White Paper c11 741610
White Paper c11 741610
White Paper c11 741610
Cisco public
Two trends are driving the design of today’s network edge: the growth in unicast streaming video,
and the continued growth of Distributed-Denial-of-Service (DDoS) attacks. The Cisco Visual
Networking Index™ has shown a 1270-percent rise in Internet traffic over the last 12 years and
projects a threefold increase in the next five years. Cisco estimates that peering bandwidth, in
particular, is growing at a steady 30-percent Compound Annual Growth Rate (CAGR), and some
carriers are already planning for growth as high as 50 percent per year. This is mostly driven by
a steady increase in unicast video streaming and other content sources versus traditional peering
expansion with other network providers.
Capacity planning for this growth in legitimate traffic is challenging enough for operators; scaling the network
edge with many more distributed peering points also risks exposing your network to inbound DDoS attacks. This
new distributed architecture increases the overall threat surface at the network edge and can potentially expose
the network core to much larger DDoS attacks.
As events in early 2018 have already proven, terabit-scale volumetric DDoS attacks are becoming more
frequent and more damaging. These types of attacks can easily overwhelm internal network capacity and even
centralized DDoS mitigation scrubbing facilities. The potential risks to a highly distributed peering architecture
are significant since they can expose the network core to significantly higher volumes of malicious traffic than
previous architectures.
© 2018 Cisco
© 2018and/or
Ciscoitsand/or
affiliates.
its affiliates.
All rightsAll
reserved.
rights reserved.
White paper
Cisco public
The DDoS threat As peering architectures become more distributed and regional, edge
landscape routers must play a more active role as the first line of defense against these
destructive volumetric attacks. The key is rapid detection using NetFlow-
DDoS attacks are a constant based telemetry combined with mitigation at the edge using Border Gateway
threat and are heavily used to Protocol Flow Specification (BGP FlowSpec) and hardware-based blocking
attack Internet-based services in the edge routers to contain the attacks before they consume critical
and networks, with the goal to resources deeper in the network.
block access to these services
and to disrupt Internet access.
The motivations behind these Keep calm, and binge on
attacks vary widely—from There is a silver lining to the traffic growth trends: the increase in peering
pranks to state-sponsored to capacity is driven primarily by content, but content providers are not the
cybercrime. A good overview source of DDoS attacks.
can be found in NETSCOUT
Arbor’s 13th Annual Worldwide Inbound Internet traffic can be coarsely separated into two categories:
Infrastructure Security Report content-peering traffic coming from content providers and Content Delivery
(Arbor WISR). Networks (CDN), and public-peering traffic originating from residential and
business endpoints connected through traditional ISPs. Netflix, Google,
According to Arbor’s Active Akamai, and Facebook are just a few examples of content sources that fall
Threat Level Analysis System into the first category. These content sources dominate today, comprising 60
(ATLAS®), which monitors more percent of all Internet traffic in 2018, and are projected to exceed 70 percent
than one-third of all Internet of all Internet traffic by 2021.
traffic, attack frequency is on
the upswing and is estimated Figure 3. CDN versus non-CDN traffic
to reach eight million in 2018,
CDN VS. NON-CDN TRAFFIC
a 16-percent increase over
2017. In addition, the proportion CDN Non-CDN Traffic
Cisco + NETSCOUT in overall peering bandwidth. As indicated in Figure 3, public peering capacity
Arbor: Proven solutions (non-CDN) is growing at a much more moderate rate compared to content
peering capacity (CDN) and the overall level of Internet transit traffic exposed
for infrastructure DDoS
to the worldwide Internet is on the decline.
protection
Continued traffic growth at more than 30 percent CAGR, comprised primarily
• Cisco is a proven peering
of content provider traffic, requires a scalable and cost-effective Internet
solution, with the Cisco®
peering solution. The peering solution must also integrate with a DDoS
Network Convergence System
mitigation solution employing best practices to mitigate all types of DDoS
(NCS) 5500 deployed at more
attacks. Unproven techniques such as applying advanced edge filtering to
than 40 service providers
cover all ingress peering traffic incurs additional cost to known good traffic
in a peering role today and
(i.e., reliably not a source of DDoS).
thousands of Cisco IOS XR
powered routers used for
peering across the world. These Decentralized peering needs decentralized
high-density edge routers
with scalable DDoS mitigation DDoS protection
capabilities have proven to be Attacks such as the Memcached reflection attack in early 2018 prove that a
effective against the landscape single attack can sustain multi-terabit traffic levels. This is more than enough
of volumetric attacks. bandwidth to overload the capacity of most network operators’ internal
• Arbor’s Threat Mitigation infrastructure and jeopardize network availability.
System (TMS®)—an intelligent
This adds a new twist to classic DDoS attacks and requires a re-evaluation
detection and mitigation system
of DDoS mitigation architectures. While traditional techniques focus primarily
(IDMS)—incorporates NetFlow-
on diverting large DDoS flows to active mitigation scrubber farms in order to
based analytics and detection,
protect users and services from attack, today’s new decentralized peering
FlowSpec-based and DPI-
architectures need an augmented approach that supports both decentralized
based DDoS mitigation, and
and centralized DDoS mitigation where it makes sense.
end-to-end workflows in a
single integrated solution with Figure 4. Decentralized DDoS attack mitigation using the peering edge router
infrastructure
Cisco routers.
• NetFlow-based detection VO
LU
M
plus BGP FlowSpec for DDoS ET
RI
C
mitigation is widely deployed,
scales to multi-terabit network
topologies, and utilizes
industry-standard protocols.
Distributed Mitigation
at Peering Sites
VOLUMETRIC
VOLUMETRIC
The motivations behind the attacks are complex (and discussed in more detail later in this paper), but in many
cases large volumetric attacks are either targeting peering infrastructure directly (including internal core bandwidth)
or serving as a smoke-screen to cover up more sophisticated and surgical attacks, such as penetration attempts
on exposed services. In the latter case, the attackers are motivated to cause as much operational disruption and
distraction as possible—including monitoring and rapidly mutating their attacks to evade static mitigation techniques.
A traditional DDoS mitigation strategy has been to implement strong (and computationally costly) application layer
traffic controls to essentially “ride it out” and absorb significant DDoS attacks. While this is the right strategy for
attacks targeting the state machine or application layer of services on the network, these attacks represent a
smaller overall proportion of DDoS traffic. These application-targeted attacks are best handled with traditional active
mitigation devices within the network. This comes at a cost, however: the capital infrastructure requirements for
building and maintaining large-scale scrubbing facilities, as well as the core bandwidth required to backhaul large
volumes of attack traffic to centralized locations in the backbone. Recent attack sizes demonstrate this approach
won’t scale as the cost of backhauling attack traffic for centralized scrubbing becomes prohibitive and transiting these
traffic volumes adds risk to core network operations. Instead, carriers need a more effective means to quickly block
this offending traffic at distributed peering locations.
1 Volumetric Attack
strikes from external
peering links.
While older attacks often included simple signatures that were consistent across all attack patterns (such as non-
standard use of TCP flags), attackers have rapidly updated their tools to prevent providers from easily filtering known-
bad attack patterns. Any solution to deal with motivated modern attackers must quickly detect new patterns (both
in sources and protocols, as well as payload patterns), and mitigate before the attackers can modify their attack
vectors. Attackers are relying on the premise that they can morph their attacks faster than providers can adapt their
mitigations. The only way to block these attacks is to quickly detect and automate mitigations.
The best solution is to automate the process of both detecting new potential attacks through intelligent network
analytics and to provide an automated workflow that instantiates mitigations directly into the edge routers using the
BGP FlowSpec protocol’s inherently dynamic capabilities.
In examining various DDoS mitigation architectures, it’s important to distinguish protection of the network
infrastructure itself from delivery of subscriber-facing DDoS mitigation services. In both scenarios, today’s distributed
peering architectures demand a multi-layer approach. Protection of the infrastructure itself should be distributed to
peering locations using a combination of advanced flow analytics and hardware-based blocking in edge routers.
For carriers that provide revenue-generating managed DDoS protection services to their downstream subscribers, a
scalable, centralized mitigation solution is also required: it provides a second layer of defense against complex and
application-layer attacks that manage to penetrate to the core network.
Modern DDoS attacks have proven to be highly dynamic in nature. Rapid and scalable flow analytics are therefore
of paramount importance in quickly detecting potentially malicious traffic patterns, signaling appropriate mitigation
strategies, and providing an automated framework for triggering mitigations, both in edge routers and intelligent DDoS
mitigation systems such as Arbor’s TMS.
Arbor’s Sightline® Traffic Routing and Analytics platform is capable of scaling to these terabit network topologies:
it consumes BGP, Simple Network Management Protocol (SNMP), and NetFlow IP Flow Information Export (IPFIX)
telemetry to provide network-wide traffic visibility; it rapidly detects potential attacks; it can signal mitigations directly
to edge routers using BGP FlowSpec as well as to Arbor’s TMS to address the full range of volumetric and more
complex state exhaustion and application layer DDoS attacks.
Edge routers must support NetFlow version 9 and IPFIX flow telemetry, as well as the ability to scale it to support the
multi-hundred gigabit traffic volumes now common at large peering sites. Platforms such as Cisco’s NCS 5500 and
ASR 9000 Series Aggregation Services Routers are designed to deliver the scale required to support today’s largest
peering environments, including maintaining the high rates of NetFlow/IPFIX telemetry essential to advanced DDoS
traffic analytics, detection, and mitigation. And filtering this large-scale traffic at the edge without blocking traffic from
legitimate sources requires high access-control-entry scale. Cisco has implemented high-efficiency ACLs in Cisco
IOS-XR Software to mitigate distributed attacks comprising thousands of source addresses.
FlowSpec Mitigation
via Cisco IOS XR
Netflow/IPFIX
Arbor
Sightline
BGP FlowSpec
BGP FlowSpec offers multiple match conditions to identify traffic based on standard, five-tuple headers such as IP
source and destination, IP protocol, Transmission Control Protocol (TCP)/User Datagram Protocol (UDP) source,
and destination ports, as well as more granular criteria such as TCP headers, Differentiated Services Code Point
(DSCP), and IP packet length. Together, these match criteria can identify each of the volumetric attack types currently
observed in the wild.
Once matched, BGP FlowSpec can also support multiple actions, including simple packet drop, policing, DSCP
marking, as well as the ability to redirect packets to a specified Virtual Route Forwarding (VRF) or IP next-hop. This
provides a powerful framework for either controlling traffic directly within the edge routers or redirecting offending
traffic to an external Integrated Database Management System (IDMS) for deep packet analysis and application
layer traffic control. These IDMS can be located either adjacent to large peering facilities, or centrally in large-scale
mitigation facilities.
FlowSpec Redirection
via Cisco IOS XR
Netflow/IPFIX
Arbor
Sightline
BGP FlowSpec
In a distributed mitigation architecture, IDMS devices such as Arbor’s TMS are deployed adjacent to peering
routers, either through dedicated appliances or virtual appliances deployed in a Virtualized Network Function (VNF)
architecture. This architecture builds on the capabilities of the distributed NetFlow and FlowSpec mitigation solution,
as malicious traffic can now be redirected to—and blocked by—intelligent mitigation resources local to the peering
site. This solves the problem of consuming internal network capacity to transport suspected attack traffic back to
centralized scrubbing centers and is a fundamentally more effective solution for infrastructure protection against
advanced attacks.
Arbor’s Sightline solution can simultaneously announce real-time traffic control directly to edge routers using BGP
FlowSpec and provide a second layer of protection against more complex attacks when paired with Arbor TMS.
These attacks can easily overwhelm even large-scale enterprise services but present as much lower overall
bandwidth attacks (typically less than 10 gigabits per second), so they are not normally considered a direct threat
to the network provider’s infrastructure. These types of attacks cannot be mitigated in stateless edge router
infrastructure: they require intelligent L3-L7 mitigation front-ending individual services and applications, and
centralized mitigation architected to protect the provider’s infrastructure and all its downstream customers and by
leveraging the economies of scale (multi-tenancy) possible with centralized scrubbing.
Figure 8. Announcing FlowSpec rules to an edge router to redirect an incoming volumetric attack to distributed intelligent
DDoS mitigation resources; dynamically redirecting inbound application-layer attack traffic from edge routers to centralized
scrubbing resources
VOLUMETRIC
FlowSpec Redirection
via Cisco IOS XR
Netflow/IPFIX Arbor
Sightline
BGP FlowSpec
APP Attack
End User
ck
ta
FlowSpec Redirection
At
Application layer and state exhaustion attacks can also be detected using security-focused flow analysis but as they
are in almost all cases low-volume-type attacks, it is often necessary to use behavioral analysis or deep packet
analysis to detect these kinds of attacks. This will usually require the use of an IDMS to provide visibility into the
application traffic and detect the specific attack vector, so application layer traffic can be protected. The traffic is
usually diverted to the IDMS using network diversion, which will then mitigate and block the attack on the device itself
or offload the blocking to the edge devices.
As the attacks can be complex in nature and a determined attacker will rapidly change the attack vector when
an earlier attack vector is mitigated, the IDMS uses a set of methods to analyze and block these kinds of attacks,
including:
• Challenge/response mechanisms
• Behavioral analysis
• Flow and rate-based analysis
• Response analysis
• Various other approaches, depending on the application under attack and the attack vector being used
Nevertheless, recent examples of attacks by determined DDoS attackers show that even in cases where such patterns
can be identified, they will simply change the attack parameters as soon as they see that the current parameters are
no longer effective and launch a different type of attack.
• For example, when the Mirai IoT botnet was introduced in October 2016, Arbor published four regex patterns to help
customers mitigate these attacks. Since then, most of the initial attack patterns have changed as the original flaws
in the attack tool were either fixed by the attackers or the attackers moved on to new attacks. As a result, Arbor
frequently updates the patterns and associated guidance to customers. Attackers usually parameterize these attack
tools, so in many cases when they launch their attacks, these specific parameters can be detected and blocked using
an IDMS.
• In addition, as part of Arbor’s ATLAS Intelligence Feed (AIF), Arbor regularly publishes new HTTP regex patterns to
match new HTTP-based botnet attacks. The number of these patterns has gradually declined as the attack tools
have become more advanced, making it more and more difficult to distinguish attackers from valid users based on
patterns in the attack payload.
• State exhaustion and application layer attack types are designed to be indistinguishable from legitimate traffic and
operate at relatively small traffic volumes. These attacks rarely have signatures that can be matched in stateless
payload filters. Active mitigations through an IDMS are typically required to identify these types of DDoS attacks.
Refer to “Appendix 2: DDoS attack types and mitigation methods” for more detail.
Arbor has already determined that actively monitored attacks can quickly change payload patterns, and attackers will
continuously update their attacks to evade detection (often in timeframes as short as half an hour). This behavior has
already rendered static pattern matching techniques, such as the ones proposed for ASIC-based payload matching
obsolete and ineffective at mitigating these attacks.
• The experience of large-scale ISPs shows that existing amplification attacks are mitigated using filters that identify
ports and packet size
• Fields within DNS can be of varying lengths, which can render static payload signatures ineffective (which rely on
matching patterns at specific offsets in the packet payload)
• Filtering based on payload signatures does not offer any additional value in blocking non-initial fragments of DNS
amplification attacks comparing to traditional FlowSpec
• DNS amplification attacks can also involve IP fragmentation, where only the initial packet contains useful protocol
headers. This poses a problem for any type of payload pattern matching, as IP fragments are highly unlikely to have
consistent payload patterns at known offsets within the fragmented packet.
As a result, the cases where payload-based filtering might make more sense versus standard IP header rules alone
do not justify investments into payload-based filtering: UDP random packet floods are not the most common attacks
and all application layer attacks, including Session Initiation Protocol (SIP)/SNMP and others represent extremely low
bandwidth. Moreover, both of them are not common attacks against infrastructure, residential, and SMB customers
that larger service providers want to protect. Typical attacks against these markets are reflection and amplification
attacks, and as analysis demonstrates, FlowSpec rules matching standard L3 and L4 headers are effective against
these attacks.
Arbor has also found that when implemented in production networks, pattern matching tends to have a high rate of
false positives for volumetric attacks. Not only is payload pattern matching likely be ineffective for mitigation, it also
introduces operational issues such as unintended and unpredictable drops of legitimate customer traffic. As these
patterns would match against packets from all sources, payload pattern packet drops would be impossible to debug
in a large production network. Arbor found these issues also occurred when using regex pattern matching, which is
significantly more flexible than the static payload patterns proposed for ASIC-based mitigation in competing routing
platforms. Further, to effectively deploy pattern-based mitigation, it’s critical to understand and protect “known-good”
traffic patterns. But in networks transporting unpredictable end-user traffic, as opposed to well-defined services, it’s
nearly impossible to accurately and consistently profile “known-good” traffic.
1. Using flow telemetry analysis supplemented with behavioral analysis to detect abnormalities and attacks. Focus on
understanding what is normal: it will simplify identification of abnormalities.
2. Using BGP FlowSpec to activate network-based blocking at the edges of the routed network once a volumetric
attack is detected.
3. Using an IDMS to detect abnormal behavior, and application layer and stateful exhaustion attacks that require
advanced and active mitigation; and using this approach in conjunction with BGP FlowSpec Offload when and
where appropriate.
If implemented successfully, these protections will force the attacker ‘network’ to behave like normal clients, rendering
the DDoS attack ineffective and allowing for the use of application-level analysis to detect any abnormal traffic or
usage patterns.
Arbor Cisco NCS 5500 Arbor Cisco NCS 5500 Arbor Cisco NCS 5500
Sightline Sightline Mitigation Sightline Mitigation
template template
BGP FlowSpec
BGP FlowSpec
Arbor Arbor
Blacklist Offload
UDP vector = drop TMS TMS
BGP FlowSpec BGP FlowSpec
UDP vector = drop IP VICTIM = REDIRECT
UDP vector = drop
IP VICTIM = REDIRECT Legitimate Legitimate
IP VICTIM = REDIRECT
traffic traffic
Complex L3-L7 Complex L3-L7 Complex L3-L7 Complex L3-L7 Complex L3-L7 Complex L3-L7
attack attack attack attack attack attack
Access Network Access Network Access Network
(Fixed/Mobile) (Fixed/Mobile) (Fixed/Mobile)
Arbor Sightline is the foundation of intelligent attack detection, using interface statistics, NetFlow, and BGP data
to identify both known and unknown attacks. Granular NetFlow data export is critical to identify attacks. Cisco, the
inventor of NetFlow, provides class-leading sampled NetFlow capabilities on both its NCS 5500 and ASR 9000 Series
routers. Once Arbor Sightline has detected an attack, automated attack mitigation is signaled to the Cisco peering
edge to block, police, or redirect traffic to an IDMS. Using the simplified traffic engineering capabilities of Segment
Routing, a router can steer traffic to a specific IDMS using either static or dynamic SR-TE Policies. Why risk using a
less effective approach?
1. Malware arms dealers 2. The DDoS mercenaries offer 3. The attackers mostly use
are either individuals or DDoS services Booter/Stresser services to
organisations which research (Booters/Stresser) for hire to launch their attacks, though
and develop attack tools that the attackers. there are some exceptions.
take advantage of security
vulnerabilities. As part of
their Q&A, often do live field
testing.
DDoS mercenaries also offer their services for hire for launching attacks on specific customers, using a
combination of “Booter” and ”Stresser” tools and by using focused attacks.
• Attempting to disconnect gaming users as a result of online gaming disputes. The attacks are either launched directly
against the target user or against the gaming infrastructure the user is connected to. These attacks are launched
using Booter and Stresser services, often with high volumes but short durations. These attackers are usually low-
skilled and mostly do not monitor the results of their attacks.
• Making a point, often during political disputes where multiple low-skill attackers join forces in launching volumetric
attacks, usually using Booter and Stresser services or simple attack tools. These attacks are often launched in
combination with application layer attacks.
• Launching a determined attack against a specific target, usually critical service provider services or enterprise
customers. A highly skilled attacker usually combines volumetric attacks with application layer attacks to hide the
more focused application layer attacks, which do the real damage.
Examples of such attacks are reflection attacks using Memcached, Network Time Protocol (NTP), DNS, and Simple
Service Discovery Protocol (SSDP), which flood the destination with large reply packets, filling up links and, in some
cases, resulting in the collapse of the target network infrastructure.
These attacks are typically very high-bandwidth (up to 100 Gbps or more and occasionally exceeding Terabits per
second) and are immediately obvious to both the target and upstream connectivity providers. This typically gets
immediate attention from security and network operations teams. Because of this, determined attackers have learned
to actively monitor the results of their attacks and often randomize their attack parameters as soon as defenders start
to block or limit the current attack vector.
Application layer and state exhaustion attacks are usually low-volume compared to volumetric attacks since they
have to conform to the protocol the application itself is using, which often involves protocol handshakes and protocol/
application compliance. This means that these attacks will primarily be launched using discrete intelligent clients,
usually IoT devices, and cannot be spoofed. The good news is defenders can identify those attacking hosts as a result
and block them using intelligent DDoS mitigation.
But even when multi-vector attacks contain identifiable patterns, a determined attacker will monitor the results of his
attack and modify it to thwart a skilled and determined defender. Because active attackers are known to continually
modify payload patterns to avoid simplistic mitigations, maintaining an ongoing list of known attack patterns quickly
becomes impractical due to scale issues and the rate at which this list must be updated. Further, since payload
patterns bring high risk of causing collateral damage, maintaining a long-lived set of payload patterns may be unwise.
DNS amplification 47, 9% of all 100 Gbps+ Yes, based on UDP DNS amplification
volumetric attacks, ports and packet length.
according to Arbor Exceptions are responses
ATLAS data based on EDNS0, e.g.
DNSSEC.
An additional FlowSpec
filter is required to block
UDP fragments to the
victim.
NTP, SSDP, 52.1% of all 100 Gbps+ Yes, based on ports and NTP, SSDP,
Memcached, volumetric attacks, packet size. An additional Memcached,
Chargen, C-LDAP, according to Arbor FlowSpec filter is required Chargen, C-LDAP,
SNMP, Portmap, ATLAS data to block UDP fragments SNMP, Portmap,
MSSQL, and other to the victim. MSSQL, and other
amplifications amplifications
TCP SYN, TCP RST, Vast majority of Less than No Yes, using a challenge/
TCP ACK session exhaustion 100Gbps response-based approach
attacks
Idle TCP, UDP Less typical Less than No, the attack Yes, using behavioural
connections attacks 10Gbps uses valid TCP and session analysis and
UDP sockets dropping inactive sessions
UDP random Less typical Less than No, if the attack is Yes, using rate-based
packet flood attacks 100Gbps destined to a valid analysis, session analysis,
active UDP socket and challenge-response
mechanisms
ICMP, GRE, and Less typical Less than Yes, if the victim Yes
other random IP attacks 100Gbps is not expecting
protocols those protocols
Although there are a few corner cases where these attacks can be generated using poorly crafted tools that utilize
static tcp.window_size or ip.ttl, the majority of attacks employ randomized TCP fields which are indistinguishable
from legitimate traffic. Therefore, patterns based on L3-L4 headers do not work. Also, Arbor’s Security Engineering
& Research Team (ASERT) has observed some attacks utilizing TCP SYN flood with payload, mostly launched by
XOR-based tools. Those attacks are not popular because they are easily detected and mitigated without payload
analysis. For example, a FlowSpec policy that drops TCP SYN traffic with a packet length of more than 40 bytes (plus
necessary encapsulations) is more than enough to successfully block in this case.
TLS (L5) Less typical attacks Around 1 Gbps, No, behavioral Yes, by analyzing the TLS session
usually less analysis of TLS setup, checking TLS extensions,
session setup is idle sessions, and so on
needed
HTTP 73% of Arbor’s WISR Around 1 Gbps, No Yes, by analyzing the rate of
enterprise respondents usually less requests, performing challenge/
observed those attacks
response, and applying known
patterns of attacks
DNS 69% of Arbor’s WISR Around 1 Gbps, No Yes, by analyzing the rate of
enterprise respondents usually less requests, performing challenge/
observed those attacks response, downloading protected
DNS zone (zone transfer), and
applying known patters of attacks
MedusaHTTP is an example of an HTTP botnet that uses a list of randomly chosen user-agent HTTP headers and
allows using a built-in browser that is indistinguishable from real user traffic. https://asert.arbornetworks.com/
medusahttp-ddos-slithers-back-spotlight/
Many DNS attacks against DNS resolver or authoritative servers are based on label prepending (so-called “water torture
attacks”) where attackers generate a lot of queries using randomized DNS subdomains. A good example is the famous
DYN attack. It is not possible to define a pattern around illegitimate queries due to the randomness of queries. It is also
not possible to define a whitelist pattern to block everything else for DNS resolvers. For DNS authoritative servers, it is
possible to whitelist the protected zone, however the scale (the number of domains in a zone cut) might be an issue.
© 2018 Cisco and/or its affiliates. All rights reserved. Cisco and the Cisco logo are trademarks or registered trademarks of Cisco and/or its affiliates in the U.S. and other countries. To view a list of
Cisco trademarks, go to this URL: https://www.cisco.com/go/trademarks. Third-party trademarks mentioned are the property of their respective owners. The use of the word partner does not imply
a partnership relationship between Cisco and any other company. (1110R) C11-741610-00 11/18