site with
what is provided. Such a user could use this test meth-
odology and observe that they could not realistically
plan to host a web server for hundreds of users using
the initial setup for some of these cloud providers. The
cases resulting in a significant number of failed HTTPS
transactions (see spread in Figure 12) may not neces-
sarily be a consequence of poor resources within that
public clouds data center - this may be something
configured on purpose by the provider to limit any
single user from taking too many resources. Neverthe-
less, as a user, this is a limitation that one should be
aware of so it could be discussed with the provider if it
is an issue.
Performance. The results for the scalability test
showed how many users the cloud could scale up to
with HTTPS sessions. We could not be sure, as a user
of the cloud, where their network or their CPU could
have been a limiting factor. To focus on CPU perfor-
mance of the clouds, we repeated the test, but config-
ured just 5% of the HTTPS transactions (insignificant
bandwidth usage) to call a CPU intensive task - POV-
Ray, a graphics program, would recalculate the light
reflections across our graphic’s surfaces. The same
scalability test was run both with and without the POV-
Ray task in parallel, and instead of measuring the
number of failed sessions, we compared the HTTPS
response times for each. This answered two questions;
1. http://www.networkworld.com/reviews/2010/040510-
Spirent
Provider D
Provider A
Provider E
Provider B
Provider F
Provider C
Internet
Figure 11: Public Cloud Setup
TestCenter
Provider
A
Figure 12: HTTPS Scalability Results
Provider
B
Provider
C
Provider
D
Provider
E
Provider
F
%
Succesful
HTTPS
Transactions
100
40
20
0
60
80
Cloud Computing Testing – Page 8 of 10
how long it take for the SQL database query to be
cloud-computing-test.html?page=1
returned to the user, when the server was processing
the POV-Ray task in parallel?
The results were surprising. The clouds which excelled
in our scalability test measured the longest HTTPS
response times when the CPU-intensive task was initi-
ated. To note, the response times are only of the stan-
dard SQL query transactions, and do not count the
POV-Ray transactions themselves. Therefore, the results
show us how heavy CPU load affects all users’
response times. In conclusion, each public cloud
provider may have its strengths, but there may also be
some weaknesses. That said, it may simply be the pref-
erence of the majority of those clouds’ customers to
have CPU intensive tasks prioritized. Our method-
ology aims to produce a tool for such an observation
to be seen by the end user.
Conclusion
The industry is pretty excited about cloud computing
and these latest advances to the data center, and
understandably so. Still, this is a developing techno-
logical area, and as such there is still a ways to go.
In our private cloud tests we found that performance of
data center bridges is not entirely straight forward, but
they succeed in convergence. Our WAN optimizer
did a great job in most cases, but may not be worth
the financial investment for all user scenarios. Firewalls
have progressed into the virtual space, yet the unpre-
dictability of virtualized resources proves to still be an
area of continued study.
Our public cloud computing tests have only begun to
scratch the surface. The services provided us with an
effective way to outsource a web server and data base
function to the cloud, but what about the performance
of these services in the cloud? How does the virtual
machine of one user affect the behavior of another
virtual machine running on the same physical server in
the same public cloud? We look forward to continuing
and expanding these tests. We hope to hear from
anyone interested in repeating these tests, or
providing suggestions for further testing.
INTEROP Las Vegas - April, 2010
This white paper report has
been distributed at the INTEROP
conference in Las Vegas in April
of 2010 alongside a live presen-
tation of the tests described
here. We would like to note that
Spirent Avalanche Virtual, which
was used in the testing, was awarded “Best of Interop”
in the Performance and Optimization category by the
Figure 13: HTTPS Performance Results
HTTPS
Response
Times
(s)
3
2
1
0
Provider
A
Provider
B
Provider
C
Provider
D
Provider
E
Provider
F
Cloud Computing Testing – Page 9 of 10
conference organizer.
EANTC AG
European Advanced Networking Test Center
Spirent Communications
Einsteinufer 17
10587 Berlin, Germany
Tel: +49 30 3180595-0
info@eantc.de
http://www.eantc.com
26750 Agoura Road
Calabasas, California 91302
Tel: +1-800-SPIRENT / +1-818-676-2683
http://www.spirent.com
This report is copyright © 2010 EANTC AG. While every reasonable effort has been made to ensure accuracy and
completeness of this publication, the authors assume no responsibility for the use of any information contained herein.
All brand names and logos mentioned here are registered trademarks of their respective companies in the United States and
other countries.
20100505 v2.1
Cloud Computing Testing – Page 10 of 10