IEEE Network - March / April 2017 - page 50

IEEE Network • March/April 2017
48
0890-8044/17/$25.00 © 2017 IEEE
A
bstract
Tremendous traffic on the Internet exhibits
ON/OFF patterns. In the past, the generation of
such ON/OFF traffic was mainly attributed to TCP
and application characteristics. Nowadays, how-
ever, much traffic is often intentionally shaped
into an ON/OFF pattern for achieving certain
optimized system performance or resulting from
characteristics of some new applications. In this
article, we present a comprehensive survey on
the ON/OFF traffic shaping in the current Inter-
net. We first introduce the motivations behind
ON/OFF traffic shaping in different fields based
on different system optimization objectives and
application characteristics. We then introduce
the impacts of ON/OFF traffic on packet drop
probability, real-time applications, and also its
interaction with TCP’s congestion control mech-
anism. We further present a survey on existing
mechanisms for mitigating these impacts in var-
ious aspects. We introduce how each of these
mechanisms works, and discuss their merits and
deficiencies. Finally, we conclude this article with
some future directions for ON/OFF traffic shaping
in the Internet.
I
ntroduction
Nowadays, tremendous traffic on the Internet
exhibits ON/OFF characteristics. ON/OFF traffic
means the existence of packet transmissions for
a short time interval followed by being idle for
another time interval, which repeats over time.
An interval with data transmissions is called an
ON interval, and one without data transmissions is
called an OFF interval. In the past, the generation
of ON/OFF traffic was mainly attributed to the
following two reasons. One is the characteristics
of applications (e.g., VoIP and web browsing). For
example, web browsing traffic exhibits an obvious
ON/OFF pattern that comes from users’ browsing
behaviors. In this application, the duration of an
ON state equals the time required for download-
ing a webpage. The other is mechanics of TCP
including slow start, loss recovery, acknowledg-
ment (ACK) compression, unused congestion win-
dow space, and bursty application programming
interface (API) calls (see the literature review in
[1] for a survey). In the past 10 years, the band-
width resources in the Internet have dramatically
increased, and many new killer applications (e.g.,
streaming services, Mapreduce in data centers)
have been deployed. These changes have led to
new trends in ON/OFF traffic shaping. Specifi-
cally, many application developers and content
providers wish to fully utilize such increased net-
work resources to trade the system performance
improvement at the user or server side while
keeping application performance satisfactory.
Consequently, much traffic is often intentionally
shaped into an ON/OFF pattern to achieve such
optimized system performance (e.g., reduced
computing overhead at the server side for data
transmissions [1] or reduced energy consump-
tion at wireless terminals for providing/receiv-
ing streaming services [2]). Also, some ON/OFF
sources come from the inherent characteristics of
new applications (e.g., Mapreduce).
The ON/OFF traffic characteristics have had a
big impact on network performance and also users’
quality of experiences (QoE). First, the burstiness of
ON/OFF traffic can cause a significant increase
in packet loss probability. For example, [3] shows
that packets at the tail of a burst for web browsing
traffic are more prone to be dropped than those
packets in the front part of such a burst. Reference
[4] shows that 40 percent of packet losses in You-
Tube traffic could be attributed to the ON/OFF
characteristics of YouTube traffic. Second, the high
burstiness of ON/OFF traffic may affect the QoE
of other real-time applications (e.g., VoIP) in the
network. Nowadays, there are usually deep buffers
at edge routers of the Internet. A large burst of
ON/OFF traffic may temporarily produce a delay
of a few hundreds of milliseconds at such edge
routers. Such a large delay may affect the perfor-
mance of real-time applications such as VoIP heavi-
ly. Third, the ON/OFF traffic pattern may make the
application layer misinterpret the available band-
width on the underlying data path and then make
inappropriate choices on traffic sending rates. For
example, an HTTP video streaming source adjusts
its encoding rate according to the available band-
width estimated by the underlying TCP. When mul-
tiple HTTP streaming flows share a common link, it
becomes difficult for the underlying TCP to accu-
rately estimate the available bandwidth on the link
since TCP is unable to perform bandwidth probing
in OFF states during which there is no traffic on the
data path. In the worst case, unfair link bandwidth
sharing or even video rate oscillation may occur
[5, 6].
In this article, we present a comprehensive
survey on the ON/OFF traffic shaping in the cur-
rent Internet. We first introduce the motivations
behind the ON/OFF traffic shaping in different
fields based on different system performance opti-
mization objectives and application characteris-
tics. We then introduce the impacts brought by
the ON/OFF traffic on packet drop probability,
real-time applications, and also its interaction with
TCP’s congestion control mechanism. We present
a survey on existing mechanisms for mitigating
these impacts in various aspects. We introduce
ON/OFF Traffic Shaping in the Internet: Motivation, Challenges, and Solutions
Yongxiang Zhao, Baoxian Zhang, Cheng Li, and Changjia Chen
Yongxiang Zhao and
Changjia Chen are with
Beijing Jiaotong University.
Baoxian Zhang is with
University of Chinese
Academy of Sciences.
Cheng Li is with Memorial
University of Newfoundland.
ACCEPTED FROM OPEN CALL
Digital Object Identifier:
10.1109/MNET.2017.1500057NM
1...,40,41,42,43,44,45,46,47,48,49 51,52,53,54,55,56,57,58,59,60,...100
Powered by FlippingBook