Dynamic right-sizing in TCP : a simulation study / Page: 2 of 8
This article is part of the collection entitled: Office of Scientific & Technical Information Technical Reports and was provided to Digital Library by the UNT Libraries Government Documents Department.
The following text was automatically extracted from the image on this page using optical character recognition software:
Dynamic Right-Sizing: A Simulation Study
Eric Weigle and Wu-chun Feng
Research & Development in Advanced Network Technology (RADIANT)
Los Alamos National Laboratory
Los Alamos, NM 87545
Abstract- Virtually all network applications requiring reliable end-to-
end communication depend on TCP. Unfortunately, the performance of any
stock TCP is abysmal over wide-area networks (WANs) and even over local-
area networks (LANs) with very high-bandwidth links. Currently, network
researchers manually optimize TCP buffer sizes to achieve acceptable per-
formance over a given connection. Unfortunately, this manual optimization
requires changes to the kernel on both end hosts involved in the network
connection (changes that are only effective for connections between these
two hosts). Furthermore, because two administrative domains must be co-
ordinated to perform this optimization, this process can be tedious and time
To address these problems, this paper illustrates the benefits of a new
technique called dynamic right-sizing. This technique dynamically and au-
tomatically determines the best buffer size, and hence flow-control window
size in TCP. Our simulation study shows that dynamic right-sizing can im-
prove the performance of flows by two orders of magnitude over stock TCP
implementations that have static flow-control windows.
Keywords: dynamic right-sizing, high-performance network-
ing, TCP, flow control, wide-area network.
Over the past decade, TCP has become the ubiquitous trans-
port protocol for the Internet.' However, stock TCP performs
abysmally over high-bandwidth or high-delay links. As a result,
the performance of application infrastructures such as computa-
tional grids  and high-volume web servers, which are built on
TCP, is crippled.
To address this problem, grid and network researchers con-
tinue to manually optimize buffer sizes to keep the network pipe
full, and thus achieve acceptable performance , . Although
such tuning can increase delivered throughput by an order of
magnitude, it requires kernel-configuration changes that cannot
be made by the end user. Instead, system administrators at the
source and destination hosts must separately configure their sys-
tems to use large buffers, a tedious and time consuming process.
Furthermore, manual tuning only works for the pair of hosts that
are tuned. We propose a straightforward modification to TCP
that automatically and transparently addresses the above prob-
lems while maintaining connection semantics and the ubiqui-
tously deployed features of TCP.
To this end, we first briefly discuss TCP, focusing on its flow-
and congestion-control mechanisms and problems with current
implementations. We then discuss our approach and its implica-
tions, followed by our experiments and results. We finish with
related work and conclusions.
'while there are many versions of TCP, we focus on TCP Reno as it is the
most commonly used and widely deployed variant. Hereafter, we mean TCP
Reno when we refer to TCP.
A. TCP Flow and Congestion Control
TCP relies on two mechanisms, flow and congestion control,
to set its transmission rate. Flow control ensures that the sender
does not overrun the receiver's available buffer space while con-
gestion control ensures that the sender does not unfairly over-
run the network's available bandwidth. TCP implements these
mechanisms via a flow-control window (fwnd) and congestion-
control window (cwnd).
Specifically, TCP calculates an effective window (ewnd) as
ewnd min(f wnd, cwnd) and then sends data at a rate of
ewnd/fRTT, where RTT is the round-trip time of the connec-
tion. While the cwnd varies dynamically as the network state
changes, fwnd has always been static. Ideally, fwnd should
vary with the bandwidth-delay product of the network.
B. The Failure of TCP
Historically, a static fwnd sufficed as all communication
occurred over networks with low bandwidth-delay products.
Setting fwnd to small values allowed acceptable performance
while wasting little memory. Today, most operating systems
set f wnd 64KB, the largest window available without
scaling . Yet bandwidth-delay products range between a
few bytes (56Kbps x 5ms -+ 36B) and a few megabytes
(622.08Mbps x 100ms -+ 7.8MB). In the first case, we can
waste over 99% of memory allocated (36B/64KB = 0.05%).
In the latter case, we waste 99% of the network bandwidth
(64KB/7.8MB = 0.80%).
During the life of a connection, network delay changes
(due to transitory queueing and congestion) implying that the
bandwidth-delay product also changes. Thus a fixed value for
fwnd is never ideal; selecting a fixed value forces an implicit
decision to (1) underallocate memory and underutilize the net-
work or (2) overallocate memory and waste system resources.
Clearly, the solution is to dynamically and transparently adapt
f wnd to achieve good performance without wasting network or
TCP suffers from problems other than the static fwnd, in-
cluding self-similar (chaotic) behavior , ,  and a slowly-
converging, additive-increase period.2 These congestion-control
problems are serious and will appear in our results, but they are
orthogonal to the problem that we addressing here, namely static
flow-control windows. However, we are working on solutions to
congestion control, as are others , , .
2For a 8MB bandwidth-delay product, convergence to the optimal band-
width can take as long as (8MB x t) x 10242 -/1500 s x 1 x
0.t00U- = 266s, or nearly four and a half minutes!
Here’s what’s next.
This article can be searched. Note: Results may vary based on the legibility of text within the document.
Tools / Downloads
Get a copy of this page or view the extracted text.
Citing and Sharing
Basic information for referencing this web page. We also provide extended guidance on usage rights, references, copying or embedding.
Reference the current page of this Article.
Weigle, E. H. (Eric H.) & Feng, W. C. (Wu-Chun). Dynamic right-sizing in TCP : a simulation study /, article, January 1, 2001; United States. (https://digital.library.unt.edu/ark:/67531/metadc930410/m1/2/: accessed April 19, 2019), University of North Texas Libraries, Digital Library, https://digital.library.unt.edu; crediting UNT Libraries Government Documents Department.