Enforcing Network-Wide Policies in the Presence

Enforcing Network-Wide Policies in the Presence
Enforcing Network-Wide Policies in the Presence
of Dynamic Middlebox Actions using FlowTags
Seyed Kaveh Fayazbakhsh∗ Luis Chiang† Vyas Sekar∗ Minlan Yu‡ Jeffrey C. Mogul⋆
† Deutsche Telekom Labs
⋆ Google
∗ Carnegie Mellon University
Middleboxes provide key security and performance
guarantees in networks. Unfortunately, the dynamic traffic modifications they induce make it difficult to reason
about network management tasks such as access control,
accounting, and diagnostics. This also makes it difficult
to integrate middleboxes into SDN-capable networks and
leverage the benefits that SDN can offer.
In response, we develop the FlowTags architecture.
FlowTags-enhanced middleboxes export tags to provide
the necessary causal context (e.g., source hosts or internal cache/miss state). SDN controllers can configure
the tag generation and tag consumption operations using
new FlowTags APIs. These operations help restore two
key SDN tenets: (i) bindings between packets and their
“origins,” and (ii) ensuring that packets follow policymandated paths.
We develop new controller mechanisms that leverage
FlowTags. We show the feasibility of minimally extending middleboxes to support FlowTags. We also show that
FlowTags imposes low overhead over traditional SDN
mechanisms. Finally, we demonstrate the early promise
of FlowTags in enabling new verification and diagnosis
Many network management tasks are implemented using custom middleboxes, such as firewalls, NATs, proxies, intrusion detection and prevention systems, and
application-level gateways [53, 54]. Even though middleboxes offer key performance and security benefits,
they introduce new challenges: (1) it is difficult to ensure
that “service-chaining” policies (e.g., web traffic should
be processed by a proxy and then a firewall) are implemented correctly [49, 50], and (2) they hinder other management functions such as performance debugging and
forensics [56]. Our conversations with enterprise operators suggest that these problems get further exacerbated
with the increasing adoption of virtualized/multi-tenant
The root cause of this problem is that traffic is
modified by dynamic and opaque middlebox behaviors. Thus, the promise of software-defined network-
USENIX Association
ing (SDN) to enforce and verify network-wide policies
(e.g., [39, 40, 44]) does not extend to networks with middleboxes. Specifically, middlebox actions violate two
key SDN tenets [24, 32]:
1. O RIGIN B INDING: There should be a strong binding
between a packet and its “origin” (i.e., the network
entity that originally created the packet);
2. PATHS F OLLOW P OLICY: Explicit policies should determine the paths that packets follow.1
For instance, NATs and load balancers dynamically
rewrite packet headers, thus violating O RIGIN B INDING.
Similarly, dynamic middlebox actions, such as responses
served from a proxy’s cache, may violate PATHS F OL LOW P OLICY. (We elaborate on these examples in §2.)
Some might argue that middleboxes can be eliminated
(e.g., [26, 54]), or that their functions can be equivalently provided in SDN switches (e.g., [41]), or that
we should replace proprietary boxes by open solutions
(e.g, [20, 52]). While these are valuable approaches,
practical technological and business concerns make them
untenable, at least for the foreseeable future. First, there
is no immediate roadmap for SDN switches to support
complex stateful processing. Second, enterprises already
have a significant deployed infrastructure that is unlikely
to go away. Furthermore, these solutions do not fundamentally address O RIGIN B INDING and PATHS F OLLOWP OLICY; they merely shift the burden elsewhere.
We take a pragmatic stance that we should attempt to
integrate middleboxes into the SDN fold as “cleanly” as
possible. Thus, our focus in this paper is to systematically (re-)enforce the O RIGIN B INDING and PATHS F OL LOW P OLICY tenets, even in the presence of dynamic
middlebox actions. We identify flow tracking as the key
to policy enforcement.2 That is, we need to reliably associate additional contextual information with a traffic flow
as it traverses the network, even if packet headers and
1 A third SDN tenet, H IGH L EVEL NAMES , states that network policies should be expressed in terms of high-level names. We do not address it in this work, mostly to retain backwards compatibility with
current middlebox configuration APIs. We believe that H IGH L EVEL NAMES can naturally follow once we restore the O RIGIN B INDING
2 We use the term “flow” in a general sense, not necessarily to refer
to an IP 5-tuple.
11th USENIX Symposium on Networked Systems Design and Implementation 533
contents are modified. This helps determine the packet’s
true endpoints rather than rewritten versions (e.g., as with
load balancers), and to provide hints about the packet’s
provenance (e.g., a cached response).
Based on this insight, we extend the SDN paradigm
with the FlowTags architecture. Because middleboxes
are in the best (and possibly the only) position to provide the relevant contextual information, FlowTags envisions simple extensions to middleboxes to add tags, carried in packet headers. SDN switches use the tags as part
of their flow matching logic for their forwarding operations. Downstream middleboxes use the tags as part of
their packet processing workflows. We retain existing
SDN switch interfaces and explicitly decouple middleboxes and switches, allowing the respective vendors to
innovate independently.
Deploying FlowTags thus has two prerequisites: (P1)
adequate header bits with SDN switch support to match
on tags and (P2) extensions to middlebox software. We
argue that (P1) is possible in IPv4; quite straightforward
in IPv6; and will become easier with recent OpenFlow
standards that allow flexible matching [9] and new switch
hardware designs [23]. As we show in §6, (P2) requires
minor code changes to middlebox software.
Contributions and roadmap: While some of these arguments appeared in an earlier position paper [28], several practical questions remained w.r.t. (1) policy abstractions to capture the dynamic middlebox scenarios; (2)
concrete controller design; (3) the viability of extending
middleboxes to support FlowTags; and (4) the practical
performance and benefits of FlowTags.
Our specific contributions in this paper are:
• We describe controller–middlebox interfaces to configure tagging capabilities (§4), and new controller
policy abstractions and rule-generation mechanisms
to explicitly configure the tagging logic (§5).
• We show that it is possible to extend five software
middleboxes to support FlowTags, each requiring less
than 75 lines of custom code in addition to a common
250-line library. (To put these numbers in context, the
middleboxes we have modified have between 2K to
over 300K lines of code.) (§6).
• We demonstrate that FlowTags enables new verification and network diagnosis methods that are otherwise
hindered due to middlebox actions (§7).
• We show that FlowTags adds little overhead over SDN
mechanisms, and that the controller is scalable (§8).
§9 discusses related work; §10 sketches future work.
Background and Motivation
In this section we present a few examples that highlight how middlebox actions violate O RIGIN B INDING
and PATHS F OLLOW P OLICY, thus making it difficult to
enforce network-wide policies and affecting other management tasks such as diagnosis. We also discuss why
some seemingly natural strawman solutions fail to address our requirements.
Motivating Scenarios
Attribution problems: Figure 1 shows two middleboxes: a NAT that translates private IPs to public IPs
and a firewall configured to block hosts H1 and H3 from
accessing specific public IPs. Ideally, we want administrators to configure firewall policies in terms of original
source IPs. Unfortunately, we do not know the privatepublic IP mappings that the NAT chooses dynamically;
i.e., the O RIGIN B INDING tenet is violated. Further, if
only traffic from H1 and H3 should be directed to the
firewall and the rest is allowed to pass through, an SDN
controller cannot install the correct forwarding rules at
switches S1 /S2 , as the NAT changes the packet headers;
i.e., PATHS F OLLOW P OLICY no longer holds.
Figure 1: Applying the blocking policy is challenging,
as the NAT hides the true packet sources.
Network diagnosis: In Figure 2, suppose the users of
hosts H1 and H3 complain about high network latency.
In order to debug and resolve this problem (e.g., determine if the middleboxes need to be scaled up [30]), the
network administrator may use a combination of hostlevel (e.g., X-Trace [29]) and network-level (e.g., [3])
logs to break down the delay for each request into persegment components as shown. Because O RIGIN B IND ING does not hold, it is difficult to correlate the logs to
track flows [50, 56].
Figure 2: Middlebox modifications make it difficult
to consistently correlate network logs for diagnosis.
Data-dependent policies: In Figure 3, the light IPS
checks simple features (e.g., headers); we want to route
suspicious packets to the heavy IPS, which runs deeper
534 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
analysis to determine if the packet is malicious. Such a
triggered architecture is quite common; e.g., rerouting
suspicious packets to dedicated packet scrubbers [12].
The problem here is that ensuring PATHS F OLLOW P OL ICY depends on the processing history; i.e., did the light
IPS flag a packet as suspicious? However, each switch
and middlebox can only make processing or forwarding
decisions with its link-local view.
$.408! !./0./!
Figure 3: S2 cannot decide if an incoming packet
should be sent to the heavy IPS or the server.
Policy violations due to middlebox actions: Figure 4
shows a proxy used in conjunction with an access control
device (ACL). Suppose we want to block H2 ’s access to
xyz.com. However, H2 may bypass the policy by accessing cached versions of xyz.com, thus evading the
ACL. The problem, therefore, is that middlebox actions
may violate PATHS F OLLOW P OLICY by introducing unforeseen paths. In this case, we may need to explicitly
route the cached responses to the ACL device as well.
Figure 4: Lack of visibility into the middlebox context (i.e., cache hit/miss in this example) makes policy
enforcement challenging.
Strawman Solutions
Next, we highlight why some seemingly natural strawman solutions fail to address the above problems. Due
to space constraints, we discuss only a few salient candidates; Table 1 summarizes their effectiveness in the
previously-presented examples.
Placement constraints: One way to ensure O RIGIN B INDING/PATHS F OLLOW P OLICY is to “hardwire” the
policy into the topology. In Figure 1, we could place the
firewall before the NAT. Similarly, for Figure 3 we could
connect the light IPS and the heavy IPS to S1 , and configure the light IPS to emit legitimate/suspicious packets
USENIX Association
(Figure 1)
(Figure 2)
Yes, if we
(e.g, [38,
(e.g., [52])
(e.g., [49])
(Figure 3)
If both
IPSes are
on S1 &
Light IPS
has 2 ports
Need IPS
(Figure 4)
Not with
Maybe, if shim is aware
Not accurate, lack of ground truth, and high overhead
Table 1: Analyzing strawman solutions vs. the motivating examples in §2.1.
on different output ports. S1 can then use the incoming port to determine if the packet should be sent to the
heavy IPS. This coupling between policy and topology,
however, violates the SDN philosophy of decoupling the
control logic from the data plane. Furthermore, this restricts flexibility to reroute under failures, load balance
across middleboxes, or customize policies for different
workloads [50].
Tunneling: Another option to ensure PATHS F OLLOWP OLICY is to set up tunneling rules, for example, using
MPLS or virtual circuit identifiers (VCIs). For instance,
we could tunnel packets from the “suspicious” output of
the light IPS to the heavy IPS in Figure 3. (Note that this
requires middleboxes to support tunnels.) Such topology/tunneling solutions may work for simple examples,
but they quickly break for more complex policies; e.g., if
there are more outputs from the light IPS. Note that even
by combining placement+tunneling, we cannot solve the
diagnosis problem in Figure 2, as it does not provide
Middlebox consolidation: At first glance, it may
seem that we can ensure PATHS F OLLOW P OLICY by running all middlebox functions on a consolidated platform [20, 52]. While consolidation provides other benefits (e.g., reduced hardware costs), it has several limitations. First, it requires a significant network infrastructure change. Second, it merely shifts the burden of
PATHS F OLLOW P OLICY to the internal routing “shim”
that routes packets between the modules. Finally, if the
individual modules are provided by different vendors, diagnosis and attribution is hard, as this shim cannot ensure
Flow correlation: Prior work attempts to heuristically correlate the payloads of the traffic entering and
leaving middleboxes to correlate flows [49]. However,
this approach can result in missed/false matches too of-
11th USENIX Symposium on Networked Systems Design and Implementation 535
$%&# '!(&!!""##
)"&*+,! &9;)!"
Figure 5: Figure 1 augmented to illustrate how tags
can solve the attribution problem.
ten to be useful for security applications [49]. Also,
such “reverse engineering” approaches fundamentally
lack ground truth. Finally, this process has high overhead, as multiple packets per flow need to be processed
at the controller in a stateful manner (e.g., when reassembling packet payloads).
As Table 1 shows, none of these strawman solutions
can address all of the motivating scenarios. In some
sense, each approach partially addresses some symptoms
of the violations of O RIGIN B INDING and PATHS F OL LOW P OLICY, but does not address the cause of the problem. Thus, despite the complexity they entail in terms of
topology hacks, routing, and middlebox and controller
upgrades, they have limited applicability and have fundamental correctness limitations.
FlowTags Overview
As we saw in the previous section, violating the O RIG IN B INDING and PATHS F OLLOW P OLICY tenets makes it
difficult to correctly implement several network management tasks. To address this problem, we propose the
FlowTags architecture. In this section, we highlight the
main intuition behind FlowTags, and then we show how
FlowTags extends the SDN paradigm.
FlowTags takes a first-principles approach to ensure
even in the presence of middlebox actions. Since the
middleboxes are in the best (and sometimes the only)
position to provide the relevant context (e.g., a proxy’s
cache hit/miss state or a NAT’s public-private IP mappings), we argue that middleboxes need to be extended
in order to be integrated into SDN frameworks.
Conceptually, middleboxes add tags to outgoing packets. These tags provide the missing bindings to ensure O RIGIN B INDING and the necessary processing context to ensure PATHS F OLLOW P OLICY. The tags are
then used in the data plane configuration of OpenFlow
switches and other downstream middleboxes.
To explain this high-level idea, let us revisit the example in Figure 1 and extend it with the relevant tags and ac-
5$#/6,(&"" "
.#!*(" 8!4,!+'9"" "
Figure 6: Interfaces between different components in
the FlowTags architecture.
tions as shown in Figure 5. We have three hosts H1 − H3
in an RFC1918 private address space; the administrator
wants to block the Internet access for H1 and H3 , and
allow H2 ’s packets to pass through without going to the
firewall. The controller (not shown) configures the NAT
to associate outgoing packets from H1 , H2 , and H3 with
the tags 1, 2, and 3, respectively, and adds these to prespecified header fields. (See §5.3). The controller configures the firewall so that it can decode the tags to map
the observed IP addresses (i.e., in “public” address space
using RFC1918 terminology) to the original hosts, thus
meeting the O RIGIN B INDING requirement. Similarly,
the controller configures the switches to allow packets
with tag 2 to pass through without going to the firewall,
thus meeting the PATHS F OLLOW P OLICY requirement.
As an added benefit, the administrator can configure firewall rules w.r.t. the original host IP addresses, without
needing to worry about the NAT-induced modifications.
This example highlights three key aspects of FlowTags. First, middleboxes (e.g., the NAT) are generators
of tags (as instructed by the controller). The packetprocessing actions of a FlowTags-enhanced middlebox
might entail adding the relevant tags into the packet
header. This is crucial for both O RIGIN B INDING and
PATHS F OLLOW P OLICY, depending on the middlebox.
Second, other middleboxes (e.g., the firewall) are consumers of tags, and their processing actions need to decode the tags. This is necessary for O RIGIN B INDING.
(In this simple example, each middlebox only generates
or only consumes tags. In general, however, a given middlebox could both consume and generate tags.)
Third, SDN-capable switches in the network use the
tags as part of their forwarding actions, in order to route
packets according to the controller’s intended policy, ensuring PATHS F OLLOW P OLICY holds.
Note that the FlowTags semantics apply in the context
of a single administrative domain. In the simple case,
we set tag bits to NULL on packets exiting the domain.3
3 More generally, if we have a domain hierarchy (e.g., “CS dept”
and “Physics dept” and “Univ” at a higher level), each sub-domain’s
egress switch can rewrite the tag to only capture higher-level semantics
(e.g, “CS” rather than “CS host A”), without revealing internal details.
536 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
This alleviates concerns that the tag bits may accidentally
leak proprietary topology or policy information. When
incoming packets arrive at an external interface, the gateway sets the tag bits appropriately (e.g., to ensure stateful
middlebox traversal) before forwarding the packet into
the domain.
Architecture and Interfaces
Next, we describe the interfaces between the controller,
middleboxes, switches, and the network administrator in
a FlowTags-enhanced SDN architecture.
Current SDN standards (e.g., OpenFlow [45]) define
the APIs between the controller and switches. As shown
in Figure 6, FlowTags adds three extensions to today’s
SDN approach:
1. FlowTags APIs between the controller and FlowTagsenhanced middleboxes, to programmatically configure their tag generation and consumption logic (§4).
2. FlowTags controller modules that configure the
tagging-related generation/consumption behavior of
the middleboxes, and the tag-related forwarding actions of SDN switches (§5).
3. FlowTags-enhanced middleboxes consume an incoming packet’s tags when processing the packet and
generate new tags based on the context (§6).
FlowTags requires neither new capabilities from SDN
switches, nor any direct interactions between middleboxes and switches. Switches continue to use traditional
SDN APIs such as OpenFlow. The only interaction between switches and middleboxes is indirect, via tags embedded inside the packet headers. We take this approach
for two reasons: (1) to allow switch and middlebox designs and their APIs to innovate independently; and (2)
to retain compatibility with existing SDN standards (e.g.,
OpenFlow). Embedding tags in the headers avoids the
need for each switch and middlebox to communicate
with the controller on every packet when making their
forwarding and processing decisions.
We retain existing configuration interfaces for customizing middlebox actions; e.g., vendor-specific languages or APIs to configure firewall/IDS rules. The advantage of FlowTags is that administrators can configure
these rules without having to worry about the impact of
intermediate middleboxes. For example, in the first scenario of §2.1, FlowTags allows the operator to specify
firewall rules with respect to the original source IPs. This
provides a cleaner mechanism, as the administrator does
not need to reason about the space of possible header values a middlebox may observe.4
4 Going forward, we want to configure the middlebox rules to ensure
the H IGH L EVEL NAMES as well [24].
USENIX Association
#" #"
*+%" "'
& ' ) &
34*5+,-. %'2"
Figure 7: Packet processing walkthrough for tag generation: 1. Tag Generation Query, 2. Tag Generation
Response, 3. Data Packet, 4. Packet-in Message, 5.
Modify Flow Entry Message, 6. Data Packet (to next
on-path switch).
8" !"
*+%" "'
12(3)*+, %%0
Figure 8: Packet processing walkthrough for tag consumption: 1. Data Packet, 2. Packet-in Message, 3.
Modify Flow Entry Message, 4. Data Packet, 5. Tag
Consumption Query, 6. Tag Consumption.
FlowTags APIs and Operation
Next, we walk through how a packet is processed in
a FlowTags-enhanced network, and describe the main
FlowTags APIs. For ease of presentation, we assume
each middlebox is connected to the rest of the network
via a switch. (FlowTags also works in a topology with
middleboxes directly chained together.) We restrict our
description to a reactive controller that responds to incoming packets, but proactive controllers are also possible.
For brevity, we only discuss the APIs pertaining to
packet processing. Analogous to the OpenFlow configuration APIs, we envision functions to obtain and set
FlowTags capabilities in middleboxes; e.g., which header
fields are used to encode the tag values (§5.3).
In general, the same middlebox can be both a generator and a consumer of tags. For clarity, we focus on these
two roles separately. We assume that a packet, before it
reaches any middlebox, starts with a NULL tag.
Middlebox tag generation, Figure 7: Before the
middlebox outputs a processed (and possibly modified)
packet, it sends the FT GENERATE QRY message to the
controller requesting a tag value to be added to the packet
(Step 1). As part of this query the middlebox provides
the relevant packet processing context: e.g., a proxy tells
the controller if this is a cached response; an IPS provides the processing verdict. The controller provides a
tag value via the FT GENERATE RSP response (Step 2).
(We defer tag semantics to the next section.)
Middlebox tag consumption, Figure 8: When a mid-
11th USENIX Symposium on Networked Systems Design and Implementation 537
dlebox receives a tag-carrying packet, it needs to “decode” this tag; e.g., an IDS needs the original IP 5tuple for scan detection. The middlebox sends the
FT CONSUME QRY message (Step 5) to the controller,
which then provides the necessary decoding rule for
mapping the tag via the FT CONSUME RSP message
(Step 6).
Switch actions: In Figure 7, when the switch receives
a packet from the middlebox with a tag (Step 3), it
queries the controller with the OFPT PACKET IN message (Step 4), and the controller provides a new flow table entry (Step 5). This determines the forwarding action; e.g., whether this packet should be routed toward
the heavy IPS in Figure 3. Similarly, when the switch
receives a packet in Figure 8 (Step 1), it requests a forwarding entry and the controller uses the tag to decide if
this packet needs to be forwarded to the middlebox.
Most types of middleboxes operate at an IP flow or
session granularity, and their dynamic modifications typically use a consistent header mapping for all packets of
a flow. Thus, analogous to OpenFlow, a middlebox needs
only once per flow. The middlebox stores the per-flow
tag rules locally, and subsequent packets in the same flow
can reuse the cached tag rules.
FlowTags Controller
In this section, we discuss how a FlowTags-enhanced
SDN controller can assign tags and tags-related “rules”
to middleboxes and switches. We begin with a policy abstraction (§5.1) that informs the semantics that tags need
to express (§5.2). Then, we discuss techniques to translate this solution into practical encodings (§5.3–§5.4). Finally, we outline the controller’s implementation (§5.5).
Dynamic Policy Graph
The input to the FlowTags controller is the policy that
the administrator wants to enforce w.r.t. middlebox actions (Figure 6). Prior work on middlebox policy focuses
on a static policy graph that maps a given traffic class
(e.g., as defined by network locations and flow header
fields) to a chain of middleboxes [30, 38, 49]. For instance, the administrator may specify that all outgoing
web traffic from location A to location B must go, in
order, through a firewall, an IDS, and a proxy. However, this static abstraction fails to capture the O RIGIN B INDING and PATHS F OLLOW P OLICY requirements in
the presence of traffic-dependent and dynamic middlebox actions. Thus, we propose the dynamic policy graph
(or DPG) abstraction.
A DPG is a directed graph with two types of nodes: (1)
In and Out nodes, and (2) logical middlebox nodes. In
and Out nodes represent network ingresses and egresses
(including “drop” nodes). Each logical middlebox rep-
(a) Dynamic policy routing
*+4 -.
*+ ,-.!6!
8799= %%&'()&('!
*+ ,-.!:
*+, -.!!
(b) Middlebox context
Figure 9: The DPGs for the examples in Figures 3 and
4. Rectangles with solid lines denote “Ingress” nodes
and with dotted lines denote “Egress” nodes. Circles denote logical middlebox functions. Each edge is
annotated with a {Class}; Context denoting the traffic
class and the processing context(s). All traffic is initialized as “{null};-”.
resents a type of middlebox function, such as “firewall.”
(For clarity, we restrict our discussion to “atomic” middlebox functions; a multi-function box will be represented using multiple nodes.) Each logical middlebox
node is given a configuration that governs its processing behavior for each traffic class (e.g., firewall rulesets
or IDS signatures). As discussed earlier, administrators
specify middlebox configurations in terms of the unmodified traffic entering the DPG, without worrying about
intermediate transformations.
Each edge in the DPG is annotated with the condition
m → m′ under which a packet needs to be steered from
node m to node m′ . This condition is defined in terms
of (1) the traffic class, and (2) the processing context of
node m, if applicable. Figure 9 shows two DPG snippets:
• Data-dependent policies: Figure 9a revisits the example in Figure 3. Here, we want all traffic to be
first processed by the light IPS. If the light IPS flags
a packet as suspicious, then it should be sent to the
heavy IPS. In this case, the edge connecting the light
IPS to the heavy IPS is labeled “*, Alarm”, where *
denotes the class of “any traffic,” and Alarm provides
the relevant processing history from the light IPS.
• Capturing effects of middlebox actions: Figure 9b
revisits the example in Figure 4, where we want to
apply an ACL only on host H2 ’s web requests. For
correct policy enforcement, the ACL must be applied
to both cached and uncached responses. Thus, both
“H2 , Hit” and “H2 , Miss” need to be on the Proxy-toACL edge. (For ease of visualization, we do not show
the policies applied to the responses coming from the
We currently assume that the administrator creates the
538 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
DPG based on domain knowledge. We discuss a mechanism to help administrators to generate DPGs in §10.
From DPG to Tag Semantics
The DPG representation helps us reason about the semantics we need to capture via tags to ensure O RIGIN B INDING and PATHS F OLLOW P OLICY.
Restoring O RIGIN B INDING: We can ensure O RIGIN B INDING if we are always able to map a packet to its
original IP 5-tuple OrigHdr as it traverses a DPG. Note
that having OrigHdr is a sufficient condition for O RIG IN B INDING: given the OrigHdr, any downstream middlebox or switch can conceptually implement the action
intended by a DPG. In some cases, such as per-flow diagnosis (Figure 2), mapping a packet to the OrigHdr might
be necessary. In other examples, a coarser identifier may
be enough; e.g., just srcIP in Figure 1.
To ensure
PATHS F OLLOW P OLICY, we essentially need to capture
the edge condition m → m′ . Recall that this condition
depends on (1) the traffic class and (2) the middlebox
context, denoted by C, from logical middlebox m (and
possibly previous logical middleboxes). Given that the
OrigHdr for O RIGIN B INDING provides the necessary
context to determine the traffic class, the only additional
required information on m → m′ is the context C.
If we assume (until §5.3) no constraints on the tag
identifier space, we can think of the controller as assigning a globally unique tag T to each “located packet”; i.e.,
a packet along with the edge on the DPG [51]. The controller maps the tag of each located packet to the information necessary for O RIGIN B INDING and PATHS F OL LOW P OLICY: T → ⟨OrigHdr, C⟩. Here, the OrigHdr
represents the original IP 5-tuple of this located packet
when it first enters the network (i.e., before any middlebox modifications) and C captures the processing context
of this located packet.
In the context of tag consumption from §4,
“dereference” tag T to obtain the OrigHdr. The middlebox can apply its processing logic based on the OrigHdr;
i.e., satisfying O RIGIN B INDING.
For tag generation at logical middlebox m,
FT GENERATE QRY provides as input to the controller:
(1) the necessary middlebox context to determine which
C will apply, and (2) the tag T of the incoming packet
that triggered this new packet to be generated. The
controller creates a new tag T ′ entry for this new located
packet and populates the entry T ′ → ⟨OrigHdr′ , C⟩ for
this new tag as follows. First, it uses OrigHdr (for the
input tag T ) to determine the value OrigHdr′ for T ′ .
In many cases (e.g., NAT), this is a simple copy. In
some cases (e.g., proxy response), the association has
to reverse the src/dst mappings in OrigHdr. Second, it
USENIX Association
associates the new tag T ′ with context C. The controller
instructs the middlebox, via FT GENERATE RSP, to
add T ′ to the packet header. Because T ′ is mapped to C,
it supports enforcement of PATHS F OLLOW P OLICY.
Encoding Tags in Headers
In practice, we need to embed the tag value in a finite number of packet-header bits. IPv6 has a 20-bit
Flow Label field, which seems ideal for this use (thus
answering the question “how should we use the flowlabel field?” [19]). For our current IPv4 prototype and
testbed, we used the 6-bit DS field (part of the 8-bit ToS),
which sufficed for our scenarios. To deploy FlowTags on
large-scale IPv4 networks, we would need to borrow bits
from fields that are not otherwise used. For example, if
VLANs are not used, we can use the 12-bit VLAN Identifier field. Or, if all traffic sets the DF (Don’t Fragment)
IP Flag, which is typical because of Path MTU Discovery, the 16-bit IP ID field is available.5
Next, we discuss how to use these bits as efficiently as
possible; §8 reports on some analysis of how many bits
might be needed in practice.
As discussed earlier, tags restore O RIGIN B INDING
and PATHS F OLLOW P OLICY. Conceptually, we need
to be able to distinguish every located packet—i.e.,
the combination of all flows and all possible paths in
the DPG. Thus, a simple upper bound on the number
of bits in each packet to distinguish between |Flows|
flows on |DPGPaths| processing paths is: log2 |Flows| +
log2 |DPGPaths|, where Flows is the set of IP flows (for
O RIGIN B INDING), and DPGPaths is the set of possible paths a packet could traverse in DPG (for PATHS F OLLOW P OLICY). However, this grows log-linearly in
the number of flows over time and the number of paths
(which could be exponential w.r.t. the graph size).
This motivates optimizations to reduce the number of
header bits necessary, which could include:
• Coarser tags: For many middlebox management
tasks, it may suffice to use a tag to identify the logical traffic class (e.g., “CS Dept User”) and the local
middlebox context (e.g., 1 bit for cache hit or miss or
1 bit for “suspicious”), rather than individual IP flows.
• Temporal reuse: We can reuse the tag assigned to a
flow after the flow expires; we can detect expiration
via explicit flow termination, or via timeouts [3, 45].
The controller tracks active tags and finds an unused
value for each new tag.
• Spatial reuse: To address O RIGIN B INDING, we only
need to ensure that the new tag does not conflict
with tags already assigned to currently active flows at
the middlebox to which this packet is destined. For
PATHS F OLLOW P OLICY, we need to: (1) capture the
5 IP ID isn’t part of the current OpenFlow spec; but it can be supported with support for flexible match options [9, 23].
11th USENIX Symposium on Networked Systems Design and Implementation 539
most recent edge on the DPG rather than the entire
path (i.e., reducing from |DPGPaths| to the node degree); and (2) ensure that the switches on the path have
no ambiguity in the forwarding decision w.r.t. other
active flows.
Putting it Together
Our current design is a reactive controller that responds to OFPT PACKET IN, FT CONSUME QRY, and
FT GENERATE QRY events from the switches and the
Initialization: Given an input DPG, we generate a data
plane realization DPGImpl; i.e., for each logical middlebox m, we need to identify candidate physical middlebox
instances, and for each edge in DPG, we find a switchlevel path between corresponding physical middleboxes.
This translation should also take into account considerations such as load balancing across middleboxes and resource constraints (e.g., switch TCAM and link capacity). While FlowTags is agnostic to the specific realization, we currently use SIMPLE [49], mostly because of
our familiarity with the system. (This procedure only
needs to run when the DPG itself changes or in case of a
network topology change. It does not run for each flow
Middlebox event handlers: For each physical middlebox instance PM i , the controller maintains two FlowTags
tables: CtrlInTagsTablei and the CtrlOutTagsTablei . The
CtrlInTagsTablei maintains the tags corresponding to all
incoming active flows into this middlebox using entries
{T → OrigHdr}. The CtrlOutTagsTablei tracks the tags
that need to be assigned to outgoing flows and maintains
a table of entries {⟨T, C⟩ → T ′ }, where T is the tag for
the incoming packet, C captures the relevant middlebox
context for this flow (e.g., cache hit/miss), and T ′ is the
output tag to be added. At bootstrap time, these structures are initialized to be empty.
The H ANDLE FT C ONSUME Q RY handler looks up
the entry for tag T in the CtrlInTagsTablei and sends
the mapping to PM i . As we will see in the next section, middleboxes keep these entries in a FlowTable-like
structure, to avoid look ups for subsequent packets. The
H ANDLE FT G ENERATE Q RY handler is slightly more
involved, as it needs the relevant middlebox context C.
Given C, the DPG, and the DPGImpl, the controller
identifies the next hop physical middlebox PM i′ for this
packet. It also determines a non-conflicting T ′ using the
logic from §5.3.
Switch and flow expiry handlers: The handlers for
OFPT PACKET IN are similar to traditional OpenFlow
handlers; the only exception is that we use the incoming tag to determine the forwarding entry. When a
flow expires, we trace the path this flow took and, for
each PM i , delete the entries in CtrlInTagsTablei and
CtrlOutTagsTablei , so that these tags can be repurposed.
We implement the FlowTags controller as a POX module [10]. The CtrlInTagsTablei and CtrlOutTagsTablei
are implemented as hash-maps. For memory efficiency
and fast look up of available tags, we maintain an auxiliary bitvector of the active tags for each middlebox and
switch interface; e.g., if we have 16-bit tags, we maintain
a 216 bit vector and choose the first available bit, using a
log-time algorithm [22]. We also implement simple optimizations to precompute shortest paths for every pair of
physical middleboxes.
FlowTags-enhanced Middleboxes
As discussed in the previous sections, FlowTags requires
middlebox support. We begin by discussing two candidate design choices for extending a middlebox to support
FlowTags. Then, we describe the conceptual operation of
a FlowTags-enhanced middlebox. We conclude this section by summarizing our experiences in extending five
software middleboxes.
Extending Middleboxes
We consider two possible ways to extend middlebox software to support FlowTags:
• Module modification: The first option is to modify
specific internal functions of the middlebox to consume and generate the tags. For instance, consider an
IDS with the scan detection module. Module modification entails patching this scan detection logic with
hooks to translate the incoming packet headers+tag to
the OrigHdr and to rewrite the scan detection logic to
use OrigHdr. Similarly, for generation, we modify the
output modules to provide the relevant context as part
• Packet rewriting: A second option is to add a
lightweight shim module that interposes on the incoming and outgoing packets to rewrite the packet
headers. For consumption, this means we modify
the packet headers so that the middlebox only sees
a packet with the true OrigHdr. For generation, this
means that the middlebox proceeds as-is and then the
shim adds the tag before the packet is sent out.
In both cases, the administrator sets up the middlebox configuration (e.g., IDS rules) as if there were no
packet modifications induced by the upstream middleboxes because FlowTags preserves the binding between
the packet’s modified header and the OrigHdr.
For consumption, we prefer packet rewriting because
it generalizes to the case where each middlebox has
multiple “consumer” modules; e.g., an IDS may apply
scan detection and signature-based rules. For generation,
540 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
t h in
Rewrite Pkt
with OrigHdr
Match in
ss packet
Name, Role
Squid [14],
Add new Tag
Figure 10: We choose a hybrid design where the
“consumption” side uses the packet rewriting and the
“generation” uses the module modification approach.
however, packet rewriting may not be sufficient, as the
shim may not have the necessary visibility into the middlebox context; e.g., in the proxy cache hit/miss case.
Thus, we use module modification in this case.
End-to-end view: Figure 10 shows a simplified view of
a FlowTags-enhanced middlebox. In general, consumption precedes generation. The reason is that the packet’s
current tag can affect the specific middlebox code paths,
and thus impacts the eventual outgoing tags.
Mirroring the controller’s CtrlInTagsTablei and
CtrlOutTagsTablei , each physical middlebox i
maintains the tag rules in the MBInTagsTablei and
MBOutTagsTablei . When a packet arrives, it first checks
if the tag value in the packet already matches an existing
tag-mapping rule in MBInTagsTablei . If there is a
match, we rewrite packet headers (see above) so that
the processing modules act as if they were operating
on OrigHdr. If there is a MBInTagsTablei miss, the
middlebox sends a FT CONSUME QRY, buffers the
packet locally, and waits for the controller’s response.
Note that the tags are logically propagated through
the processing contexts (not shown for clarity). For example, most middleboxes follow a connection-oriented
model with a data structure maintaining per-flow or perconnection state; we augment this structure to propagate
the tag value. Thus, we can causally relate an outgoing packet (e.g., a NAT-ed packet or a proxy cached response) to an incoming packet.
When a specific middlebox function or module
is about to send a packet forward, it checks the
MBOutTagsTablei to add the outgoing tag value. If there
is a miss, it sends the FT GENERATE QRY, providing
the necessary module-specific context and the tag (from
the connection data structure) for the incoming packet
that caused this outgoing packet to be generated.
Experiences in Extending Middleboxes
Given this high-level view, next we describe our experiences in modifying five software middleboxes that span
a broad spectrum of management functions. (Our choice
was admittedly constrained by the availability of the mid-
USENIX Association
Modified /
Total LOC
75 / 216K
Snort [13],
Balance [1],
PRADS [11],
iptables [6],
45 / 336K
60 / 2K
Key Modules
Client and Server
Side Connection,
Forward, Cache
Decode, Detect,
Client and Server
25 / 15K
55 / 42K
Conn Map
Table 2: Summary of the middleboxes we have added
FlowTags support to along with the number of lines
of code and the main modules to be updated. We use
a common library (≈ 250 lines) that implements routines for communicating to the controller.
dlebox source code.) Table 2 summarizes these middleboxes and the modifications necessary.
Our current approach to extend middleboxes is semimanual and involved a combination of call graph analysis [7, 17] and traffic injection and logging techniques [2,
4, 5, 15]. Based on these heuristics, we identify the suitable “chokepoints” to add the FlowTags logic. Developing techniques to automatically extend middleboxes is an
interesting direction for future work.
• Squid: Squid [14] is a popular proxy/cache. We modified the functions in charge of communicating with
the client, remote server, and those handling cache
lookup. We used the packet modification shim for
incoming packets, and applied module modification
to handle the possible packet output cases, based on
cache hit and miss events.
• Snort: Snort [13] is an IDS/IPS that provides many
functions—logging, packet inspection, packet filtering, and scan detection. Similar to Squid, we applied the packet rewriting step for tag consumption
and module modification for tag generation as follows. When a packet is processed and a “verdict”
(e.g., OK vs. alarm) is issued, the tag value is generated based on the type of the event (e.g., outcome of
a matched alert rule).
• Balance: Balance [1] is a TCP-level load balancer
that distributes incoming TCP connections over a
given a set of destinations (i.e., servers). In this case,
we simply read/write the tag bits in the header fields.
• PRADS: PRADS [11] is passive monitor that gathers
traffic information and infers what hosts and services
exist in the network. Since this is a passive device,
we only need the packet rewriting step to restore the
(modified) packet’s OrigHdr.
• NAT via iptables: We have registered appropriate
tagging functions with iptables [6] hook points, while
11th USENIX Symposium on Networked Systems Design and Implementation 541
Src / Time(s)
H1 / 0
H1 / 0.3
H1 / 0.6
H1 / 0.8
DPG path
L-IPS alarm
(a) In Figure 3, we configure Snort as the light IPS (L-IPS) to flag
hosts sending more than 3 packets/sec and send them to the heavy
Host / URL
H1 / Dept
H2 / CNN
H2 / Dept
H1 / CNN
DPG path
always allow
miss, allow
hit, drop
hit, allow
(b) In Figure 4, we use Squid as the proxy and Snort as the ACL and
block H2 ’s access to the Dept site.
Figure 11: Request trace snippets for validating the
example scenarios in Figure 3 and Figure 4.
it is configured as a source NAT. The goal is to maintain 5-tuple visibility via tagging. We added hooks
for tag consumption and tag generation into the PREROUTING and the POSTROUTING chains, which
are the input and output checkpoints, respectively.
Validation and Use Cases
Next, we describe how we can validate uses of FlowTags.
We also discuss how FlowTags can be an enabler for new
diagnostic and verification capabilities.
Testing: Checking if a network configuration correctly
implements the intended DPG is challenging—we need
to capture stateful middlebox semantics, reason about
timing implications (e.g., cache timeouts), and the impact of dynamic modifications. (Even advanced network
testing tools do not capture these effects [39, 57].) Automating this step is outside the scope of this paper, and
we use a semi-manual approach for our examples.
Given the DPG, we start from each ingress and enumerate all paths to all “egress” or “drop” nodes. For each
path, we manually compose a request trace that traverses
the required branch points; e.g., will we see a cache hit?
Then, we emulate this request trace in our small testbed
using Mininet [33]. (See §8 for details.) Since there is
no other traffic, we use per-interface logs to verify that
packets follow the intended path.
Figure 11 shows an example with one set of request sequences for each scenario in Figures 3 and 4. To emulate
Figure 3, we use Snort as the light IPS to flag any host
sending more than 3 packets/second as suspicious, and
direct such hosts’ traffic to the heavy IPS for deep packet
inspection (also Snort). Figure 11(a) shows the request
trace and the corresponding transitions it triggers.
To emulate Figure 4, we use Squid as the proxy and
Snort as the (web)ACL device. We want to route all H2 ’s
web requests through ACL and configure Snort to block
$ "
.:3";3":/" .:3";3"=/"
&'()'*"!+(,'-".!*,/" !'0(12,"!+(,'-".!*,"3"45!"#"$%!"673"&'(89"$%!"67/"
Figure 12: Disconnect between header-space analysis
and the intended processing semantics in Figure 3.
H2 ’s access to the department website. Figure 11(b)
shows the sequence of web requests to exercise different
DPG paths.
We have validated the other possible paths in these examples, and in other scenarios from §2. We do not show
these due to space constraints.
FlowTags-enabled diagnosis: We revisit the diagnosis
example of Figure 2, with twenty user requests flowing
through the NAT and LB. We simulated a simple “red
team-blue team” test. One student (“red”) synthetically
introduced a 100ms delay inside the NAT or LB code
for half the flows. The other student (“blue”) was responsible for attributing the delays. Because of dynamic
header rewriting, the “blue” team could not diagnose delays using packet logs. We repeated the experiment with
FlowTags-enhanced middleboxes. In this case, the FlowTags controller assigns a globally unique tag to each request. Thus, the “blue” team could successfully track
a flow through the network and identify the bottleneck
middlebox using the packet logs at each hop.
Extending verification tools: Verification tools such
as Header Space Analysis (HSA) [39] check correctness
(e.g., reachability) by modeling a network as the composition of header-processing functions. While this works
for traditional switches/routers, it fails for middleboxes,
as they operate at higher semantic layers. While a full
discussion of such tools is outside the scope of this paper, we present an example illustrating how FlowTags
addresses this issue.
Figure 12 extends the example in Figure 3 to show
both header-space annotations and DPG-based semantic annotations. Here, a header-space annotation (solid
boxes) of ⟨Src⟩ describes a packet from Src, so ⟨∗⟩ models a packet from any source. A DPG annotation (dashed
boxes) of ⟨Src, L, H⟩ describes a packet from Src for
which Light IPS returns L and Heavy IPS returns H, so
⟨∗, 0, ∗⟩ indicates a packet from any source that is flagged
by Light IPS as not OK; our policy wants such suspicious
packets to go via Heavy IPS, while ⟨∗, 1, ∗⟩ packets need
no further checking.
Recall from §2 that we cannot implement this policy,
542 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
in this topology, using existing mechanisms (i.e., without
FlowTags). What if we rewired the toplogy by adding
the (dashed) link Light IPS → Heavy IPS? Even with
this hardwired topology, tools like HSA incorrectly conclude that “all” packets exit the network (the output edge
is labeled ⟨∗⟩), because HSA models middleboxes as
“wildcard”-producing blackboxes [39].
FlowTags bridges the gap between “header space,” in
which verification tools operate, and “semantic space,” in
which the policy operates. Instead of modeling middleboxes as blackboxes, or reverse-engineering their functions, in FlowTags we treat them as functions operating
on tag bits in an (extended) header space. Then, we apply
HSA on this extended header space to reason if the network implements the reachability defined by the DPG.
Performance Evaluation
We frame questions regarding the performance and scalability of FlowTags:
• Q1: What overhead does support for FlowTags add to
middlebox processing?
• Q2: Is the FlowTags controller fast and scalable?
• Q3: What is the overhead of FlowTags over traditional
• Q4: How many tag bits do we need in practice?
Setup: For Q1 and Q2, we run each middlebox and
POX controller in isolation on a single core in a 32-core
2.6 Ghz Xeon server with 64 GB RAM. For Q3, we use
Mininet [33] on the same server, configured to use 24
cores and 32 GB RAM to model the network switches
and hosts. We augment Mininet with middleboxes running as external virtual appliances. Each middlebox runs
as a VM configured with 2GB RAM on one core. (We
can run at most 28 middlebox instances, due to the maximum number of PCI interfaces that can be plugged in
using KVM [8]). We emulate the example topologies
from §2, and larger PoP-level ISP topologies from RocketFuel [55]. Our default DPG has an average path length
of 3.
Q1 Middlebox overhead: We configure each middlebox to run with the default configuration. We vary the
offered load (up to 100 Mbps) and measure the perpacket processing latency. Overall, the overhead was low
(<1%) and independent of the offered load (not shown).
We also analyzed the additional memory and CPU usage
using atop; it was < 0.5% across all experiments (not
Q2 Controller scalability: Table 3 shows the running
time for the H ANDLE FT G ENERATE Q RY. (This is the
most complex FlowTags processing step; other functions
take negligible time.) The time is linear as a function
of topology size with the baseline algorithms, but almost
constant using the optimization to pre-compute reacha-
USENIX Association
Topology (#nodes)
Abilene (11)
Geant (22)
Telstra (44)
Sprint (52)
Verizon (70)
AT&T (115)
Baseline (ms)
Optimized (ms)
Table 3: Time to run H ANDLE FT G ENERATE Q RY.
Figure 13: Breakdown of flow processing time in different topologies (annotated with #nodes).
bility information. This implies that a single-thread POX
≈ 35K middlebox queries
controller can handle 0.028ms
per second (more than three times larger than the peak
number of flows per second reported in [24]).
We also varied the DPG complexity along three axes:
number of nodes, node degrees, and distance between adjacent DPG nodes in terms of number of switches. With
route pre-computation, the controller processing time is
independent of the DPG complexity (not shown).
Q3 End-to-end overhead: Figure 13 shows the breakdown of different components of the flow setup time in a
FlowTags-enhanced network (i.e., mirroring the steps in
Figure 7) for different Rocketfuel topologies. Since our
goal is to compare the FlowTags vs. SDN operations, we
do not show round-trip times to the controller here, as it
is deployment-specific [35].6 Since all values are close
to the average, we do not show error bars. We can see
that the FlowTags operations add negligible overhead.
In fact, the middlebox tag processing is so small that it
might be hard to see in the figure.
We also measure the reduction in TCP throughput a
flow experiences in a FlowTags-enhanced network, compared to a traditional SDN network with middleboxes
(but without FlowTags). We vary two parameters: (1)
controller RTT and (2) the number of packets per flow.
As we can see in Table 4, except for very small flows (2
packets), the throughput reduction is <4%.
Q4 Number of tag bits: To analyze the benefits of
spatial and temporal reuse, we consider the worst case,
where we want to diagnose each IP flow. We use
packet traces from CAIDA (Chicago and San Jose traces,
2013 [16]) and a flow-level enterprise trace [18]. We sim6 FlowTags adds 1 more RTT per middlebox, but this can be avoided
by pre-fetching rules for the switches and middleboxes.
11th USENIX Symposium on Networked Systems Design and Implementation 543
Flow size (#packets)
Reduction in throughput (%)
1ms RTT
10ms RTT
20ms RTT
Table 4: Reduction in TCP throughput with FlowTags relative to a pure SDN network.
(spatial, temporal)
(No spatial, 30 sec)
(Spatial, 30 sec)
(Spatial, 10 sec)
(Spatial, 5 sec)
(Spatial, 1 sec)
Number of bits
CAIDA trace
Enterprise trace
Table 5: Effect of spatial and temporal reuse of tags.
ulate the traces across the RocketFuel topologies, using a
gravity model to map flows to ingress/egress nodes [55].
Table 5 shows the number of bits necessary with
different reuse strategies, on the AT&T topology from
RocketFuel.7 The results are similar across other topologies (not shown). We see that temporal reuse offers the
most reduction. Spatial reuse helps only a little; this is
because with a gravity-model workload, there is typically
a “hotspot” with many concurrent flows. To put this in
the context of §5.3, using the (Spatial, 1 sec) configuration, tags can fit in the IPv6 FlowLabel, and would fit in
the IPv4 IP ID field.
Related Work
We have already discussed several candidate solutions
and tools for verification and diagnosis (e.g., [34, 39]).
Here, we focus on other classes of related work.
Middlebox policy routing: Prior work has focused on
orthogonal aspects of policy enforcement such as middlebox load balancing (e.g., [42, 49]) or compact data
plane strategies (e.g,. [27]). While these are candidates
for translating the DPG to a DPGImpl (§5), they do not
provide reliable mechanisms to address dynamic middlebox actions.
Middlebox-SDN integration: OpenMB [31] focuses
on exposing the internal state (e.g., cache contents and
connection state) of middleboxes to enable (virtual) middlebox migration and recovery. This requires significantly more instrumentation and vendor support compared to FlowTags, which only requires externally relevant mappings. Stratos [30] and Slick [21] focus on using SDN to dynamically instantiate new middlebox modules in response to workload changes. The functionality
these provide is orthogonal to FlowTags.
7 Even though the number of flows varies across traces, they require
the same number of bits, as the values of ceil(log2 (# f lows)) are the
Tag-based solutions: Tagging is widely used to implement Layer2/3 functions, such as MPLS labels or
virtual circuit identifiers (VCIs). In the SDN context, tags have been used to avoid loops [49], reduce
FlowTable sizes [27], or provide virtualized network
views [46]. Tags in FlowTags capture higher-layer semantics to address O RIGIN B INDING and PATHS F OL LOW P OLICY. Unlike these Layer2/3 mechanisms where
switches are generators and consumers of tags, FlowTags
middleboxes generate and consume tags, and switches
are consumers.
Tracing and provenance: The idea of flow tracking
has parallels in the systems (e.g., tracing [29]), databases
(e.g., provenance [58]), and security (e.g., taint tracking [47, 48]) literature. Our specific contribution is to
use flow tracking for integrating middleboxes into SDNcapable networks.
Conclusions and Future Work
The dynamic, traffic-dependent, and hidden actions of
middleboxes make it hard to systematically enforce and
verify network-wide policies, and to do network diagnosis. We are not alone in recognizing the significance
of this problem—others, including the recent IETF network service chaining working group, mirror several of
our concerns [37, 43, 50].
The insight behind FlowTags is that the crux of these
problems lies in violation of two key SDN tenets—
by middlebox actions. We argue that middleboxes are
in the best (and possibly the only) vantage point to restore these tenets, and make a case for minimally extending middleboxes to provide the necessary context,
via tags embedded inside packet headers. We design new
SDN APIs and controller modules to configure this tagrelated behavior. We showed a scalable proof-of-concept
controller, and the viability of adding FlowTags support,
with minimal changes, to five canonical middleboxes.
We also demonstrated that the overhead of FlowTags is
comparable to traditional SDN mechanisms.
We believe that there are three natural directions for
future work: automating DPG generation via model
refinement techniques (e.g., [25]); automating middlebox extension using appropriate programming-languages
techniques; and, performing holistic testing of the network while accounting for switches and middleboxes.
We would like to thank our shepherd Ben Zhao and the
NSDI reviewers for their feedback. This work was supported in part by grant number N00014-13-1-0048 from
the Office of Naval Research and by Intel Labs’ University Research Office.
544 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
[1] Balance.
[2] Bit-Twist.
[3] Cisco systems netflow services export version 9. RFC
[4] httperf.
[5] iperf. https://code.google.com/p/iperf/.
[6] iptables.
[7] KCachegrind.
[8] KVM.
[9] Openflow
[10] POX Controller.
[11] PRADS.
[12] Prolexic. www.prolexic.com.
[13] Snort. http://www.snort.org/.
[14] Squid. http://www.squid-cache.org/.
[15] tcpdump. http://www.tcpdump.org/.
[16] The Cooperative Association for Internet Data Analysis
(caida). http://www.caida.org/.
[17] Valgrind. http://www.valgrind.org/.
[18] Vast Challenge.
[19] S. Amante, B. Carpenter, S. Jiang, and J. Rajahalme.
Ipv6 flow label update. http://rmv6tf.org/wpcontent/uploads/2012/11/rmv6tf-flowlabel11.pdf.
[20] J. W. Anderson, R. Braud, R. Kapoor, G. Porter, and
A. Vahdat. xOMB: extensible open middleboxes with
commodity servers. In Proc. ANCS, 2012.
[21] B. Anwer, T. Benson, N. Feamster, D. Levin, and J. Rexford. A Slick Control Plane for Network Middleboxes. In
Proc. ONS, research track, 2012.
[22] G. Banga and J. Mogul. Scalable kernel performance for
Internet servers under realistic loads. In Proc. USENIX
ATC, 1998.
[23] P. Bosshar, G. Gibb, H.-S. Kim, G. Varghese, N. McKeown, M. Izzard, F. Mujica, and M. Horowitz. Forwarding
Metamorphosis: Fast Programmable Match-Action Processing in Hardware for SDN. In Proc. SIGCOMM, 2013.
USENIX Association
[24] M. Casado, M. J. Freedman, J. Pettit, J. Luo, N. McKeown, and S. Shenker. Ethane: Taking control of the enterprise. In Proc. SIGCOMM, 2007.
[25] E. Clarke, O. Grumberg, S. Jha, Y. Lu, and H. Veith.
Counterexample-guided abstraction refinement. In Proc.
CAV, 2000.
[26] C. Dixon, H. Uppal, V. Brajkovic, D. Brandon, T. Anderson, and A. Krishnamurthy. ETTM: a scalable fault
tolerant network manager. In Proc. NSDI, 2011.
[27] L. Erran Li, Z. M. Mao, and J. Rexford. CellSDN:
Software-defined cellular networks. In Techinical Report,
Princeton University.
[28] S. K. Fayazbakhsh, V. Sekar, M. Yu, and J. C. Mogul.
FlowTags: Enforcing network-wide policies in the presence of dynamic middlebox actions. In Proc. HotSDN,
[29] R. Fonseca, G. Porter, R. H. Katz, S. Shenker, and I. Stoica. X-Trace: a pervasive network tracing framework. In
Proc. NSDI, 2007.
[30] A. Gember, A. Krishnamurthy, S. S. John, R. Grandl,
X. Gao, A. Anand, T. Benson, A. Akella, and V. Sekar.
Stratos: A network-aware orchestration layer for middleboxes in the cloud. CoRR, abs/1305.0209, 2013.
[31] A. Gember, P. Prabhu, Z. Ghadiyali, and A. Akella. Toward software-defined middlebox networking. In Proc.
HotNets-XI, 2012.
[32] N. Gude, T. Koponen, J. Pettit, B. Pfaff, M. Casado,
N. McKeown, and S. Shenker. NOX: towards an operating system for networks. In CCR, 2008.
[33] N. Handigol, B. Heller, V. Jeyakumar, B. Lantz, and
N. McKeown. Reproducible network experiments using
container-based emulation. In Proc. CoNext, 2012.
[34] N. Handigol, B. Heller, V. Jeyakumar, D. Maziéres, and
N. McKeown. Where is the debugger for my softwaredefined network? In Proc. HotSDN, 2012.
[35] B. Heller, R. Sherwood, and N. McKeown. The Controller Placement Problem. In Proc. HotSDN, 2012.
[36] X. Jin, L. Erran Li, L. Vanbever, and J. Rexford. Softcell:
Scalable and flexible cellular core network architecture.
In Proc. CoNext, 2013.
[37] W. John, K. Pentikousis, G. Agapiou, E. Jacob, M. Kind,
A. Manzalini, F. Risso, D. Staessens, R. Steinert, and
C. Meirosu. Research directions in network service chaining. In Proc. IEEE SDN4FNS, 2013.
[38] D. A. Joseph, A. Tavakoli, and I. Stoica. A policy-aware
switching layer for data centers. In Proc. SIGCOMM,
[39] P. Kazemian, G. Varghese, and N. McKeown. Header
space analysis: static checking for networks. In Proc.
NSDI, 2012.
[40] A. Khurshid, W. Zhou, M. Caesar, and P. B. Godfrey. Veriflow: verifying network-wide invariants in real time. In
Proc. NSDI, 2013.
11th USENIX Symposium on Networked Systems Design and Implementation 545
[41] J. Lee, J. Tourrilhes, P. Sharma, and S. Banerjee. No More
Middlebox: Integrate Processing into Network. In Proc.
SIGCOMM posters, 2010.
[42] L. Li, V. Liaghat, H. Zhao, M. Hajiaghay, D. Li, G. Wilfong, Y. Yang, and C. Guo. PACE: Policy-Aware Application Cloud Embedding. In Proc. IEEE INFOCOM,
[56] W. Wu, G. Wang, A. Akella, and A. Shaikh. Virtual network diagnosis as a service. In Proc. SoCC, 2013.
[57] H. Zeng, P. Kazemian, G. Varghese, and N. McKeown.
Automatic test packet generation. In Proc. CoNext, 2012.
[58] W. Zhou, M. Sherr, T. Tao, X. Li, B. T. Loo, and Y. Mao.
Efficient querying and maintenance of network provenance at internet-scale. In Proc. SIGMOD, 2010.
[43] L. MacVittie.
Service chaining and unintended
[44] N. McKeown.
Mind the Gap: SIGCOMM’12
Keynote. http://www.youtube.com/watch?v=
[45] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar,
L. Peterson, J. Rexford, S. Shenker, and J. Turner. OpenFlow: enabling innovation in campus networks. CCR,
March 2008.
[46] C. Monsanto, J. Reich, N. Foster, J. Rexford, and
D. Walker. Composing Software Defined Networks. In
Proc. NSDI, 2013.
[47] Y. Mundada, A. Ramachandran, M. B. Tariq, and
N. Feamster. Practical Data-Leak Prevention for Legacy
Applications in Enterprise Networks. Technical Report
http://hdl.handle.net/1853/36612, 2011.
[48] J. Newsome and D. Song. Dynamic Taint Analysis for
Automatic Detection, Analysis, and Signature Generation of Exploits on Commodity Software. In Proc. NDSS,
[49] Z. Qazi, C. Tu, L. Chiang, R. Miao, V. Sekar, and
M. Yu. SIMPLE-fying Middlebox Policy Enforcement
Using SDN. In Proc. SIGCOMM, 2013.
[50] P. Quinn, J. Guichard, S. Kumar, P. Agarwal, R. Manur, A. Chauhan, N. Leyman, M. Boucadir, C. Jacquenet,
M. Smith, N. Yadav, T. Nadeau, K. Gray, B. Mcconnell,
and K. Glavin. Network service chaining problem statement. http://tools.ietf.org/html/draftquinn-nsc-problem-statement-03.
[51] M. Reitblatt, N. Foster, J. Rexford, C. Schlesinger, and
D. Walker. Abstractions for network update. In Proc.
SIGCOMM, 2012.
[52] V. Sekar, N. Egi, S. Ratnasamy, M. K. Reiter, and G. Shi.
Design and implementation of a consolidated middlebox
architecture. In Proc. NSDI, 2012.
[53] V. Sekar, S. Ratnasamy, M. K. Reiter, N. Egi, and G. Shi.
The middlebox manifesto: enabling innovation in middlebox deployment. In Proc. HotNets, 2011.
[54] J. Sherry, S. Hasan, C. Scott, A. Krishnamurthy, S. Ratnasamy, and V. Sekar. Making middleboxes someone
else’s problem: Network processing as a cloud service.
In Proc. SIGCOMM, 2012.
[55] N. Spring, R. Mahajan, and D. Wetherall. Measuring
ISP Topologies with Rocketfuel. In Proc. of ACM SIGCOMM, 2002.
546 11th USENIX Symposium on Networked Systems Design and Implementation
USENIX Association
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF