﻿IEEE/ACM TRANSACTIONS ON NETWORKING 1
Anomalous Loss Performance for Mixed Real-Time
and TCP Traffic in Routers With Very Small Buffers
Arun Vishwanath, Student Member, IEEE, Vijay Sivaraman, Member, IEEE, and
George N. Rouskas, Senior Member, IEEE, Member, ACM
Abstract—In the past few years there has been vigorous debate
regarding the size of buffers required at core Internet routers. Recent
arguments supported by theory and experimentation show
that under certain conditions, core router buffer sizes of a few tens
of packets suffice for realizing acceptable end-to-end TCP throughputs.
This is a significant step toward the realization of optical
packet switched (OPS) networks, which are inherently limited in
their ability to buffer optical signals. However, prior studies have
largely ignored the presence of real-time traffic, which is increasing
in importance as a source of revenue for Internet service providers.
In this paper, we study the interaction that happens between realtime
(open-loop) and TCP (closed-loop) traffic when they multiplex
at buffers of very small size (few tens of packets) and make
a significant discovery—namely that in a specific range of buffer
size, real-time traffic losses increase as buffer size becomes larger.
Our contributions pertaining to this anomalous behavior are threefold.
First, we exhibit this anomalous loss performance for realtime
traffic via extensive simulations using synthetic traffic and
real video traces. Second, we develop quantitative models that reveal
the dynamics of buffer sharing between real-time and TCP
traffic that lead to this behavior. Third, we show how various factors
such as the nature of real-time traffic, mixture of long-lived
and short-lived TCP flows, and packet sizes impact the severity
of the anomaly. Our study is the first to consider interactions between
real-time and TCP traffic in very small (potentially all-optical)
buffers and informs router manufacturers and network operators
of the factors to consider when dimensioning such small
buffer sizes for desired performance balance between real-time and
TCP traffic.
Index Terms—Anomalous loss performance, mixed TCP and
real-time traffic, optical packet switched (OPS) networks, routers
with very small buffers.
I. INTRODUCTION
I
N RECENT years, there has been vigorous debate on how
large buffers at an Internet router should be. Conventional
wisdom, attributed to [1], holds that a router should be able to
store a round-trip-time worth of data so as to keep the output link
fully utilized while TCP ramps up its window size after a loss
Manuscript received June 12, 2009; revised January 07, 2010 and July 31,
2010; accepted October 27, 2010; approved by IEEE/ACM TRANSACTIONS ON
NETWORKING Editor I. Keslassy. This work is an extended version of papers
presented at the IEEE Conference on Computer Communications (INFOCOM),
Rio de Janeiro, Brazil, April 19–25, 2009, and the IEEE International Workshop
on Quality of Service (IWQoS), Enschede, The Netherlands, June 2–4, 2008.
A. Vishwanath and V. Sivaraman are with the School of Electrical Engineering
and Telecommunications, University of New South Wales, Sydney,
NSW 2052, Australia (e-mail: arunv@ee.unsw.edu.au; vijay@unsw.edu.au).
G. N. Rouskas is with the Department of Computer Science, North Carolina
State University, Raleigh, NC 27695-8206 USA (e-mail: rouskas@ncsu.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TNET.2010.2091721
event. Equivalently, this rule of thumb mandates buffer size
, where is the average round-trip time of a TCP
connection flowing through the router and is the capacity of
the bottleneck link. For typical
ms, a router with
a Gb/s link would require 10 Gb of buffering, which
poses a considerable challenge to router design.
This buffer sizing rule was first challenged in 2004 by researchers
from Stanford University [2], [3] who showed that
when a large number of long-lived TCP flows multiplex
at a bottleneck link, synchronization does not arise, and near-
100% utilization of the bottleneck link can be achieved with
only
buffers. This means that a router
carrying 10 000 TCP flows needs only buffer 10 000 packets instead
of the million packets required by the rule of thumb.
Since 2004, several new arguments on buffer sizing have been
put forth. Stanford University researchers further proposed in
[4]–[6] that under certain conditions, which they believe hold in
today’s Internet, as few as 20–50 packet buffers suffice for TCP
traffic to realize acceptable link utilization, a claim supported
by their experimental results at Sprint ATL and Verizon Communications
[7]. A measurement study on a Sprint backbone
router also found the queue size to seldom exceed 10 packets [8],
while the choice of 50 packet buffers is recommended in [9] to
guarantee overall stability. These initial results show the feasibility
of building all-optical networks that can be operated at
70%–80% utilization using routers having very small packet
buffers. Clearly, the aforementioned results have significant implications
from an all-optical router design point of view, where
buffering presents a very important but difficult operation, since
data is to be retained in the optical domain.
Researchers from Georgia Tech [10] revisited the ongoing
buffer sizing debate from the perspective of average per-flow
TCP throughput rather than focusing purely on link utilization.
The authors present evidence to suggest that the output/input
capacity ratio at a router’s interface largely governs the amount
of buffering needed at that interface. If this ratio is greater
than 1, then the loss rate falls exponentially, and only a very
small amount of buffering is needed, which corroborates with
the results reported in [6]. However, the concern is that if
the output/input capacity ratio is lower than 1, then the loss
rate follows a power-law reduction and significant buffering
is needed. Researchers from the University of Illinois at
Urbana–Champaign also arrive at a similar conclusion in [11].
Other studies have considered factors such as application layer
performance [12], [13] and fairness [14] influencing buffer
sizing. In late 2008, the Stanford group presented experimental
results validating the applicability of routers with very small
buffers in the core of the Internet [15]. For a comprehensive
IEEE Proof
Web Version
1063-6692/$26.00 © 2010 IEEE2 IEEE/ACM TRANSACTIONS ON NETWORKING
survey on this topic of router buffer sizing, we refer the reader
to our recent survey paper [16].
A. Motivation
While most prior studies on buffer sizing have focused on
electronic Internet routers, some earlier works such as [6] and
[17] have applied buffer sizing principles to optical switches.
However, they focus entirely on TCP traffic performance and
ignore the performance implications for real-time traffic. From
the observation of traffic in the Internet core, it is widely
accepted that TCP constitutes nearly 85%–90% of the traffic,
while real-time (UDP) accounts for about 5%–10%. This has
led all previous work to largely ignore the impact of very small
buffers on UDP’s performance. In this paper, we focus our
attention on what happens at a router with very small buffers
when both open-loop UDP and closed-loop TCP traffic coexist,
and we show why it is important to address their joint performance.
We use the term real-time, UDP, and open-loop traffic
interchangeably.
To understand the dynamics of buffer occupancy at a bottleneck-link
router, we mixed a small fraction of UDP traffic with
TCP traffic and measured the UDP packet loss and end-to-end
TCP throughput. Before starting our simulations, our intuition
was that in the regime of very small buffers (up to 50 kB):
1) UDP packet loss would fall monotonically with buffer size;
2) end-to-end TCP throughput would increase with buffer
size to saturation.
Surprisingly, our observation was contrary to our intuition.
We found that there exists a certain continuous region of buffer
sizes (typically starting from 8–10 kB or so) wherein the performance
of real-time traffic degrades with increasing buffer
size. In other words, packet loss for real-time traffic increases as
the buffer size increases within this region. We call this region
of buffer size an “anomalous region” with respect to real-time
traffic. More surprisingly, we found that when there are a sufficiently
large number of TCP flows, this performance degradation
for UDP traffic does not come with any significant improvement
in end-to-end TCP throughput, and in fact the inflection
point occurs around the buffer size region corresponding to
when TCP has nearly attained its saturation throughput.
This phenomenon is important for a number of reasons and
forms the motivation for the study in this paper. First, as realtime
multimedia applications such as online gaming, interactive
audio-video services, VoIP, and IPTV proliferate in the Internet
and become part of an ISP’s revenue stream, router buffer sizing
studies cannot afford to ignore the performance impact on realtime
traffic when it is multiplexed with TCP traffic, which to
the best of our knowledge has not been undertaken before in the
context of very small buffers.
Second, in the regime of very small buffers, it is prudent to
size router buffers so as to balance the performance of TCP and
UDP traffic. Operating the router buffers in the “anomalous region”
can result in increased UDP packet loss, with only a marginal
improvement in end-to-end TCP throughput, which is undesirable
from a network operator’s point of view.
Third, our results have important implications for emerging
optical packet switched networks since buffering of packets in
the optical domain remains a complex and expensive operation.
Optical buffering using fiber delay lines (FDLs) [18] is impractical
due to the bulk of large spools (1 km of fiber buffers light for
only s) and the large optical crossbar sizes needed to switch
packets in and out of the delay lines (see, for example, the shared
memory architecture in [19]). Recent advances in integrated
photonic circuits have enabled optical first-in–first-out (FIFO)
buffers [20]–[22], though their size is limited to a few tens of
packets and is not expected to increase dramatically in the near
future. The anomaly revealed by our study shows that the investment
made in deploying larger optical buffers has the potential
to negatively impact quality of service and lead to worse performance,
which could be a significant concern for the operator of
the network.
Finally, recent work, such as that in [23] and [24], has proposed
the use of adaptive strategies wherein routers adapt their
buffer size on the fly (depending upon the prevailing network
conditions) so as to achieve a desired loss rate and link utilization.
However, these studies rely on the existence of a monotonic
relationship between packet loss and buffer size, i.e., that
loss rate decreases as buffer size increases. Since our results indicate
that this is not always the case, it is important that further
studies be wary of the nonmonotonicity that may exist when
modeling the interaction between TCP and UDP traffic in the
regime of very small buffers.
B. Contributions of This Work
In the context of loss performance for real-time (UDP) traffic
multiplexed with TCP traffic in routers with very small buffers,
our contributions are threefold.
1) We demonstrate via extensive simulations the existence of
anomalous loss performance (namely an increase in loss
with increasing router buffer size) for real-time traffic.
Our simulations consider real video traces, synthetic
short-range dependent Poisson and long-range dependent
fractional Brownian motion models, and several thousand
TCP flows.
2) We develop quantitative models that help explain
the anomaly. Our first model intuitively captures the
buffer-sharing dynamics between real-time and TCP
traffic and shows the impact of the latter’s greedy nature
on the effective buffers available to the former. Our second
model takes a more rigorous Markov chain-based approach
and allows explicit numerical evaluation of packet
loss. Both models validate the anomalous loss seen in
simulations and provide an analytical handle to explore
the phenomenon in greater depth.
3) We illustrate the impact of several system parameters on
the severity of the anomaly. Significant ones include the
relative mix of short/long-lived TCP flows, characteristics
of real-time traffic, and distribution of packet sizes.
We believe the phenomenon studied in this paper adds a new
dimension to the ongoing debate on buffer sizing, including in
the context of optical packet switches, and our results aid switch
manufacturers and network operators in selecting small buffer
sizes that achieve desired performance balance between TCP
and real-time traffic.
The rest of this paper is organized as follows. In Section II, we
introduce the anomalous loss behavior using real video traces.
IEEE Proof
Web VersionVISHWANATH et al.: ANOMALOUS LOSS PERFORMANCE FOR MIXED REAL-TIME AND TCP TRAFFIC 3
Fig. 1. ns-2 simulation topology.
In Section III, we develop an intuitive analytical model that captures
this phenomenon succinctly. In Section IV, we present
a more rigorous analysis using an M/M/1/B queueing model,
which further validates the anomaly observed in simulations. In
Section V, we study how various network design factors such
as real-time and TCP traffic characteristics affect the anomalous
loss performance. In Section VI, we investigate the impact
of varying UDP packet sizes on the anomaly. We conclude the
paper in Section VII and point to directions for future work.
II. ANOMALY
To illustrate the anomalous loss behavior, we consider a
simple dumbbell topology (Fig. 1) that is commonly used
to analyze the performance of various congestion control
algorithms, including TCP. Such a topology captures packet
queueing effects at the bottleneck link, which is the dominant
factor in end-to-end loss and delay variation for each flow, while
abstracting the rest of the flow path by an access link on either
side of the bottleneck link. We use ns-2 [25] (ver. 2.30) for our
simulations and consider 1000 TCP flows, corresponding to
each source–destination pair - - .
Furthermore, we use TCP-Reno in all our simulations, consistent
with the TCP version used in previous related work on
buffer sizing, and employ FIFO queue with drop-tail queue
management, which is commonly used in most routers today.
UDP traffic is generated between nodes - - .It
suffices to have a single UDP flow since open-loop traffic can,
without loss of generality, be aggregated. Multiple UDP flows
traversing the bottleneck link can thus be modeled as a single
UDP flow that represents the aggregate of all individual UDP
flows passing through that bottleneck link. However, we need
multiple TCP flows since they each react independently to the
prevailing network condition and the state of the buffers.
The propagation delay on the UDP access link is chosen at
5 ms, while it is uniformly distributed between ms on
the TCP access links. The propagation delay on the bottleneck
link is 50 ms, thus round-trip times vary between 102
and 150 ms. All TCP sources start at random times between
s. UDP source starts at time 0 s. The simulation duration
is 800 s, and performance measurements are recorded after 200 s
to allow for the stabilization of all TCP flows.
Buffer size at the bottleneck router is varied in terms of
kilobytes. To set the packet sizes, we draw on the fact that several
real-time applications, e.g., online gaming [26], [27], use
small UDP packets since they require extremely low latencies.
The study showed that almost all packets were under 200 bytes.
Our experiments using Skype and Yahoo! Messenger showed
that for interactive voice chat, UDP packet sizes were between
150–200 bytes. Also, traces obtained at a trans-Pacific 150-Mb/s
link [28] suggest that average UDP packet sizes are smaller than
average TCP packet sizes. Therefore, in all our simulations, we
fix the TCP packet size at 1000 bytes and simulate fixed- and
variable-size UDP packets in the range of bytes.
Akin to the traffic in the Internet core, we want to keep
the fraction of UDP traffic to within 3%–10% as well. We
performed simulations using various movie traces such as Star
Wars, Jurassic Park I, Die Hard III, Silence of the Lambs,
Aladdin, etc. For brevity, we present results from only a subset
of the movies mentioned. Results for the movies not described
here closely follow the ones described. All the movie traces
have been profiled and are known to exhibit self-similar and
long-range-dependent traffic characteristics.
We first illustrate the phenomenon using the video traffic
trace from the movie Star Wars, obtained from [29] and references
therein. The mean rate is 374.4 kb/s, and the peak rate is
4.446 Mb/s, with the peak-rate-to-mean-rate ratio being nearly
12. The packet size is fixed at 200 bytes. We set the bottleneck
link at 10 Mb/s and the TCP access links at 1 Mb/s, while
the UDP access link is kept at 100 Mb/s. The bottleneck link
was only 10 Mb/s because the mean rate of the video trace
(UDP) is low (374.4 kb/s), and we want to keep the fraction
of UDP traffic feeding into the core to within 3%–10% of the
bottleneck-link rate (to be consistent with the nature of Internet
traffic today). In this example, the video traffic constitutes
3.75% of the bottleneck-link rate. Subsequent sections of the
paper will present results considering higher bottleneck-link
rates as well.
We have a high-speed access link for UDP since UDP traffic
feeding into the core can be an aggregate of many individual
UDP streams. TCP traffic on the 1-Mb/s access link models
traffic from a typical home user. Fig. 2 shows the UDP packet
loss and TCP throughput curves as a function of buffer size at
the bottleneck router in the range of 1–50 kB. We see that TCP
quickly ramps up to nearly 9.6 Mb/s with only about 8 kB of
buffering, reaching close to its saturation throughput. Simultaneously,
UDP packet loss falls rapidly as well. Up to this point,
both TCP and UDP behave as expected. However, as the buffer
size increases beyond 8 kB till about 24 kB, UDP performance
degrades as its packet loss increases with buffer size in this region.
The loss at 24 kB of buffering is approximately 30% more
than the loss at 8 kB of buffering. There is, however, no appreciable
increase in TCP throughput.
To verify that the anomalous UDP loss was not unique to
the packet trace we had chosen, we performed our simulations
using several other traces. For instance, we used video traces
from the movie Jurassic Park I obtained from [30], with packet
sizes uniformly distributed in the range bytes. The
IEEE Proof
Web Version
video traffic contributes 7.7% of the bottleneck-link rate. Fig. 3
shows the corresponding UDP packet loss curves as a function
of buffer size corresponding to this scenario and clearly indicates
the presence of a region of buffer size in which loss increases
with buffer size. We also observed this behavior when4 IEEE/ACM TRANSACTIONS ON NETWORKING
Fig. 2. Star Wars 200-byte packets: UDP loss and TCP throughput.
Fig. 3. Jurassic Park [150; 300]-byte packets: UDP loss and TCP throughput.
UDP traffic is generated from synthetic models (discussed in
Section V), suggesting that the anomaly is fundamental and not
just a coincidence. Moreover, in the absence of TCP traffic, UDP
loss was observed to drop monotonically with buffer size, confirming
that the anomaly arises due to the interaction of openand
closed-loop traffic at the bottleneck router with very small
buffers.
Through our results in this study, we hope to bring the
anomaly to the attention of optical switch vendors and service
providers who could make considerable investments in
incorporating optical buffering in their packet switches, only
to obtain potentially worse performance if they inadvertently
operate their buffer sizes in this anomalous region. Given that
each extra kilobyte of optical buffering can add significantly
to the cost of the optical switch (this was confirmed in our
interactions with Bell Labs researchers who are prototyping an
all-optical packet switch called IRIS [31]), manufacturers and
operators should be wary of the potential for negative returns
on this investment.
Having observed the anomalous loss behavior in several simulation
scenarios, we now develop quantitative models that can
help explain this phenomenon.
Fig. 4.
CDF of average congestion window size for 1000 and 200 TCP flows.
III. INTUITIVE MODEL OF THE ANOMALY
We begin with an intuitive explanation of why we think the
anomaly happens, and this helps us develop a simplistic yet effective
model that quantifies the buffer-sharing dynamics between
TCP and real-time traffic.
To explain the anomaly, we start by characterizing the dynamics
of TCP congestion control as a function of bottlenecklink
buffer size. We show in Fig. 4 the cumulative distribution
function (CDF) of the average window size of the TCP flows
(obtained from simulation) for various buffer sizes at the bottleneck
link under two different flow settings (both of which
carry 5% Poisson UDP traffic): one in which 1000 flows share
a 200-Mb/s link, and the other in which 200 flows share the
200-Mb/s bottleneck link (the TCP per-flow rates are 200 kb/s
and 1 Mb/s, respectively, for the two scenarios, representative of
different provisioning scenarios, and we note that the anomaly is
seen in both the settings). We observe from the figure that when
buffers at the bottleneck link are extremely small (say in the
range 2–5 kB), the congestion window size of the TCP flows are
also small—for example, we see that when bottleneck buffers
are 2 kB, average congestion window size is 1.8–4 packets when
there are 1000 flows, and 6–9 packets when there are 200 flows.
As the bottleneck buffer grows larger, say to 10–25 kB, TCP
average congestion window sizes also increase—for example,
the figure shows that when bottleneck buffers are say 20 kB, average
congestion window sizes are in the range 2.1–5.6 packets
for the first scenario with 1000 TCP flows, and in the range
12–20 packets for the second scenario with 200 TCP flows. This
represents a factor 1.2–2 increase in average flow congestion
window size as buffers increase from 2 to 20 kB.
The significance of TCP’s average congestion window size
on the anomaly is as follows. When TCP congestion window is
very small (due to very small bottleneck buffers of say 1–5 kB),
each TCP flow transmits only a few packets in each round-trip
time, and is therefore mostly idle. Consequently, the buffers at
the bottleneck link are often devoid of TCP packets, allowing
UDP packets to enjoy use of these buffers for the most part.
In this region, therefore, TCP and UDP predominantly “timeshare”
the buffers, and UDP loss decreases with buffer size,
much like it would if TCP traffic were nonexistent. On the other
IEEE Proof
Web VersionVISHWANATH et al.: ANOMALOUS LOSS PERFORMANCE FOR MIXED REAL-TIME AND TCP TRAFFIC 5
hand, when a larger fraction of the TCP flows are able to increase
their congestion window (equivalently a smaller fraction
of the TCP flows remains idle), due to bottleneck buffers
being larger (say in the range 10–25 kB corresponding to the
anomaly), TCP traffic makes more use of the buffers at the bottleneck
link, leaving a smaller fraction of the buffers for UDP
traffic to use. The aggressive nature of TCP in increasing its
congestion window to probe for additional bandwidth causes
the “space sharing” of bottleneck-link buffers between TCP and
UDP in this region to be skewed in favor of TCP, leaving lesser
buffers available to UDP traffic even as buffer size increases.
We now try to quantify this intuition via a simple analytical
model that captures the transition from time sharing to space
sharing of the bottleneck-link buffers between TCP and realtime
traffic. We make the assumption that there is a sufficiently
large number of TCP flows sharing the bottleneck link, and that
they have sufficiently large round-trip times such that the delaybandwidth
product is larger than the buffering available at the
bottleneck link. Moreover, TCP is assumed to contribute a vast
majority of the overall traffic on the link (this is consistent with
observations that nearly 85%–90% of today’s Internet traffic is
carried by TCP). Under such circumstances, we first make the
following observation.
Observation: TCP’s usage of the bottleneck buffers increases
exponentially with the size of the buffer. More formally, let
denote the buffer size (in kilobytes) at the bottleneck link, and
the probability that at an arbitrary instant of time the
buffers at the bottleneck link are devoid of TCP traffic. Then
where is a constant (with same unit as ) dependent on
system parameters such as link capacity, number of TCP flows,
round-trip times, ratio of long-lived to short-lived TCP flows,
etc. The constant can be inferred from the plot of the natural
logarithm of as a function of , which yields a straight
line. The slope of the line corresponds to .
This behavior has been observed in the past by various
researchers, by direct measurement of idle buffer probabilities
[32, Sec. III] as well as indirectly via measurement of
TCP throughput [6, Fig. 1]. The latter has shown roughly
exponential rise in TCP throughput with bottleneck buffer size,
confirming that TCP’s loss in throughput (which arises from
an idle buffer) falls exponentially with buffer size. We also
validated this via extensive simulations (shown in Fig. 5 and in
various other TCP plots in later sections) in ns-2. One thousand
TCP flows with random round-trip times from a chosen range
were multiplexed at a bottleneck link, and the idle buffer probability
was measured as a function of bottleneck-link buffer size.
The large number of flows, coupled with randomness in their
round-trip times, ensures that the TCP flows do not synchronize
their congestion windows. Fig. 5 plots on log-scale the idle
buffer probability as a function of bottleneck buffer size for two
ranges of round-trip times and shows fairly linear behavior in
the range of 5–50 packets (each packet was 1 kB), confirming
the exponential fall as per (1).
Having quantified TCP’s usage of the bottleneck buffers,
we now consider a small fraction (say 5%–10%) of
(1)
Fig. 5. Probability of idle buffer versus buffer size for TCP traffic.
real-time (UDP) traffic multiplexed with TCP traffic at the
bottleneck link. The small volume of UDP traffic does not
alter TCP behavior significantly. However, TCP’s usage of
the buffer does significantly impact loss for UDP traffic. If we
assume the buffer is very small (a few tens of kilobytes), we
can approximate the buffer as being in one of two states: idle
(empty) or busy (full). With the objective of estimating the
“effective” buffers space available to UDP traffic, we identify
the following two components.
• Fair-share: During periods of time when TCP and UDP
packets coexist in the buffer, the buffer capacity is shared
by them in proportion to their respective rates. The firstin–first-out
nature of service implies that the average time
spent by a packet in the system is independent of whether
the packet is UDP or TCP, and Little’s law can be invoked
to infer that the average number of waiting packets of a
class is proportional to the arrival rate of that class. UDP
packets therefore have, on average, access to a “fair share”
of the buffers, namely , where denotes the fraction of
total traffic that is UDP.
• Time-share: Whenever the buffer is devoid of TCP traffic
(i.e., with probability ), UDP packets have access
to the remaining buffer space as well. We call
this the “time-share” portion since this portion of the
buffer is shared in time between UDP and TCP traffic. The
time-share portion of buffers available to UDP is therefore
.
Combining the fair-share and time-share components and
invoking (1) gives us an estimate of the total “effective”
buffers available to UDP traffic
IEEE Proof
To illustrate the significance of this equation, we plot it for
(i.e., 5% UDP traffic) and kB (consistent
from Fig. 5). Fig. 6 shows the total effective buffers for UDP, as
well as the fair-share and time-share components. The fair-share
component increases linearly with buffer size, while the
time-share component
rises to a peak and then
falls again (readers may notice a shape similar to the Aloha protocol’s
throughput curve). This happens because smaller buffers
Web Version
(2)6 IEEE/ACM TRANSACTIONS ON NETWORKING
Fig. 6. Effective buffers for UDP traffic.
are more available for UDP to time-share, but as buffers get
larger, TCP permits exponentially diminishing opportunity for
time sharing. The total effective buffers for UDP, being the sum
of the above two components, can therefore show anomalous behavior,
i.e., a region where larger real buffers can yield smaller
effective buffers for UDP. For any realistic UDP traffic model
(note that our analytical model does not make any specific assumption
about the UDP traffic model), the smaller effective
buffers will result in higher loss, which is of serious concern to
network designers and operators who operate their router buffer
sizes in this region.
The model presented is highly simplified and ignores several
aspects of TCP dynamics as well as real-time traffic characteristics.
It nevertheless provides valuable insight into the
anomaly and will be used in later sections for a quantitative
understanding of the impact of various parameters on the
severity of the anomaly.
IV. MARKOV MODEL OF THE ANOMALY
It is challenging in general to mathematically analyze finite
buffer systems in which several thousand feedback-based adaptive
TCP flows interact with stochastic real-time traffic. In what
follows, we develop a realistic yet rigorous Markov model based
on some simplifications.
1) Assumption: TCP Packet Arrivals Are Poisson: If a large
number (potentially thousands) of long-lived TCP flows multiplex
at a bottleneck link, it is believed [2] they do not synchronize
their window dynamics behavior and can thus be treated as
independent flows. Combined with the fact that each TCP flow’s
window will be quite small (since bottleneck buffers are small),
implying that each flow will only generate a small amount of
traffic per RTT, the aggregation of a large number of such independent
flows can reasonably be assumed to yield Poisson
traffic. Prior studies on buffer sizing have also employed this
assumption [11]. We further validate this assumption using ns-2
simulations on the dumbbell topology shown in Fig. 1.
Two thousand TCP flows with random RTTs and start
times are multiplexed at the bottleneck link, which operates at
200 Mb/s. TCP packet size is fixed at 1 kB. Two sets of buffer
sizes are chosen at the bottleneck router : 50 kB representing
the very small buffer case, and 6250 kB ms Mb/s
representing the traditional delay-bandwidth product rule. The
Fig. 7. TCP traffic burstiness at different timescales for buffer sizes of 6250
and 50 kB.
simulation is run for 151 s, and data in the range 15–150 s is
used in all the computations. Our objective is to test if TCP
packet arrivals to the bottleneck-link buffer are near-Poisson.
To do so, we measure the burstiness of the arrival traffic at
various timescales. Burstiness at timescale is quantified by
, the coefficient of variation (i.e., ratio of standard deviation
to mean) of traffic volume measured over time intervals of
size . Log-log plots of versus are routinely used in the
literature to depict traffic burstiness over various timescales as
an indicator of self-similarity of traffic traces and to show the
influence of the Hurst parameter . We will use estimates of
obtained from the burstiness plots as a measure of how close
the traffic is to being short-range-dependent (i.e., Poisson).
Fig. 7 shows the traffic burstiness as a function of the
timescale (in microseconds) for four different combinations
of TCP flows and buffer size. The top two curves correspond
to large (delay-bandwidth) buffer size of 6250 kB, while the
bottom two are for small buffers of 50 kB (the two curves in each
set correspond respectively to short- and long-lived flows, with
the former generated using a model described in Section VI).
Both sets of curves are fairly similar till timescale of s
ms. However, at timescales beyond a millisecond, the curves
for large buffers flatten significantly, indicating onset of longrange
dependence with a Hurst parameter estimated at approximately
0.8. This shows that when buffers are large, TCP traffic is
significantly bursty at timescales commensurate with the time it
takes to empty/fill the buffer, and the buffer dynamics cannot be
modeled with a Poisson assumption (due to the exhibited long
range dependence).
By contrast, the bottom two curves (corresponding to
small buffers) in the plot do not change slope significantly at
timescales beyond 1 ms, and they retain a slope close to that
of short-range-dependent traffic. This lends credence to our
assumption that small bottleneck buffers help reduce synchronization
among flows and make the traffic akin to Poisson.
To further affirm the Poisson nature of aggregated TCP traffic
in the core, we investigated the impact of increasing the number
of TCP flows. Fig. 8 plots the TCP burstiness for 100, 500, 1000,
and 2000 long-lived flows for fixed buffer size of 50 kB at the
IEEE Proof
Web VersionVISHWANATH et al.: ANOMALOUS LOSS PERFORMANCE FOR MIXED REAL-TIME AND TCP TRAFFIC 7
Fig. 8. TCP traffic burstiness at different timescales with varying number of
flows and 50-kB buffer size.
bottleneck router in the simulation topology of Fig. 1. It can be
clearly noted that at timescales beyond a millisecond, the burstiness
curve falls more steeply (and hence the traffic exhibits reducing
long range dependence) as the number of TCP flows increases.
This shows that the TCP arrival process becomes more
Poisson as the number of flows increases (due to scalability and
memory issues, we were unable to scale our simulations to a
larger number of flows).
In summary, if bottleneck buffers are small, and if the number
of TCP flows multiplexed is large, the assumption that aggregate
TCP traffic arrivals to the bottleneck link are Poisson is well
justified.
2) Assumption: UDP Packet Arrivals Are Also Poisson: Stochastic
studies such as [33]–[35] have shown that the aggregation
of traffic from a large number of independent flows (as can
be expected at a core link) converges to Poisson. This important
result makes the analysis tractable (though the phenomenon
of anomalous loss is observed even if UDP arrivals are
non-Poisson).
3) Claim: UDP packets are, on average. smaller in size than
TCP packets, as discussed in Section II and as reported in several
measurements of traffic in the Internet core [28]. Consistent with
our example presented in Fig. 2, we choose average TCP and
UDP packet sizes to be 1000 and 200 bytes.
4) Claim: The aggregate TCP rate increases exponentially
with bottleneck-link buffer size, as demonstrated in Fig. 5 and
discussed in the Section III. Denoting the bottleneck buffer size
as (in kilobytes), the TCP throughput is given by
where denotes the saturation throughput of TCP (when
buffer size is large).
In order to construct a Markov chain model, we make
the further assumption that packet transmission times are
exponentially distributed (we will relax this assumption in
Section IV-A). We can then model the FIFO queue at the
bottleneck-link router as an M/M/1 system with finite buffer
and with two classes of customers.
IEEE Proof
(3)
1) UDP arrivals are Poisson at fixed rate (denoted by )
and require exponential service time with unit mean (the
service rate is normalized to average UDP packet size).
2) TCP arrivals are Poisson at rate derived from (3),
where each TCP packet arrival brings a bulk of five customers
(corresponding to the packet size ratio 1000/200),
each requiring exponential service time with unit average.
For illustrative purposes, let us consider the buffer size to
be 3 kB. Then, we can model the state of the system as the
number of customers in the FIFO queue. Fig. 9 shows the resulting
Markov chain. A transition from state to state
corresponds to the arrival of a TCP packet, whereas a transition
from state to state corresponds to the arrival of a UDP
packet.
Denoting
to be the corresponding
buffer size in bytes, and the number of states in the Markov
chain, then
If represents the steady-state probability of the queue being
in state (i.e., the probability that the queue contains customers),
then we can write the global balance equations as follows:
Web Version
(4)
(5)
(6)
(7)
(8)
(9)
The above equations and the normalizing constraint
form a set of linear equations that can be
solved to compute the probability that an incoming UDP packet
will be dropped, which in this example is . Obtaining balance
equations as the buffer size increases is straightforward,
and the resulting set of linear equations is easily solvable numerically
(in MATLAB) to get the UDP packet loss probability.
The analytical result shown in this paper chooses model parameters
to match the simulation setting as closely as possible.
The normalized UDP rate is set to
(i.e., 5% of link
capacity), and the TCP saturation throughput
(so that TCP and UDP customers have a combined maximum
rate less than the service rate of in order to guarantee
stability). The constant kB is consistent with what is
obtained from simulations in Fig. 5.
Fig. 10 plots the UDP loss (on log-scale) obtained from
solving the M/M/1 chain with bulk arrivals and finite buffers,
as well as the TCP rate in (3), as a function of buffer size .It
can be observed that in the region of 1–8 kB of buffering, UDP
loss falls monotonically with buffer size. However, in the buffer
size region between 9–30 kB, UDP packet loss increases with8 IEEE/ACM TRANSACTIONS ON NETWORKING
Fig. 9. Markov chain state transition diagram for buffer occupancy with buffer size = 3000 bytes.
Fig. 10.
Anomalous UDP loss results from the M/M/1/B analytical model.
increasing buffer size, showing that the model is able to predict
the anomaly found in simulations.
A. M/D/1/B Analysis
We refine the M/M/1/B model by relaxing the assumption
that packet sizes are exponentially distributed. It has been
recently observed that Internet packet sizes have a bimodal
distribution [37], [38]: with peaks at large packets (1500-byte
TCP data) and small packets (typically 40-byte TCP ACK).
Real-time and other streams generate intermediate packet
sizes (200–500 bytes). To develop a model that is tractable yet
reflective of these dominant modes, we employ an M/D/1/B
model in which packet sizes are bimodal: large for TCP
packets (1000 bytes) and small for UDP packets (say 200 bytes).
For brevity, we do not explain the derivation of this model here
and instead refer the interested reader to our paper [39] for the
detailed analysis. It suffices to state here that the results from
the M/D/1/B analysis are qualitatively similar to the M/M/1/B
results presented, and both chains predict the inflection point to
occur at around 8 kB. Moreover, the M/D/1/B model validates
the anomaly from the realistic scenario of having bimodal
packet sizes and not relying on exponential service times.
V. IMPACT OF REAL-TIME TRAFFIC CHARACTERISTICS
Using simulations (all performed on the dumbbell topology
shown in Fig. 1) as well as the analytical models developed, we
now investigate the impact of various system parameters on the
nature of the observed anomalous loss performance. This section
studies how real-time traffic characteristics such as traffic
model and intensity affect the anomaly, while subsequent sections
consider the effect of TCP traffic characteristics and packet
size distributions.
A. Traffic Model
It may be recalled that our intuitive model in Section III did
not make any specific assumptions about the stochastics of UDP
Fig. 11.
Poisson: UDP packet loss and TCP throughput from simulation.
traffic. We now validate via simulation that the anomaly is agnostic
to whether the UDP traffic exhibits short- or long-range
dependence. The two models are described next.
1) Poisson: We start with the well-known Poisson model as
the UDP traffic source. The core link (we use the term “core
link” to refer to the bottleneck link, and vice versa) bandwidth
is set at 100 Mb/s. TCP access links are at 10 Mb/s each, while
the UDP access link operates at 100 Mb/s. The average rate
of Poisson traffic is 5 Mb/s, constituting about 5% of the total
bottleneck-link bandwidth. Fig. 11 shows the UDP packet loss
(on log-scale) and the corresponding TCP throughput curves
when the buffer size at the bottleneck router is varied from 1 to
50 kB. TCP is able to quickly ramp up to nearly 93 Mb/s with
just about 11 kB of buffering, corresponding to nearly 98% of its
saturation throughput. We note from the figure that up to 11 kB,
UDP packet loss falls with increasing buffer size. In addition,
further increase in buffer size leads to an increase in UDP packet
loss. The loss at 30 kB of buffering is 50% more than the loss at
11 kB of buffering, while there is only a negligible increase in
TCP throughput.
2) fBm: It is widely believed that Internet traffic is
not Poisson in nature, but tends to exhibit self-similar and
long-range-dependent properties. To see if the phenomenon
also occurs under this scenario, we generated fBm traffic at the
same average rate of 5 Mb/s. Other parameters are the same as
before. The fBm model used is similar to our previous work
in [40] and [41]. The traffic model combines a constant mean
arrival rate with fractional Gaussian noise (fGn) characterized
by zero mean, variance , and Hurst parameter .
We use our filtering method in [42] to generate, for a chosen ,
a sequence of normalized fGn (zero mean and unit variance).
A discretization interval is chosen, and each then denotes
the amount of traffic in addition to the constant rate stream
IEEE Proof
Web VersionVISHWANATH et al.: ANOMALOUS LOSS PERFORMANCE FOR MIXED REAL-TIME AND TCP TRAFFIC 9
Fig. 12. fBm: UDP packet loss and TCP throughput from simulation.
that arrives in the th interval. Specifically, the traffic (in
bits) arriving in the th interval of length s is computed
using where denotes in bits
per second the constant rate stream, and is a scaling factor
that determines the instantaneous burstiness. For this work, we
set the Hurst parameter at , and the discretization
interval s. The scaling factor is chosen to satisfy
, which corresponds to moderate burstiness
(around 16% of the samples are truncated), and is then
adjusted to give the desired mean traffic rate. The fluid traffic
is then packetized into fixed-length packets (of size 200 bytes)
before being fed into the simulations.
We plot the UDP packet loss (on log-scale) and the TCP
throughput curves as a function of buffer size in Fig. 12. Here,
too, as in the case of the Poisson traffic model, TCP attains 98%
of its saturation throughput with only about 11 kB of buffering.
UDP packet loss is the lowest at this point. An increase in buffer
size negatively affects UDP packet loss, but results in only a
marginal improvement in TCP throughput. The loss at 30 kB of
buffering is nearly 50% more than the loss at 11 kB of buffering.
B. Traffic Intensity
Having observed the anomalous loss phenomenon for both
short- and long-range-dependent UDP traffic models, we now
explore if the anomaly is affected by the relative volumes of
real-time and TCP traffic. This helps us understand if ISPs who
carry a relatively larger fraction of real-time traffic are more
susceptible to this anomalous loss performance, and also if the
anomaly would change if real-time applications were to become
more widespread, thereby changing the traffic mix in the
Internet. With this aim, we simulated 1000 TCP flows on a
200-Mb/s core link with (long-range-dependent) UDP traffic
rate set at 5%, 10%, 15%, and 20% of the core-link capacity.
The resulting plots of UDP loss and TCP throughput are shown
in Figs. 13 and 14. We observe from Fig. 13 that when the UDP
rate is 5%, the inflection point is clearly seen to exist at about
9 kB. However, as the fraction of UDP traffic increases, the inflection
point gradually shifts to the left, while the magnitude of
the anomaly seems to diminish.
Fig. 13.
UDP loss with varying fBm rate.
Fig. 14. TCP throughput for varying fBm rates.
To see if the anomaly vanishes entirely at high UDP rates, we
simulated three scenarios, corresponding to 80- and 90-Mb/s average
UDP rates on a 100-Mb/s core link, and 180-Mb/s average
UDP rate on a 200-Mb/s core link, each with 1000 TCP flows.
The fraction of UDP traffic is thus very high at 80%–90%. The
resulting UDP loss curves are plotted in Fig. 15. Clearly, we can
see that the UDP loss curves do not exhibit a point of inflection,
i.e., UDP loss falls monotonically with increasing buffer size
and there is no anomalous loss.
A qualitative explanation for why the anomaly vanishes at
high UDP rates follows. Referring back to the case when UDP
rates are low, increasing buffers in the anomalous region gave
TCP an exponentially larger opportunity to use the overall
buffers, while giving UDP only a minimal fair share of extra
buffering, the net effect being a reduction in the effective buffers
available to UDP. Now, when UDP rates are high, increasing
the buffers at the bottleneck link gives UDP substantially
more buffers as its fair share (in proportion to its rate) while
diminishing the opportunity for TCP to time-share the buffers
with UDP. This results in a net positive gain in the effective
buffers available to UDP. As a result, the UDP packet loss falls
monotonically with increasing buffer size.
IEEE Proof
Web Version10 IEEE/ACM TRANSACTIONS ON NETWORKING
Fig. 15.
UDP loss falling monotonically with buffer size by simulation.
This intuition can be quantified using our buffer-sharing
model developed in Section III. Referring to (2), recall that
represents the fraction of UDP traffic, and the two terms in the
summation represent respectively the fair-share and time-share
components on the effective-buffers available to UDP. We plot
in Fig. 16(a) and (b) the fair-share component, the time-share
component, and the effective buffers when kB, for two
sets of UDP rates, namely 10% (low) and 80% (high). From
the figures, and also from Fig. 6 that plots these values for
(5% UDP traffic), we note that the shape of the curves
corresponding to the time-share component and the effective
buffers available to UDP changes as the UDP rate increases.
The presence of the time-share component is less pronounced,
while that for effective buffers approaches a straight line at
higher rates. To explain the change in the nature of these curves,
we note that from (2), as increases, the fair-share component
begins to dominate over the time-share component
since
becomes negligible (tends toward
0) at large . This implies that the effect of the time-share
component on the effective buffers available to UDP falls with
increasing UDP rate (as seen in the figures). Consequently,
increases linearly with buffer size , which implies that
the effective buffers available to UDP increase as the real buffer
size increases, thus yielding a straight line with slope . This
helps explain why at high UDP rates, simulations show that the
packet loss falls monotonically with buffer size.
VI. IMPACT OF TCP CHARACTERISTICS
Our Markov chain analysis relied on the assumption that
TCP’s usage of buffers increases exponentially with buffer
size. Though this has been observed when all TCP flows are
long-lived ([32, Sec. III], [6, Fig. 1], Fig. 5), the reader may
wonder if similar behavior is seen when many of the TCP
flows are short-lived (or equivalently, the number of TCP
flows is time-varying). This is an important consideration
since measurement based studies at the core of the Internet
suggest that a large number of TCP flows (e.g., HTTP requests)
are short-lived (“mice”) and carry only a small volume of
traffic, while a small number of TCP flows (e.g., FTP) are
long-lived (“elephants”) and carry a large volume of traffic.
We studied UDP loss for such TCP traffic mixes by simulating
them in ns-2 over the dumbbell topology shown in
Fig. 1. In Fig. 17, we plot the TCP empty buffer probability
on a 200-Mb/s core link for four different ratios of long-lived
to short-lived flows. The total number of TCP flows is kept
constant at 2000. In order to incorporate realistic TCP traffic,
we consider the closed-loop flow arrival model described in
[10] and [43], operating as follows. A given number of users
(up to a maximum of 2000 in our example) perform successive
file transfers to their respective destination nodes. The size of
the file to be transferred follows a Pareto distribution with mean
100 kB and shape parameter 1.5. These chosen values are representative
of Internet traffic and comparable to measurement
data. After each file transfer, the user transitions into an idle or
OFF state, or as the authors of [10] suggest, a “thinking period.”
The duration of the thinking period is exponentially distributed
with mean 1 s. It is widely believed that Internet traffic exhibits
self-similar and long-range-dependent characteristics. It can be
noted that the above traffic generation mechanism, which is a
combination of several ON–OFF sources with Pareto-distributed
ON periods, is in fact long-range-dependent [44].
Fig. 17 plots the empty buffer probability for the four different
ratios of long-lived to short-lived flows. Our first observation
is that the empty buffer probability falls fairly linearly (on
log-scale) with buffer size (in the region of 8–50 kB) irrespective
of the traffic mix. This satisfies one of the important assumptions
required by our analytical models shown in (1) and (3),
and thus renders our models valid for mixes of short and long
TCP flows. However, the slope of the linear region, which in
turn determines the required by the model, does seem to depend
on the relative fractions of short- and long-lived flows. The
figure shows that as the fraction of long-lived flows increases
from 0 to 1, the value of decreases correspondingly from 14
down to 7 kB. Intuitively, this is because short-lived flows do
not generate sufficient traffic to continuously saturate the link,
and most of them remain in the slow-start phase without entering
into the congestion avoidance mode during the entire file
transfer process. However, with the increase in the number of
long-lived flows, there is a corresponding increase in the buffer
occupancy since long-lived flows always have data to send and
are more likely to be in the congestion-avoidance mode. This
results in the core link being saturated and reduces the probability
of the buffer being empty, explaining why reduces as
the number of long-lived flows increases.
Having observed how changes with the long–short TCP
flow mix, we study the corresponding impact on the performance
predicted by our analytical model. Fig. 18 shows the
TCP throughput curves obtained from the M/M/1/B analysis as
a function of core-link buffer size for different values of ,
ranging from 1 to 14 kB. The key point to note from the figure
is that as increases, TCP requires bigger buffers to attain saturation
throughput, which is 0.94 or 94% of the core-link rate
since this analysis plot considers the presence of 0.05 or 5%
UDP traffic. In other words, the smaller the , the faster TCP
rises, thus needing fewer buffers to attain saturation throughput.
Fig. 19 shows the impact of on the UDP loss prediction
from our model. We observe that the UDP packet loss (and
hence the severity of the anomaly) is more pronounced when
IEEE Proof
Web VersionVISHWANATH et al.: ANOMALOUS LOSS PERFORMANCE FOR MIXED REAL-TIME AND TCP TRAFFIC 11
Fig. 16. Effective buffers for UDP (from analysis) when the fraction of UDP traffic is 10% (low) and 80% (high). (a) B at 10% rate. (b) B at 80% rate.
Fig. 17. Empty buffer probability with varying number of short- and long-lived
TCP flows.
Fig. 18.
TCP throughput from M/M/1/B analysis.
is larger. Moreover, the inflection point, i.e., the point at which
UDP packet loss begins to increase, shifts slightly to the right
as increases.
To verify if these observations are corroborated in simulation,
we multiplex 10-Mb/s Poisson traffic with the four mixes
of short- and long-lived TCP flows and record the UDP packet
loss as a function of buffer size. Fig. 20 shows that as the number
of long-lived flows increases, there is a corresponding increase
in losses for UDP traffic. Referring back to our analytical model,
this can be argued as follows: When the fraction of long-lived
TCP flows increases from 0 to 1, the core-link buffer occupancy
due to TCP traffic alone increases, reducing from 14 to 7 kB.
Fig. 19.
UDP loss from M/M/1/B analysis.
IEEE Proof
Fig. 20. UDP loss for varying number of short and long TCP flows.
This in turn permits less opportunity for UDP traffic to access
the buffers, leading to higher UDP losses. Though the prediction
from analysis qualitatively matches simulation and even predicts
the inflection point with reasonable accuracy, we notice a
quantitative discrepancy between them, particularly when short
TCP flows dominate. This is because our assumption that TCP
arrivals are Poisson is not very accurate when we do not have a
large number of TCP flows (refer to Fig. 8) and when the volume
of short-lived TCP traffic is significant (refer to Fig. 7) since the
short-lived TCP flows used in simulation have long-range-dependent
characteristics. We believe that if we have potentially
tens of thousands of TCP flows multiplexing at a router with
very small buffers, the loss estimates obtained from the model
Web Version12 IEEE/ACM TRANSACTIONS ON NETWORKING
Fig. 21.
UDP packet loss from simulation with varying packet size ratios.
will be more accurate. Nevertheless, our model validates the
anomaly even in the presence of short/long-lived TCP mixes
and gives a good indication on how the severity of the anomaly
changes with the TCP traffic mix. Given that TCP traffic is by itself
notoriously difficult to analyze, let alone thousands of TCP
flows interacting with real-time traffic, we believe our model
offers sufficient accuracy to be a valuable aid in sizing routers
(potentially all-optical) with very small buffers.
Our study also evaluates the impact of several other TCP parameters
such as round-trip times, number of flows, and corelink
capacity on the severity of the anomaly. With a large enough
number of flows, these parameters did not seem to have a very
dominant effect on the anomaly, and hence we do not discuss
them here. The interested reader is referred to our paper [36] for
a more detailed discussion of the impact of these parameters.
VII. IMPACT OF PACKET SIZE DISTRIBUTION
In this section, we investigate the impact of varying
TCP/UDP packet size ratios on the anomalous loss performance
and also point out the implications for TCP ACK
(acknowledgement) packet losses.
We consider UDP packet sizes of 40, 100, 200, 500, and 1000
bytes while fixing TCP packets at 1000 Bytes. TCP flows, 1000
in number, along with 5% (i.e., 5 Mb/s) Poisson UDP traffic,
are multiplexed on the dumbbell topology with a 100-Mb/s core
link. Fig. 21 shows the UDP packet loss observed in simulation
as a function of core-link buffer size for different packet size
ratios.
The figure indicates that UDP losses are higher when UDP
packets are larger (i.e., ratio of TCP to UDP mean packet
size gets smaller). This by itself is not surprising since bigger
packets constitute burstier arrivals for a given mean bit rate
(they are equivalent to a bulk arrival of several smaller packets),
and moreover larger packets are dropped in their entirety even
if a large part (but not whole) of the packet can be accommodated
in the buffer. What is, however, interesting to note in the
simulation plot is that reducing the UDP packet size makes the
anomaly more pronounced: When TCP and UDP packets have
identical sizes (1000 bytes), the anomaly is not witnessed in
simulation, but when UDP packets are only 40 bytes long, the
anomaly is quite severe.
We compare the simulation results against prediction from
our Markov chain model. We use the same kB that was
Fig. 22. M/M/1/B UDP packet loss results with varying packet size ratios.
observed in simulation of this scenario (already shown in Fig. 5)
and employ in our M/M/1/B chain a bulk arrival size equal to the
TCP/UDP packet size ratio. Fig. 22 shows the UDP loss curves
obtained via analysis for different packet size ratios. Although
the analysis and simulation curves are not a perfect match, it
nevertheless corroborates with the results obtained from simulation
and indicates the following.
1) First, the analysis predicts correctly that as the TCP/UDP
packet size ratio gets larger, losses for UDP traffic become
smaller. The bottom curve in Fig. 22 depicts losses for
40-byte UDP packets (i.e., ratio 25) and clearly shows that
loss corresponding to this packet size is significantly lower
than loss for larger UDP packets.
2) Second, the analysis seems to be fairly accurate in predicting
the inflection point. While the simulations suggest
that the inflection point occurs at 10 kB, the analysis predicts
it to happen at around 8 kB. Furthermore, that losses
(for all the packet size ratios except 25) fall again beyond
30 kB of buffers is predicted successfully by analysis.
3) Third, when TCP and UDP packets are of equal or
near-equal size, both simulation and analysis show that
the anomaly is insignificant, and the anomaly increases in
severity as the disparity in packet sizes increases.
The importance of packet size to the anomalous loss performance
also has an implication for TCP ACK packets that
are typically 40 bytes long. We therefore undertook a simulation
study of whether TCP ACK packets will also exhibit similar
anomalous behavior. We simulated 1000 bidirectional TCP
flows (without UDP) on the dumbbell topology and recorded
the ACK packet drops at routers and (see Fig. 1). The simulation
parameters are identical to the setup described earlier.
In Fig. 23, we plot the ACK packet loss probability as a function
of core-link buffer size. ACK drops in the forward direction
correspond to losses at , while the losses in the reverse direction
correspond to losses at . Clearly, ACK packets also suffer
from the anomaly and indeed match well with the analytical estimate
plotted in Fig. 22.
IEEE Proof
Web Version
VIII. CONCLUSION AND FUTURE WORK
The subject of router buffer sizing has received considerable
attention over the past few years. Researchers have questioned
the use of the rule of thumb and have argued that few tens of
packets of buffering suffice at core Internet routers for TCPVISHWANATH et al.: ANOMALOUS LOSS PERFORMANCE FOR MIXED REAL-TIME AND TCP TRAFFIC 13
Fig. 23.
Losses for TCP ACK packets from simulation.
traffic to realize acceptable link utilization. However, the research
has been primarily TCP-centric since over 85%–90%
of today’s Internet traffic is carried by TCP. Although realtime
(UDP) traffic accounts for only about 5%–10%, we note
that its popularity—through the prolific use of online gaming,
real-time video conferencing, VoIP, and many other multimedia
applications—is growing in the Internet. As such, we believe
that the study of router buffer sizing should not focus on TCP
alone, but should also consider its impact on the performance of
real-time traffic.
As a first step, in this paper, we examined the dynamics of
UDP and TCP interaction at a bottleneck-link router equipped
with very small buffers. We observed a curious phenomenon:
Losses for real-time traffic do not fall monotonically with buffer
size as one would expect. Instead, there exists an inflection point
beyond which loss increases with increasing buffer size. We
showed the existence of the anomalous loss behavior using real
video traffic traces, short-range-dependent Poisson traffic, and
long-range-dependent fBm traffic models. Furthermore, we developed
simple analytical models that gave insights into why the
anomaly exists under certain circumstances. We also presented
scenarios describing when the anomaly does not exist. Through
extensive simulations, we investigated the impact of various
factors such as fraction of UDP traffic, ratio of long-lived to
short-lived TCP flows, and packet sizes on the anomaly. The
effect of these factors on the inflection point were studied in
conjunction with the analytical models.
It is apparent that emerging optical packet switched networks
will be capable of buffering no more than a few tens of kilobytes
of data. Given this stringent constraint and the fact that
adding extra buffering adds significantly to the cost of the optical
switch, the anomalous behavior revealed in our study can
be of serious concern to optical switch manufacturers and service
providers who could make considerable investment in these
optical packet switches, only to realize worse performance if
they inadvertently operate their buffer sizes in this anomalous
region.
Several aspects of the problem require further investigation.
In our analytical models, we have assumed TCP arrivals to be
Poisson. There is scope to refine this assumption, particularly in
the presence of short-lived flows that are known to have some
long-range-dependent characteristics in their activity durations.
We can also seek to refine our Markov chain-based model to
incorporate other UDP traffic models. The impact of the ACK
drop anomaly on TCP throughput and average flow completion
times warrants a deeper understanding. Finally, we are in the
process of developing an experimental test-bed on which we
can reproduce the anomaly seen in simulations and analysis.
Commercial switches and routers today have hidden buffers
that do not permit buffer sizes to be configured as low as a
few tens of packets. For our experimentation, we are therefore
using the FPGA-based routing platform called NetFPGA (refer
Web site) recently developed at Stanford University that allows
fine-grained control of buffer sizes and has been used successfully
in their buffer-sizing studies [15]. We hope to be able to
report results from such a test-bed in the near future.
REFERENCES
[1] C. Villamizar and C. Song, “High performance TCP in ANSNet,”
Comput. Commun. Rev., vol. 24, no. 5, pp. 45–60, 1994.
[2] G. Appenzeller, I. Keslassy, and N. McKeown, “Sizing router buffers,”
in Proc. ACM SIGCOMM, 2004, pp. 281–292.
[3] G. Appenzeller, “Sizing router buffers,” Ph.D. dissertation, Stanford
Univ., Stanford, CA, 2005[Dept?].
[4] D. Wischik and N. McKeown, “Part I: Buffer sizes for core routers,”
Comput. Commun. Rev., vol. 35, no. 3, pp. 75–78, Jul. 2005.
[5] M. Enachescu, Y. Ganjali, A. Goel, N. McKeown, and T. Roughgarden,
“Part III: Routers with very small buffers,” Comput. Commun. Rev., vol.
35, no. 3, pp. 83–90, Jul. 2005.
[6] M. Enachescu, Y. Ganjali, A. Goel, N. McKeown, and T. Roughgarden,
“Routers with very small buffers,” in Proc. IEEE INFOCOM,
Barcelona, Spain, Apr. 2006, pp. 1–11.
[7] Y. Ganjali and N. McKeown, “Update on buffer sizing in Internet
routers,” Comput. Commun. Rev., vol. 36, no. 5, pp. 67–70, Oct. 2006.
[8] N. Hohn, D. Veitch, K. Papagiannaki, and C. Diot, “Bridging router
performance and queuing theory,” in Proc. ACM SIGMETRICS, New
York, Jun. 2004, pp. 355–366.
[9] G. Raina, D. Towsley, and D. Wischik, “Part II: Control theory for
buffer sizing,” Comput. Commun. Rev., vol. 35, no. 2, pp. 79–82, Jul.
2005.
[10] R. S. Prasad, C. Dovrolis, and M. Thottan, “Router buffer sizing
revisited: The role of the output/input capacity ratio,” in Proc. ACM
CoNEXT, New York, Dec. 2007, Article no. 15.
[11] A. Lakshmikantha, R. Srikant, and C. Beck, “Impact of file arrivals and
departures on buffer sizing in core routers,” in Proc. IEEE INFOCOM,
Phoenix, AZ, Apr. 2008, pp. 86–90.
[12] A. Dhamdhere and C. Dovrolis, “Open issues in router buffer sizing,”
Comput. Commun. Rev., vol. 36, no. 1, pp. 87–92, Jan. 2006.
[13] G. Vu-Brugier, R. S. Stanojevic, D. J. Leith, and R. N. Shorten, “A critique
of recently proposed buffer-sizing strategies,” Comput. Commun.
Rev., vol. 37, no. 1, pp. 43–47, Jan. 2007.
[14] M. Wang and Y. Ganjali, “The effects of fairness in buffer sizing,” in
Proc. IFIP NETWORKING, Atlanta, GA, May 2007, pp. 867–878.
[15] N. Beheshti, Y. Ganjali, M. Ghobadi, N. McKeown, and G. Salmon,
“Experimental study of router buffer sizing,” in Proc. ACM/USENIX
IMC, Oct. 2008, pp. 197–210.
[16] A. Vishwanath, V. Sivaraman, and M. Thottan, “Perspectives on router
buffer sizing: Recent results and open problems,” Comput. Commun.
Rev., vol. 39, no. 2, pp. 34–39, Apr. 2009.
[17] N. Beheshti, Y. Ganjali, R. Rajaduray, D. Blumenthal, and N.
McKeown, “Buffer sizing in all-optical packet switches,” in Proc.
IEEE/OSA OFC, Mar. 2006, Paper OThF8.
[18] D. Hunter, M. Chia, and I. Andonovic, “Buffering in optical packet
switches,” J. Lightw. Technol., vol. 16, no. 12, pp. 2081–2094, Dec.
1998.
[19] S. Yao, S. Dixit, and B. Mukherjee, “Advances in photonic packet
switching: An overview,” IEEE Commun. Mag., vol. 38, no. 2, pp.
84–94, Feb. 2000.
[20] H. Park, E. F. Burmeister, S. Bjorlin, and J. E. Bowers, “40-Gb/s optical
buffer design and simulations,” in Proc. NUSOD, Aug. 2004, pp.
19–20.
IEEE Proof
Web Version14 IEEE/ACM TRANSACTIONS ON NETWORKING
[21] E. F. Burmeister, J. P. Mack, H. N. Poulsen, J. Klamkin, L. A. Coldren,
D. J. Blumenthal, and J. E. Bowers, “SOA gate array recirculating
buffer for optical packet switching,” in Proc. IEEE/OSA OFC,
Feb. 2008, Paper OWE4.
[22] J. D. LeGrange, J. E. Simsarian, P. Bernasconi, D. T. Neilson, L. Buhl,
and J. Gripp, “Demonstration of an integrated buffer for an all-optical
packet router,” in Proc. IEEE/OSA OFC, Mar. 2009, Paper OMU5.
[23] Y. Zhang and D. Loguinov, “ABS: Adaptive buffer sizing for heterogeneous
networks,” in Proc. IEEE IWQoS, Enschede, The Netherlands,
Jun. 2008, pp. 90–99.
[24] R. Stanojevic, R. Shorten, and C. Kellet, “Adaptive tuning of drop-tail
buffers for reducing queueing delays,” IEEE Commun. Lett., vol. 10,
pp. 570–572, Jul. 2006.
[25] “The network simulator—ns-2,” [Online]. Available: http://www.isi.
edu/nsnam/ns/[Author? Year?]
[26] W. Feng, F. Chang, W. Feng, and J. Walpole, “Provisioning on-line
games: A traffic analysis of a busy counter-strike server,” in Proc. ACM
SIGCOMM IMW, Nov. 2002, pp. 151–156.
[27] W. Feng, F. Chang, W. Feng, and J. Walpole, “A traffic characterization
of popular on-line games,” IEEE/ACM Trans. Netw., vol. 13, no. 3, pp.
488–500, Jun. 2005.
[28] “Packet traces from measurement and analysis on the WIDE Internet
backbone,” [Online]. Available: http://tracer.csl.sony.co.jp/
mawi[Author? Year?]
[29] V. Markovski, F. Xue, and L. Trajkovic, “Simulation and analysis of
packet loss in video transfers using user datagram protocol,” J. Supercomput.,
vol. 20, no. 2, pp. 175–196, 2001.
[30] “Video traffic traces for performance evaluation,” [Online]. Available:
http://trace.eas.asu.edu/TRACE/ltvt.html[Author? Year?]
[31] P. Bernasconi, J. Gripp, D. Neilson, J. Simsarian, D. Stiliadis, A.
Varma, and M. Zirngibl, “Architecture of an integrated router interconnected
spectrally (IRIS),” in Proc. IEEE High Perform. Switching
Routing, Jun. 2006, pp. 71–78.
[32] L. Andrew, T. Cui, J. Sun, M. Zukerman, K. Ho, and S. Chan, “Buffer
sizing for nonhomogeneous TCP sources,” IEEE Commun. Lett., vol.
9, no. 6, pp. 567–569, Jun. 2005.
[33] G. Raina and D. Wischik, “Buffer sizes for large multiplexers: TCP
queueing theory and instability analysis,” in Proc. Next Generation Internet
Netw., Rome, Italy, Apr. 2005, pp. 173–180.
[34] J. Cao and K. Ramanan, “A Poisson limit for buffer overflow probabilities,”
in Proc. IEEE INFOCOM, New York, Jun. 2002, vol. 2, pp.
994–1003.
[35] J. Cruise, “Poisson convergence, in large deviations, for the superposition
of independent point processes,” Ann. Oper. Res., vol. 170, no. 1,
pp. 79–94, Sep. 2009.
[36] A. Vishwanath and V. Sivaraman, “Routers with very small buffers:
Anomalous loss performance for mixed real-time and TCP traffic,” in
Proc. IEEE IWQoS, Enschede, The Netherlands, Jun. 2008, pp. 80–89.
[37] R. Sinha, C. Papadopoulos, and J. Heidemann, “Internet packet size
distributions: Some observations,” USC/Information Sciences Institute,
Marina del Rey, CA, Tech. Rep. ISI-TR-2007-643, 2007 [Online].
Available: http://www.isi.edu/~johnh/PAPERS/Sinha07a.html
[38] “CAIDA Internet packet size distribution,” CAIDA, La Jolla,
CA, 2010 [Online]. Available: http://www.caida.org/research/
traffic-analysis/pkt_size_distribution/graphs.xml
[39] A. Vishwanath, V. Sivaraman, and G. N. Rouskas, “Considerations for
sizing buffers in optical packet switched networks,” in Proc. IEEE IN-
FOCOM, Rio de Janeiro, Brazil, Apr. 2009, pp. 1323–1331.
[40] V. Sivaraman, H. ElGindy, D. Moreland, and D. Ostry, “Packet pacing
in short buffer optical packet switched networks,” in Proc. IEEE IN-
FOCOM, Barcelona, Spain, Apr. 2006, pp. 1–11.
[41] V. Sivaraman, H. ElGindy, D. Moreland, and D. Ostry, “Packet pacing
in small buffer optical packet switched networks,” IEEE/ACM Trans.
Netw., vol. 17, no. 2, pp. 1066–1079, Aug. 2009.
[42] D. Ostry, “Synthesis of accurate fractional Gaussian noise by filtering,”
IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1609–1623, Apr. 2006.
[43] B. Schroeder, A. Wierman, and M. Harchol-Balter, “Open versus
closed: A cautionary tale,” in Proc. USENIX NSDI, May 2006, p. 18.
[44] W. Willinger, M. S. Taqqu, R. Sherman, and D. V. Wilson, “Self-similarity
through high-variability: Statistical analysis of ethernet LAN
traffic at the source level,” in Proc. ACM SIGCOMM, Aug./Sep. 1995,
pp. 110–113.
Arun Vishwanath (S’04) is currently a Ph.D.
candidate in the School of Electrical Engineering
and Telecommunications, University of New South
Wales, Sydney, Australia.
He was a visiting Ph.D. student with the Department
of Computer Science, North Carolina State University,
Raleigh, in 2008. His research interests include
energy-efficient network architecture and design,
optical networks, and router buffer sizing.
Vijay Sivaraman (M’94) received the B.Tech. degree
from the Indian Institute of Technology, Delhi,
India, in 1994; the M.S. degree from North Carolina
State University, Raleigh, in 1996; and the Ph.D. degree
from the University of California, Los Angeles,
in 2000, all in computer science.
He has worked with Bell Labs[City?] and a
Silicon Valley startup manufacturing optical switchrouters.
He is now a Senior Lecturer with the University
of New South Wales, Sydney, Australia. His
research interests include optical networking, packet
switching, and QoS routing.
George N. Rouskas (S’92–M’95–SM ’01) received
the Diploma in computer engineering from the
National Technical University of Athens (NTUA),
Athens, Greece, in 1989, and the M.S. and Ph.D.
degrees in computer science from the College
of Computing, Georgia Institute of Technology,
Atlanta, in 1991 and 1994, respectively.
He is a Professor of computer science with North
Carolina State University, Raleigh. He is co-editor
of the book Traffic Grooming for Optical Networks:
Foundations, Techniques and Frontiers (Springer
2008). His research interests include network architectures and protocols,
optical networks, multicast communication, and performance evaluation.
Dr. Rouskas is a Member of the Association for Computing Machinery
(ACM) and the Technical Chamber of Greece. He is a recipient of a 1997
National Science Foundation (NSF) Faculty Early Career Development (CA-
REER) Award.
IEEE Proof
Web Version