Anda di halaman 1dari 7

LUO_LAYOUT.

qxp_Layout 1 5/15/14 1:55 PM Page 4

CoLoR: An Information-Centric Internet


Architecture for Innovations
Hongbin Luo, Zhe Chen, Jianbo Cui, Hongke Zhang, Moshe Zukerman, and Chunming Qiao

Abstract
In this article, we describe an information-centric Internet architecture called CoLoR
that couples service location and inter-domain routing while decoupling them from
forwarding. Preliminary results based on implementation and analysis show that
CoLoR is promising since it satisfies many requirements of the future Internet, includ-
ing being information-centric, encouraging innovations, and providing efficient sup-
port for mobility, multicast, multi-homing, and middleboxes.

T he Internet has achieved enormous success since its


inception. However, despite the rapid increase in the
number of users and applications, it faces many chal-
lenges, including poor scalability and security. Unfor-
tunately, these architectural deficiencies cannot be remedied
by incremental changes to the current Internet architecture.
Therefore, in recent years there has been an increasing
4. Encouraging innovations: The future Internet architecture
should allow each network to use its preferred network
architecture and routing mechanism so that different net-
work technologies can be simultaneously deployed and con-
tested, thus encouraging innovations.
5. Enhanced security: The current Internet employs a default-
on model [9] wherein any host is able to send packets to a
amount of effort in developing clean slate redesigns of the remote host, which makes the current Internet vulnerable
Internet architecture, aimed at rectifying one or more of these to distributed denial of service attacks. Therefore, the
problems through non-incremental changes [15, references future Internet should offer receivers the ability to control
therein]. For example, the authors of [24] aimed at designing incoming traffic, especially to refuse unwanted traffic.
an information-centric Internet. The authors of [6, 7] argued 6. Enhanced scalability: The future Internet should provide
that a future Internet should be designed to encourage com- better routing scalability than the current Internet. In par-
petition and innovations. Raychaudhuri et al. [8] aimed at ticular, the routing table size should be significantly smaller
designing an Internet that provides better support for mobility than that in the current Internet.
and security. 7. Ease of traffic matrix estimation: It is difficult to estimate
In this article, we also pursue a clean slate Internet archi- traffic matrices in the current Internet. However, since traf-
tecture, focusing on the following attributes, which constitute fic matrices are critical input to many aspects of network
the design objectives of the new architecture: management such as traffic engineering and network provi-
1. Information-centric: While the current Internet is host-cen- sioning, the future Internet should make it easy to accurate-
tric (i.e., packets are sent to a specific hosts IP address), it ly estimate traffic matrices in real time.
is being used overwhelmingly for information retrieval. 8. Deployability: Although we aim for a clean slate design, the
Accordingly, there is an increasing consensus that the new architecture should be deployable without incurring
future Internet should be information-centric. That is, con- significant cost.
tents should be assigned names that are location-indepen- While many Internet architectures [18,10] have been pro-
dent and application-independent. posed in the past, none of them has achieved all the above men-
2. Efficient support for mobility: With the rapid increase in the tioned design objectives. In this article, we propose a future
number of mobile devices, the future Internet architecture Internet architecture that has all the above attributes. It is based
should efficiently support mobility. on Coupling service Location and inter-domain routing while
3. Efficient support for multi-homing: In multi-homing, a host decoupling them from forwarding. Accordingly, we name it
(or network) is simultaneously attached to multiple net- CoLoR for short. CoLoR is created by synthesizing many ideas
works. It is difficult to efficiently support multi-homing in borrowed from existing work in a cohesive manner. For exam-
the current Internet because it causes a serious routing scal- ple, we borrow the idea of using self-certifying node identifiers
ability problem. A future Internet architecture is expected from MobilityFirst [8] and AIP [11]. From DONA [2], we bor-
to support multi-homing more efficiently. row the idea of naming data with self-certifying identifiers to
achieve data authenticity and persistency. From FII [6] and
Pathlet routing [12], we borrow the idea of using paths (or path
Hongbin Luo, Zhe Chen, Jianbo Cui, and Hongke Zhang are with Beijing segments) for inter-domain routing. Similarly, we borrow the
Jiaotong University. name-based routing mechanism from DONA [2] and CCN [4].
In addition to designing CoLoR, we also build a proof-of-
Moshe Zukerman is with the City University of Hong Kong. concept prototype of CoLoR and present results from experi-
ments carried on the prototype to demonstrate CoLoRs
Chunming Qiao is with the State University of New York at Buffalo. feasibility in small-scale deployments.

4 0890-8044/14/$25.00 2014 IEEE IEEE Network May/June 2014


LUO_LAYOUT.qxp_Layout 1 5/15/14 1:55 PM Page 5

The remainder of the article is organized as follows. We domains, as long as the PID is unique in both domains.
first describe the CoLoR architecture. We then report our Unlike [6, 12], where the PID of the path between two
proof-of-concept prototype of CoLoR and explain how domains is advertised throughout the Internet, the PID of the
CoLoR satisfies our design objectives. Finally, we conclude path between two domains is known only by the two domains
the article and outline future work. for the purpose of enhancing security. In addition, we antici-
pate that the PIDs between two domains are randomly chosen
so that it is difficult for attackers to correctly guess them, thus
The CoLoR Architecture further improving security. While the length of PIDs is 32 bits
We now describe the design details of the CoLoR architec- in our implementation, optimizing this length is nontrivial as
ture. We begin with the network topology, naming, and rout- it requires trading off efficiency and security. Intuitively, a
ing. We then describe how to register services, how to locate longer length leads to higher bandwidth consumption, but
services, and how to determine inter-domain paths for ser- makes it harder for attackers to correctly guess the PIDs
vices. Finally, we describe how data packets are forwarded between two domains.
and discuss how to cache services in CoLoR.
Routing
Network Topology In the current Internet, both inter-domain and intra-domain
As in the current Internet, CoLoR assumes that a future routing rely on IP addresses. As a result, without global coor-
Internet will continue to be organized around domains that dination, it is difficult for domains to deploy novel networking
have domain-level provider/customer/peer relationships. Every technologies. To overcome this drawback, CoLoR separates
domain has a logical resource manager (with possibly many intra-domain routing from inter-domain routing. Domains in
physical incarnations), which maintains a registration table CoLoR are free to adopt their preferred network architec-
that stores the reachability information of contents. For brevi- tures and intra-domain routing mechanisms, such as IPv4 and
ty, we denote a resource manager (RM) associated with a IPv6, thus encouraging innovations of network technologies.
domain X by RMX. The relationships between RMs of domains This may make it difficult to efficiently support host mobility
are defined by domain-level relationships: RMX is the provider, since a mobile device needs to adapt to different network
customer, or peer (or, alternatively, parent, child, or peer) of technologies. However, there are at least two approaches for
RMY if domain X is the provider, customer, or peer of domain dealing with this issue. First, as advocated in [6], we can
Y, respectively. Note that the RM in CoLoR is similar to the require that each intra-domain networking technology sup-
resource handler (RH) in DONA. However, as shown later, ports a method of bootstrapping so that a host can learn
an RM in CoLoR needs to choose path identifiers if there are about the environment and determine where its local
multiple path identifiers between two domains. On the con- resources are. Second, we may split every domain into a core
trary, an RH in DONA does not need to do this. and an edge, with border routers connecting the edge and the
In addition, we assume that every node in CoLoR knows core. End hosts are located at the edge and connected to the
the location of its local RM through some local configura- core through border routers by a common approach. This has
tions, similar to the way in which an end host knows its local the benefit that border routers and other routers in the core
domain name service (DNS) server in the Internet today. can use any technology chosen by the domain.
Inter-domain routing in CoLoR relies on paths negotiated
Naming by domains. In particular, two domains can negotiate a set of
Unlike the current Internet, which uses only one namespace paths depending on their local policies. The two endpoints of
(i.e., IP address), CoLoR uses two global namespaces: service each path are located at each of the two domains. For exam-
identifiers (SIDs) and node identifiers (NIDs). To be informa- ple, domains D5 and D6 in Fig. 1 negotiate three paths, P1, P2,
tion-centric, CoLoR assigns every content a persistent and and P 3, as indicated by the bold dashed lines in Fig. 1. The
unique SID that is application-independent and location-inde- two endpoints of P 1 are R 8 and R 7 located in D 5 and D 6 ,
pendent. By persistent, we mean that the name for a content respectively. For a path that begins at a domain, the RM and
remains valid as long as the underlying content is available. By the routers (which connect hosts or other domains) in the
unique, we mean that two different pieces of content should domain maintains the paths endpoint located at the domain
have different SIDs. As in DONA [2], CoLoR uses self-certi- and the domain identifier at which the other endpoint is locat-
fying SIDs to achieve content authenticity. Meanwhile, NIDs ed. For example, the inter-domain routing table of R 6 is
are used to name nodes in the network. Every node in the shown in the upper left corner of Fig. 1.
network has a unique and persistent NID that is independent
of the location of the node. As in [8, 11], NIDs in CoLoR are Service Registration
flat self-certifying labels, so authenticating a node does not In CoLoR, users send out GET messages to obtain their
require an external authority such as ICANN [8], thus improv- desired contents, and the network routes the GET messages
ing security and privacy. Note that the current Internet uses to the nearest nodes holding the desired contents. This
hierarchical IP addresses, since its global routing heavily relies requires a mechanism to propagate the reachability informa-
on the aggregability of IP addresses to achieve scalability. In tion of SIDs. To achieve this, the RMs in CoLoR comprise an
contrast, NIDs in CoLoR are not used for global routing, but overlay to propagate the reachability information of SIDs.
for authentication. The mechanism used to propagate the reachability informa-
CoLoR also uses two local namespaces: intra-domain rout- tion of SIDs may be similar to that of DONA [2] or CCN [4].
ing locators and path identifiers (PIDs). Intra-domain routing In this article, we use a mechanism similar to that in DONA.
locators are used for routing within a domain. Domains in In particular, a node holding a copy of a content registers the
CoLoR may use different intra-domain routing locators. For contents SID to the RM in the same domain if it is autho-
example, Domain A may use IPv6 addresses for local routing, rized to do so. Depending on the local policy, the RM may
while Domain B may use IPv4 addresses. PIDs are used to also register the content to its parent/peer RMs. Figure 1
name (virtual) paths between domains. As in [6, 12], two illustrates the registration process. When RM 3 receives the
domains could negotiate the number of paths between them. registration message for the SID from RM1, it stores an entry
For a given path, its PID is also negotiated between the two for the SID into its registration table. In addition, it also sends

IEEE Network May/June 2014 5


LUO_LAYOUT.qxp_Layout 1 5/15/14 1:55 PM Page 6

PID R D Pref
RM6
P1 R7 D5 60
D6 RM finds at least one entry for the SID, it chooses
P2 R7 D5 30 R7 an entry based on its local policy (e.g., traffic engi-
neering). If the node contained in the entry is in the
P3 R7 D5 10
same domain as the RM, the RM directly forwards
P5 R6 D3 100 R6 the GET message to the node using the routing
P5 mechanism in that domain. If the node is in another
P1 (iii)
R8 domain, the RM chooses a path toward that domain
(3) RM5
R5 (iv) based on its local policy. Afterward, the RM appends
D3 RM3 the PID of the chosen path onto the GET message
(v) P2
R2 P3 and forwards the GET message to the node.
(5) Figure 1 illustrates how a GET message is for-
R10 D5 warded from a client to a node hosting the desired
R4 R9 content represented by an SID, assuming that the
P6 (2)
(ii) content has been registered as described earlier.
RM1 P7 R11 When node C wants the content, it sends a GET
R1 P4
R2 RM2 message to its local RM (i.e., RM 4 ). When RM 4
receives the GET message, it cannot find an entry
(vi) for the SID. Accordingly, it appends path P4 onto the
R12 (i)
D1 D2 GET message and sends the resulting message to its
D4
A (1) B C RM4 parent RM (i.e., RM5). Similarly, RM5 appends path
(4) P1 (or P2 or P3, depending on D5s local policy) onto
the received GET message and sends the resulting
(i) SID1 C (vi) SID1 C P4 P1 P5 P6 message to its parent RM (i.e., RM 6 ). When RM 6
receives the GET message, it finds an entry for the
SID in its registration table. Since RM6 is not in the
(ii) SID1 C P4 (v) SID1 C P4 P1 P5 P6
same domain with RM3, it appends path P5 onto the
received GET message and sends the resulting mes-
(iii) SID1 C P4 P1 (iv) SID1 C P4 P1 P5
sage to RM3. Since RM3 can find two entries for the
SID, it chooses the entry pointing to domain D 1
Figure 1. Illustration of network topology and service registration in based on its local policy, appends path P 6 onto the
CoLoR. received GET message, and sends the GET message
to RM 1 . When RM 1 receives the GET message, it
finds that a local node (i.e., node A) holds the
a registration message to its parent RM (i.e., RM 6 ). When desired content. Accordingly, it sends the GET message to
RM6 receives the registration message, it stores an entry for node A. For clarity, we show the GET messages in each step
the SID in its registration table. When RM3 receives a regis- at the bottom of Fig. 1.
tration message for the SID from RM2, it finds that there is an
entry for the SID, but the registration message comes from Packet Forwarding
RM2 instead of RM1. Thus, it adds an entry for the SID to its Once the source node storing the desired content receives the
registration table. Note that now RM3 does not send a regis- GET message, it knows the SID of the desired content, the
tration message for the SID to RM6, since it has registered the clients NID, and the inter-domain paths to be used to reach
SID already. the client. As a result, the source node encapsulates the pack-
ets of the content with a header that carries the clients NID,
Service Location and Inter-Domain Routing the SID, and the PIDs.
Recall that, for security, the PIDs between two domains are Recall that for every path that begins at a domain, the
not advertised to other domains. Accordingly, a content routers and the RM in the domain maintain the endpoint of
source cannot know the inter-domain paths to a content con- the path located at the domain and the domain identifier at
sumer, so it cannot send data packets to the consumer. To which the other endpoint is located. Accordingly, the source
address this issue, CoLoR determines the inter-domain paths node then sends the packets to the border router associated
during the service location process. In particular, when a with the first PID (or path) by using the routing mechanism of
client wants to obtain a content represented by an SID, it its domain. The packet is then forwarded by the border router
sends out a GET message to its local RM. The GET message to the other end of the first path, assuming that the routers
should contain the SID and the clients NID. If the RM can- along the path are able to forward packets by using PIDs.
not find an entry for the SID, it first chooses a parent RM When the other end of the first path receives the packet, it
based on its local policy and then chooses a PID destined to strips out the PID of the first path, obtains the PID of the
the corresponding parent domain. Note that when domain A second path, and forwards the packet to the border router
has multiple PIDs to domain B, it can choose which PID to toward the client. In this way, the packets of the desired con-
use based on its local policy. However, the best PID for tent are sent to the client.
domain A may not be the best one for domain B, since the We now illustrate packet forwarding in CoLoR using Fig. 2,
path from the ingress node (or the node holding the SID) to by assuming that domains D1 and D3 in Fig. 1 use IP and mul-
the egress node corresponding to the chosen PID in domain B tiprotocol label switching (MPLS) for local routing, respec-
may be congested. Therefore, it is necessary for the two tively. When node A receives the GET message, it first
domains to negotiate with each other in order to ensure the encapsulates the data with a header that contains SID1, the
best performance for both domains. However, how to nego- clients NID (i.e., C), and the PIDs. Since domain D1 uses IP
tiate is beyond the scope of this article. for local routing, node A then encapsulates the packet with an
The RM then appends the PID of the chosen path onto the outer IP header the source and destination addresses of which
GET message and sends it to the chosen parent RM. If the are IP 2 and IP 1 , respectively, as illustrated in Fig. 2a. The

6 IEEE Network May/June 2014


LUO_LAYOUT.qxp_Layout 1 5/15/14 1:55 PM Page 7

P5
R1 (d)
and PID 2 between node R 2 in D 1 and node R 5 in
(b) R5
IP1 D 2 . All nodes are implemented using CLICK to
RM1 (a) D3 (c)
D1 perform their functionalities according to the
RM3
A
designs described earlier. Since we do not use
IP2 P6 R2
R4 routing protocols to disseminate routing informa-
tion, we manually configure routing tables of all
nodes; the IP addresses of each node are shown in
(a) IP1 IP2 P6 P5 P1 P4 C SID1 Data Fig. 3. Similarly, the PID tables of all nodes are
manually configured for simplicity.
(b) P6 P5 P1 P4 C SID1 Data
Experimental Results
We now present the experimental results obtained
(c) MPLS LSP1 P5 P1 P4 C SID1 Data
with the prototype. In our experimentation, we let
every server register 50,000 SIDs with its local RM,
(d) P5 P1 P4 C SID1 Data which then forwards the registration messages to
the other RM. Thus, every RM stores 100,000 SID
Figure 2. Illustration of packet forwarding in CoLoR. entries. To reduce the traffic load of the client, the
content corresponding to every SID is 1024 bytes,
so each GET message is replied to by one data
packet. The client then sends GET messages for
packet will be forwarded to node R 1 by IP routing. R 1 then SIDs randomly chosen from the 100,000 SIDs to RM 1. The
strips out the outer IP header and sends the packet to path intervals between subsequent GET messages follow the expo-
P6, as illustrated in Fig. 2. When node R2 receives the packet, nential distribution the mean of which equals 1 ms, and we
it encapsulates the packet with an outer MPLS header that run the experiment for 1000 s. In addition, for contents pro-
contains the LSP between node R 2 and node R 5 (i.e., label vided by server 2, RM 1 chooses path PID 1 with 70 percent
switching path 1 LSP1), since domain D 3 uses MPLS for probability and path PID 2 with 30 percent probability. We
local routing, as illustrated by Fig. 2c. When R5 receives the find that packets are forwarded as expected, which not only
packet, it strips out the MPLS header and sends the packet to verifies the correctness of the CoLoR design, but also demon-
path P5, as illustrated by Fig. 2. In this way, the packet is ulti- strates CoLoRs feasibility in this small-scale deployment.
mately sent to the client, C. Although the performance from this proof-of-concept imple-
From the above descriptions, one can see that CoLoR cou- mentation of CoLoR is not indicative of the performance in a
ples service location with inter-domain routing, but decouples large-scale deployment, studying the delay for processing
them from forwarding. GET messages at RM1 enables us to gain better insight into
CoLoRs performance.
Content Caching Figure 4 shows the distribution of the delay for processing
Since contents are assigned persistent and unique names, they GET messages at RM1. From this figure, we observe that the
can be cached by routers during the packet forwarding pro- processing delay ranges from 0.35 to 0.65 ms, with its mean
cess. In addition, to make the best usage of cached contents, a and median being 0.529 and 0.426 ms, respectively. Note that
router should register its cached contents to its local RM so RMs use sequential search to look up an SID among 100,000
that the local RM can forward subsequent GET messages for entries. Thus, the processing delay could be significantly
the cached contents to the router. In this way, CoLoR not reduced if modern search algorithms are used. For example,
only improves resource utilization efficiency, but also reduces [2] assumes that an RH is able to process 40,000 GET mes-
the delay for users to obtain contents. In addition,
the caching and service registration primitives
make CoLoR efficient in supporting multicast and
mobility. RM1 RM2
IP5: 1000.4000::1
IP1: 10.0.2.1 IP6: 1000.4000::2
Implementation IP2: 10.0.2.2
IP3: 10.0.5.1
10.0.3.2 2000:1000::1 IP7: 1000.5000::1
2000:1000::2 IP8: 1000.5000::2
To validate the proposed design and also evaluate IP4: 10.0.5.2 10.0.3.1 IP9: 2000.2000::2
IP0: 10.0.4.1
its performance, we built a Linux-based prototype IP10: 10.0.4.2
of CoLoR using CLICK [13], which is an open R2 PID2 R5 2000:2000::1
source software platform. In this section, we first IP0
describe the prototype, followed by evaluation IP2 IP8
D1 D2
IP10 IP9
results from the prototype. IP1 IP7
IP4 IP6
Prototype Design IP3
R3 PID1 R4 IP5 R6
Figure 3 shows the network topology of the proto- R1
type, which consists of two domains and 11 nodes, 10.0.6.2 2000:3000::1
10.0.1.2
including one client, two servers, two RMs, and six
routers. The 11 nodes are placed in two domains, 10.0.1.1 10.0.6.1 2000:3000::2
D 1 and D 2; and each domain has one server, one
RM, and three routers, as shown in Fig. 3. The
Client Server1 Server2
client is placed at D1. D1 and D2 use IPv4 address-
es and IPv6 addresses for local routing, respective-
ly. The two domains are connected by two paths:
PID 1 between node R 3 in D 1 and node R 4 in D 2 , Figure 3. The topology of the prototype.

IEEE Network May/June 2014 7


LUO_LAYOUT.qxp_Layout 1 5/15/14 1:55 PM Page 8

0.14
Mean = 529 s
Median = 426 s
0.12
These results demonstrate the accuracy of CoLoR
in estimating traffic matrices.

0.1 Security
The empirical probability

Security is achieved in CoLoR as follows. First,


the use of self-certifying NIDs makes it efficient
0.08
and effective to authenticate the identity of a
node without relying on an external authority.
0.06 Second, the inter-domain PIDs are determined
during the service location process. This in turn
makes it convenient to deploy mechanisms such
0.04 as TVA [14] to defend against denial of service
attacks, even if an attacker knows the PIDs
0.02 toward the victim. Third, packet forwarding in
CoLoR is intrinsically based on loose source
routing. However, since the PIDs are not adver-
0 tised throughout the Internet, it is almost impos-
350 400 450 500 550 600 650
The delay for processing GET messages at an RM1 (s)
sible for an attacker to correctly guess the PIDs
between two domains. Therefore, a malicious
node can know the PIDs toward a victim only in
Figure 4. The delay in processing GET messages at RM1. one of the following two cases. In one case, the
GET message from the victim is forwarded to the
malicious node. In the other case, the malicious
sages, which implies that the delay for processing a GET mes- node is compromised by an attacker and learns the PIDs
sage will be 25 ms. toward the victim from the attacker. In this case, however, the
malicious node must be in the same domain with the attacker
since the PIDs from different domains to the victim are differ-
Other Features of CoLoR ent, which in turn makes it easy to trace the attacker. In both
From the above descriptions, it is clear that CoLoR is an cases, we can use DoS defense mechanisms such as TVA to
information-centric Internet architecture and naturally prevent the victim from being attacked. A more detailed anal-
encourages innovations of network technologies. In this sec- ysis of CoLoRs security can be found in [15].
tion, we describe how CoLoR satisfies the remaining design
objectives. Mobility
Mobility in CoLoR follows naturally from its registration and
Traffic Matrices Estimation the GET primitives. A mobile node holding a content can
CoLoR makes it easy to accurately estimate traffic matrices in unregister from a previous location and register with its new
real time. Indeed, when an ingress border router receives a location. Once the necessary registration state is installed at
data packet, it knows the egress border router by querying the the new location, subsequent GET messages for the content
inter-domain routing table. Accordingly, to estimate the traffic will be routed to the new location. On the other hand, a
from the ingress border router to the egress border router, mobile node requesting a content could re-send a GET mes-
the ingress border router only needs to: sage for the content, and the GET message is routed to a
Maintain a counter that records the number of bytes (or nearby copy of the desired content, possibly cached at a near-
packets) between the two border routers by router, due to the caching mechanism discussed above. In
Count the number of bytes of the packet (or the packet) addition, the mobility within a domain could be addressed by
when forwarding a packet the mobility management approach of the domain.
Note that these tasks can be done in real time, and an accu- For ongoing communications, when a consumer host roams
rate measure of the traffic matrix can be obtained. To esti- from domain A to a neighboring domain B, domain A may
mate the traffic matrices of a domain, we only need to let the maintain a mobility anchor, and the packets to the host are
routers in the domain: first routed to the anchor. When the anchor receives the pack-
Estimate the traffic matrix from the router to other routers ets destined to the host, it then appends the PID between
in the domain domain A and domain B, and forwards the packets to domain
Report the traffic matrices B in which the host is located. On the other hand, when a
Figure 5 shows the estimated traffic matrices from node R3 source host roams from domain A to a neighboring domain B,
to node R1, from node R2 to node R1, from node R6 to node the host is able to know the PID(s) from domain A to domain
R5, and from node R6 to node R4 in our prototype. From Fig. B. Accordingly, when host A sends out a packet, it should first
5, we observe that the traffic from node R3 to R1 is about 850 append one PID from domain A to domain B onto the outer
packets/s, the traffic from node R 6 to node R 4 is about 350 header of the packets. This way, the packets are first sent
packets/s, the traffic from node R 2 to node R 1 is about 150 from domain B to domain A and then forwarded to the corre-
packets/s, and the traffic from node R 6 to node R 5 is about sponding node.
150 packets/s. Note that these numbers match with the actual
traffic matrices very well. For example, the actual traffic from Multi-Homing
node R 6 to node R 5 is 150 packets/s because the number of CoLoR supports multi-homing efficiently without affecting
GET messages sent to server 2 is 500/s, and 30 percent of the the scalability of the Internet. First, when a network A adds a
replying data packets are sent to PID2. Similarly, the actual connection to another network B, the PIDs between the two
traffic from node R 6 to node R 4 is 350 packets because the networks are not advertised throughout the Internet. There-
number of GET messages sent to server 2 is 500/s, and 70 fore, the inter-domain routing tables of other networks are
percent of the replying data packets are being sent to PID1. not affected. Second, it is true that the RM in network A may

8 IEEE Network May/June 2014


LUO_LAYOUT.qxp_Layout 1 5/15/14 1:55 PM Page 9

1000

900
register some SIDs to the RM in network B. How-
ever, this does not affect the scalability of the whole 800
Internet since this does not increase the number of R3R1
SIDs with which tier 1 networks need to deal. 700
In addition, since packet forwarding across

Rate (packets/s)
domains is based on paths, a multi-homed domain 600
R6R4
is able to control its incoming traffic by forwarding
500
GET messages to different providers. For example,
for load balancing, the RM may forward some 400
GET messages to one provider domain, but for-
ward the rest of the GET messages to the other 300
R6R5
provider domain. Furthermore, the per-request
nature of GET messages makes it easy for multi- 200
homed domains/hosts to realize fine-grained traffic
100
engineering over incoming traffic, thus making the
R2R1
best use of multi-homing. We have evaluated the 0
effect of CoLoRs ability to efficiently support traf- 0 100 200 300 400 500 600 700 800 900 1000
fic engineering in our prototype. Figure 6 shows the Time (s)
number of packets on paths PID1 and PID2. From
this figure, we observe that the number of packets Figure 5. The traffic matrices estimated from the prototype.
traversing path PID1 is about 350 packets/s. On the
other hand, the number of packets traversing path
PID 2 is about 150 packets/s. This implies that
CoLoR efficiently supports traffic engineering 500
since, as described before, RM1 chooses path PID1
with 70 percent probability and path PID2 with 30 450
percent probability when it forwards GET messages
400
for contents provided by server 2.
Note also that when a host has two interfaces 350
connecting to two different networks, A and B, the
Rate (packets/s)

host can decide by itself to which network a GET 300


message should be forwarded so as to make the PID1
best usage of multi-homing. However, when the 250
host has decided to forward a GET message to net-
200
work A, which has two provider networks, the RM
in network A can decide where to forward the GET 150
message based on its local policy.
100
Scalability PID2
The scalability of CoLoR is determined by two 50
aspects: the routing table size and the huge number
0
of SIDs with which a tier 1 autonomous system 0 100 200 300 400 500 600 700 800 900 1000
(AS) must handle. Time (s)
1) Routing table size: A router in CoLoR main-
tains two routing tables: an intra-domain routing
Figure 6. The number of packets on paths PID1 and PID2.
table and an inter-domain routing table. Since
domains in CoLoR are free to adopt their pre-
ferred routing mechanisms, we anticipate that the
intra-domain routing table size in a domain is well within the pointed out that it is possible to design a distributed-hash-
domains control. For the inter-domain routing table size, a table-based name resolution system for 1016 flat SIDs with an
router in a domain maintains the paths that connect the average resolution delay below 100 ms. Therefore, CoLoR is
domain to the domains parent/customer/peer domains. From feasible at the scale of the current Internet.
[16], a domain (i.e., AS174) had at most 4060 neighbors as of
August 1, 2013. Therefore, even if two domains have 10 PIDs, Deployment
the inter-domain routing table size is at most 40,600, which is We note that CoLoR may not be incrementally deployable as
significantly less than the current global routing table size end hosts need to be updated in order to send registration
(more than 480,000 as of August 1, 2013). and GET messages. Nevertheless, the above analysis and
2) Dealing with SIDs: In [2], it is shown that resource han- implementation results demonstrate CoLoRs ease of deploya-
dlers (RHs) in DONA are capable of processing REGISTER bility. Since CoLoR allows each network to use their chosen
and FIND messages if DONA is deployed at the scale of the network architectures and routing mechanisms, existing net-
current Internet. Given the fact that DONA only caches con- works are only required to update their border routers and
tents on RHs, but CoLoR can cache contents on all nodes build an RM in order to accommodate CoLoR. Since the
within a domain, we anticipate that a domain in CoLoR will number of border routers in a network is relatively small,
cache more contents than DONA, thus reducing the number most routers are not required to be updated, thus significantly
of messages that an RM needs to process. Accordingly, the reducing the cost of deploying CoLoR.
processing overhead of RMs in CoLoR should be less than For CoLoR to interoperate with the current Internet, we
that of RHs in DONA. Furthermore, the authors in [17] considered two cases. In the first case, a single domain

IEEE Network May/June 2014 9


LUO_LAYOUT.qxp_Layout 1 5/15/14 1:55 PM Page 10

employs the CoLoR architecture. In this case, we employ a [6] T. Koponen et al., Architecting for Innovation, ACM SIGCOMM CCR,
proxy at the border of the domain and assign the domain a vol. 41, no. 3, July 2011, pp. 2426.
[7] J. Chuang, Loci of Competition for Future Internet Architecture, IEEE
unique domain name, say, www.CoLoR.com. We then let the Commun. Mag., vol. 49, no. 7, July 2011, pp. 3843.
proxy register contents in the domain to the current Internet. [8] D. Raychaudhuri, K. Nagaraja, and A. Venkataramani, MobilityFirst: A
Specifically, given a piece of content with a name SID, we Robust And Trustworthy Mobility-Centric Architecture for the Future Inter-
assign it a URL, www.CoLoR.com/SID. In this way, the nodes net, ACM Mobile Computing and Commun. Rev. , vol. 16, no. 3, July
2012, pp. 213.
in the current Internet can obtain the content hosted in the [9] H. Ballani et al., Off by Default!, Proc. ACM HotNets05, Nov. 2005,
domain. Similarly, when the nodes in the domain want to College Park, MD.
obtain the contents in the current Internet, we let the nodes [10] H. Luo et al., An Incrementally Deployable Network Architecture to Support
send the requests to the proxy, which then obtains the con- Both Data-Centric and Host-Centric Services, IEEE Network, Nov. 2013.
[11] D. G. Andersen et al., Accountable Internet Protocol (AIP), Proc. SIG-
tents from the current Internet. In the second case, several COMM 08, Aug. 2008, Seattle, WA.
isolated domains employ CoLoR. In this case, each of the [12] P. B. Godfrey et al. , Pathlet Routing, Proc. SIGCOMM 09 , Aug.
domains can interoperate with the current Internet, as stated 2009, Barcelona, Spain.
above. On the other hand, these domains can communicate [13] E. Kohler et al., The Click Modular Router, ACM Trans. Computer Sys-
tems, vol. 18, no. 3, Aug. 2000, pp. 26397.
with each other by using IP tunnels between border routers, [14] X. Yang, D. Wetherall, and T. Anderson, TVA: A DoS-Limiting Network Archi-
and each IP tunnel is viewed as a virtual path represented by tecture, IEEE/ACM Trans. Net., vol. 16, no. 6, Dec. 2008, pp. 126780.
a PID in CoLoR. [15] Z. Chen et al., Security Analysis of a Future Internet Architecture, Proc.
8th Wksp. Secure Network Protocolsa, Oct. 2013, Gottingen, Germany.
[16] BGP Peer Report, http://bgp.he.net/report/peers.
Summary and Future Work [17] C. Dannewitz, M. DAmbrosio, and V. Vercellone, Hierarchical DHT-
Based Name Resolution for information-Centric Networks, Computer
In this article, we have presented CoLoR, an information-cen- Commun., vol. 36, no. 7, Apr. 2013, pp. 73649.
tric Internet architecture that satisfies many attractive features
of the future Internet, such as being information-centric, effi- Biographies
cient support for mobility, traffic engineering, multi-homing, HONGBIN LUO [SM06, M07] (hbluo@bjtu.edu.cn) is a professor at the School
of Electronic and Information Engineering, Beijing Jiaotong University. He has
enhanced security, and innovation encouragement. We have authored more than 50 peer-reviewed papers in leading journals, such as
implemented CoLoR in a prototype and verified its feasibility. IEEE/ACM Transactions on Networking , and conference proceedings. His
While we have shown that CoLoR is promising and feasi- research interests are in the areas of routing, Internet architecture, and optical
ble, there are still many open questions. For example, one networking. He is an Editor of IEEE Communications Letters and has been a
Technical Program Committee member of many conferences.
open problem is how should the RM in a domain coordinate
nodes in the domain in order to cache as many contents as ZHE CHEN (11120069@bjtu.edu.cn) is a Ph.D. student at the School of Elec-
possible. Another question is, in the case where many tronic and Information Engineering, Beijing Jiaotong University. His research
provider/peer RMs could provide a desired content, how interests are in the areas of information-centric networking and software
defined networking.
should an RM forward a GET message for the desired con-
tent. In addition, since the RM in a domain processes regis- JIANBO CUI (12120053@bjtu.edu.cn) is a Masters student at the School of
tration/GET messages, an important question is how to secure Electronic and Information Engineering, Beijing Jiaotong University. His
the RM and make it robust. Furthermore, how should two research interests are in the areas of routing and software defined net-
working.
neighbor domains negotiate paths? Should two domains re-
negotiate paths to prevent attackers from collecting the paths HONGKE ZHANG (hkzhang@bjtu.edu.cn) is a professor of the School of Elec-
between neighboring domains? If so, how should they be re- tronic and Information Engineering, Beijing Jiaotong University. He has pub-
negotiated, and when? Similarly, the hop-by-hop inter-domain lished more than 100 research papers in the areas of communications,
computer networks, and information theory. He is the author of eight books
path selection in CoLoR means that the routing path will fol- written in Chinese and is Chief Scientist of a National Basic Research Pro-
low the resolution tree, which may cause path stretch. Hence, gram of China.
it is also interesting to investigate how large the path stretch is
and how to avoid/reduce the path stretch. These and other MOSHE ZUKERMAN [M87, S91, F07] (m.zu@cityu.edu.hk) received his Ph.D.
from the University of California, Los Angeles, in 1985. In 19861997, he
issues need to be addressed as future work. was with Telstra Research Laboratories and during 19972008 with the Uni-
versity of Melbourne, Australia. Since 2008, he has been with the City Univer-
Acknowledgments sity of Hong Kong. He has also served on various journal editorial boards
We thank the anonymous reviewers for their invaluable com- and conference Technical Program Committees. His research interests include
performance analysis of telecommunications networks, network design, and
ments that have improved the article. This work was support- traffic engineering.
ed in part by the 973 Program of China under Grant No.
2013CB329100, in part by NSFC under Grant Nos. 61271200, CHUNMING QIAO [f09] (qiao@computer.org) is a professor with the State Uni-
and 61232017, in part by the Ph.D. Programs Foundation of versity of New York, Buffalo, where he directs the Laboratory for Advanced
Network Design, Analysis, and Research, which conducts cutting-edge
the Ministry of Education of China under Grant No. research work on optical networks, wireless and mobile networks, survivable
20130009110014, in part by NCET under Grant No. NCET- networks, and TCP/IP technologies. He has published more than 80 and 120
12-0767, and in part by the Fundamental Research Funds for papers in leading technical journals and conference proceedings, respectively.
the Central Universities under Grant No. 2014JBM011. His pioneering research on optical Internet, in particular the optical burst
switching paradigm, is internationally acclaimed. In addition, his work on inte-
grated cellular and ad hoc relaying systems, started in 1999, is recognized
References as the harbinger for todays push toward the convergence between heteroge-
[1] H. Balakrishnan et al., A Layered Naming Architecture for the Internet, neous wireless technologies, and has been featured in BusinessWeek and
Proc. SIGCOMM 04, Aug. 2004, Portland, OR. Wireless Europe, as well as on the websites of New Scientists and CBC. He
[2] T. Koponen et al., A Data-Oriented (and Beyond) Network Architecture, has given several keynotes, tutorials, and invited talks on the above research
Proc. SIGCOMM 07, Aug. 2007, Kyoto, Japan. topics. He has chaired and co-chaired a dozen international conferences and
[3] E. Nordstrom et al., Serval: An End-Host Stack for Service-Centric Net- workshops. He has been an editor of IEEE Transactions on Networking and a
working, Proc. 9th USENIX Symp. Networked System Design and Imple- Guest Editor for several IEEE Journal on Selected Areas in Communications
mentation, Apr. 2012, San Jose, CA. issues, as well as Chair of the IEEE Technical Committee on High Speed Net-
[4] V. Jacobson et al., Networking Named Content, Proc. ACM CoNEXT works. Currently, he is on the editorial boards of several journals and maga-
09, Dec. 2009, Rome, Italy. zines, including IEEE Transactions on Parallel and Distributed Systems, and
[5] G. Xylomenos et al. , A Survery of Information-Centric Networking chairs the IEEE Subcommittee on Integrated Fiber and Wireless Technologies
Research, to appear, IEEE Commun. Surveys and Turorials. (FiWi), which he founded.

10 IEEE Network May/June 2014

Anda mungkin juga menyukai