Joint Hop-by-hop and Receiver-Driven Interest Control Protocol for Content-Centric Networks

Massimo Gallo (Orange labs), joint work with:

Giovanna Carofiglio (Bell Labs), Luca Muscariello (Orange labs). CCNxCon 2012 - September 13th, 2012 - Sophia Antipolis

1.  Transport issues in CCN

2.  Receiver-driven Interest control

3.  Hop-by-hop Interest control: design and analysis

4.  Performance Evaluation

5.  Hop-by-hop Interest control: benefits

6.  Conclusions

AGENDA

CCN Transport - KEY ASPECTS UNIQUE ENDPOINT AT THE RECEIVER No connection instantiation, multiple senders for the same content retrieval, unknown a priori at the receiver

MULTIPLE SOURCES Data can be retrieved by multiple repositories, but also intermediate caches

PULL-BASED POINT TO MULTIPOINT RETRIEVAL Interests for the same content retrieval can be forwarded in a point-to-multipoint fashion: better throughput, better traffic load balancing if multiple sources in parallel

USER

CACHE

CACHE

CACHE

CACHE

CACHE

CACHE

CACHE

CACHE

REPO

Interests

Data

@ the receiver

CCN receiver is the unique flow endpoint:

•  knows application requirements, end-to-end round trip delay per packet retrieval, receiver buffer (flow size)

•  is the best place where to control content retrieval over multiple paths

@ network nodes

CCN nodes know Interest/Data rates of flows (identified by content name) per interface:

•  Interest/Data traverse the same nodes in opposite directions

•  provide hop-by-hop Interest control

•  handle bursty traffic and react faster

CCN congestion control mechanisms

Receiver-driven Interest Control Protocol (ICP) DESIGN

•  One Interest per Data packet, in the order decided by the application

•  Window-based Additive Increase Multiplicative Decrease (AIMD):

!  W is increased by !/W at each Data packet reception

!  W is decreased by "W at each timer expiration (a timer is set at the receiver for each Interest sent out) and no more than once in a time interval equal to the timer duration

•  Adaptive timer expiration value, ! , based on RTT estimates over a history of samples

! reflects the average virtual RTT and may be associated to a path

Hop-by-Hop Interest Control - OBJECTIVES

Interest control at network node: •  anticipate congestion detection by monitoring Interest/Data rate

•  trigger rate reduction via Interest shaping before timer expiration at the receiver.

•  control PIT entries according to Upstream resources

Basic Idea If Interest rate> fair rate at a given interface, one can queue and delay Interests at output interface to reduce Data queuing at the bottleneck

Hop-by-Hop Interest Control - DESIGN

•  One virtual queue per flow at each output interface, identified by the content name •  One credit counter per virtual queue initialized to B Data bytes that the flow can transmit with no additional

delay •  The counter is:

•  incremented at the estimated fair rate •  decremented by forwarded Interests

Hop-by-Hop Interest Control - DESIGN (cont’d)

bottlenecked ?

Yes No

send Interest (no additional delay)

queueInterest in a drop tail FIFO served

at !i(t)

@ interest arrival (after CS/PIT/FIB lookup)

•  is the rate of non-bottlenecked flows (total rate of non shaped flows, counting the size of the corresponding Data packets)

•  is the # of bottlenecked flows (# of non empty queues) "

Shaping rate:

Shaping algorithm:

Hop-by-Hop Interest Control - ANALYSIS

Main Result

We prove that HR-ICP is stable and converges to the max-min fair rate of ICP, where the shaping queue Qs

i (t) replaces Qi(t) in the ICP system

Interest Window

Delivery Time [s]

Throughput [Mbps]

Losses [%]

W [pkts] With HbH

W/o HbH

With HbH

W/o HbH

With HbH

W/o HbH

2 2.42 2.42 16.30 16.30 0 0

10 1.00 1.00 39.70 39.60 0 0

15 1.00 2.08 39.60 19.20 0 11.20

20 1.00 1.90 39.60 20.90 0 15.30

ICP 1.00 1.00 39.80 39.80 0 0

Performance Evaluation Impact of Interest Control on User Performance

•  Hop-by-hop Interest Shaping is not enough •  Interest shaping reduces Data packet losses

•  Implementation oh Interest shaping mechanism in CCNPL-Sim (C++ event driven simulator for the CCN architecture)

• Two hops network, single content retrieval. ICP vs constant window, w or w/o Hop by Hop Interest control:

Results

•  HR-ICP queues Interests before the bottleneck link (Q1s not Q2, Q2 is zero )

•  ICP flows almost not affected w H2H, get the fair rate, W slightly reduced

•  Greedy CBR flow looses (CBR rate - fair rate)

Performance Evaluation The benefits of HR-ICP over ICP

Results

Repository

C1= 100Mbps C2= 40 Mbps

Three flows: •  2 ICP (t1=0s, t1=0.5s) •  CBR (t=1s - avg 40Mbps).

Performance Evaluation Prioritization of real-time and delay sensitive traffic

Three flows: •  ICP 1 bottlenecked at 4 (t=0s), •  ICP 2 bottlenecked at 2 (t=0.5s), •  4 Interests in batch every 10ms (t=0.5s, avg 5Mbps - peak 100

Mbps). Results: •  HR-ICP queues Interests before

bottleneck, •  ICP flows almost not affected, while

the new flow gets priority along the request path.

Repository

C1= 100Mbps C2= 40 Mbps C3= 100 Mbps

Repository

C4= 20 Mbps

HBH Interest shaping - benefits

"  Interest not Data Control

"  Early Congestion Detection

"  Protection from misbehaving receivers

"  Scalability/Feasibility

"  Delay-sensitive flows protection

"  No interest losses

"  Additional traffic control opportunities

Conclusions

We show that #  Hop-by-hop Interest shaping enhances rate and congestion control

performance #  Compared to alternative solutions for CCN Interest control, our proposal brings additional benefits due to:

$  the coupling with a rate/fairness optimal receiver control, $  the positioning of Interest shaper at output interfaces, $  the shaping mechanism based on Data max-min fair rate.

Future works definition of traffic control mechanisms for the management of a multipath communication, coupled with an Interest forwarding policy.

Questions

CCN Simulator soon available at: http://perso.rd.francetelecom.fr/muscariello/sim

http://code.google.com/p/ccnpl-sim