Dynamic right-sizing in TCP : a simulation study / Page: 3 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:
II. DYNAMIC RIGHT-SIZING
Ideally, the transmission window (ewnd) should be limited
only by the congestion window (cwnd). Recall that ewnd
min(f wnd, cwnd), so network utilization is maximized when
fwnd > cwnd, and memory use is most efficient when
fwnd = cwnd.
For simulation purposes, we could simply set f wnd =
cwnd. In our implementation, however, we provide some lee-
way (Icwnd- f wnd <6) due to the nonzero overhead of mem-
ory allocation and deallocation. For 6 sufficiently small, we can
reproduce the performance of dynamically sizing flow-control
windows in these simulations.
We refer to the process of dynamically sizing flow-control
windows as dynamic right-sizing  and use the algorithm be-
low to implement it.
Let awnd = receiver's advertised window (which is used
to set f wnd at the sender)
cwnd = sender's congestion window
maxwnd = maximum amount of memory that any
one connection can utilize.
At the source,
fwnd = ?nin(cwnd, awnd, maxwnd).
At the destination,
awnd > 2 x cwnd if sender in slow start
awnd > cwnd + 1 if sender in additive increase.
We modify the sender only by adding the parameter
maxwnd, limiting the sending rate when administrators are un-
willing to dedicate the resources required. With this formula,
maxwnd can be viewed as the maximum flow window or the
maximum congestion window. We use maxwnd to simulate
the problem of sharing limited memory among multiple con-
nections. In practice, it is unlikely to be used; instead a fairness
algorithm would control memory allocation among competing
The receiver must infer the TCP state and congestion window
of the sender. This can be done by observing time stamps on
received packets and calculating mean transfer rates over the last
several RTTs. To conserve memory, awnd can be reduced a few
RTTs after a multiplicative decrease is detected (enough time
for packets buffered in the network to reach the receiver).
A. Benefits of Dynamic Right-Sizing
The major benefits of dynamic right-sizing include improved
memory and network performance. Our approach also gives full
and valid interoperability with other TCP implementations and
transparency to the user (i.e., no complex administration) and is
also provably fair.
Other approaches either affect TCP semantics  or are spe-
cialized non-TCP protocols that are obviously not interopera-
ble, particularly over the wide-area network (WAN), e.g., VIA
. For significant performance improvements, TCP imple-
mentations require the window-scaling extensions , but this
technique works without them.
Often, system administration is just an irritating inconve-
nience, but sometimes it is more than that. When firewalls are
involved, connections between hosts on opposite sides are of-
ten broken into two connections - one from the source to the
firewall and another from the firewall to the destination. So, re-
gardless of the settings at the hosts, the firewall will be the bot-
tleneck. Because firewalls are typically controlled by security-
minded administrators, changing the firewall configuration can
be difficult. If the firewall is under the control of a different insti-
tution, it may simply be impossible. Using dynamic right-sizing
(as part of TCP) in the firewall makes this problem disappear.
Our approach is fair with respect to TCP Reno; we implement
dynamic right-sizing in TCP Reno and do not alter its additive-
increase, multiplicative-decrease (AIMD) mechanism. Special-
ized protocols like VIA cannot make this claim. Although flows
using dynamic right-sizing acquire more bandwidth than those
without it, that does not make it unfair. Flows that properly
configure both endpoints should receive better performance than
misconfigured flows. Dynamic right-sizing automatically makes
the proper adjustments to flow-control windows, so we expect
better performance with dynamically right-sized flows than for
generally misconfigured static flows.
More rigorously, we use the following measure of fairness:
Flows should acquire bandwidth proportional to the resources
(in our case, memory) they dedicate to the connection. Thus,
even if dynamic right-sizing reduces the share of bandwidth uti-
lized by static flows, dynamic right-sizing would still be fair
(since all the static flows would have to do to compete is to in-
crease the memory allotted to their flow-control windows).
In this study, we simulate and analyze three general cases
that show the effects of incremental adoption of dynamic right-
. All connections use static flow-control windows, i.e., today's
. Some connections use static flow-control windows while oth-
ers use dynamic flow-control windows.
. All connections use dynamic flow-control windows.
Figure 1 shows the generic topology over which we run our
simulation experiments. Links run at 100Mbps with 10ms delay
in one set of experiments; in the other 1Gbps with 16ms delay.
The RTTs for these experiments are 60ms and 96ms, respec-
tively, representative of physically long connections present in
the global Internet.
S3 RI R2 D3
Fig. 1. Generic Topology
Buffer space available in network queues heavily influences
TCP Reno's performance. This is due to the bursty nature of the
traffic generated by TCP Reno ,  causing dramatic varia-
tion in queue lengths. To study these effects, we simulate queue
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/m1/3/: accessed April 24, 2019), University of North Texas Libraries, Digital Library, https://digital.library.unt.edu; crediting UNT Libraries Government Documents Department.