Dynamic right-sizing in TCP : a simulation study / Page: 4 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:
sizes of 100, 500, and 16,384 packets in routers R1 and R2. The
latter queue size represents an ideal router where packets are
never dropped. Host buffers are set to not drop packets locally.
All nodes used a drop-tail queueing algorithm; use of random
early detection (RED) gateways did not noticeably change the
For static flows, we set fwnd to the default of 64KB. All
flows are one-way, constant bit rate (CBR) transmissions at the
link speed (100 Mbps or 1000 Mbps) from sources Si to desti-
nations Dj where i, j E [1, n]. Use of other traffic models such
as Poisson or Pareto, or the inclusion of reverse path traffic [141
did not produce significantly different results. We believe that it
is unlikely that traffic patterns alone would cause significant dif-
ferences between TCP with and without dynamic right-sizing.
IV. RESULTS AND ANALYSIS
The experimental results in this section will show that TCP
flows with dynamic right-sizing outperform flows without it.
As a uniform basis for comparison, we track the delivered
bandwidth at the receiver (as measured by the number of ac-
knowledgement packets returned to the sender) as a function of
maxwnd. We also consider performance with respect to link
utilization, packet loss, and fairness among flows.
We structure our discussion as follows. We first discuss the
current situation on the Internet where all flow-control windows
are static in size, i.e., static flows. We then consider a head-to-
head comparison of static flows vs. dynamic flows, i.e., dynami-
cally right-sized flows. Lastly, we evaluate situations that would
exist with incremental adoption of dynamic right-sizing.
A. Static Only
Over the 100Mbps topology, at most 1,666,666 packets can
be sent during our 200-second simulation. This is shown as the
maximum value on the y-axis in Figure 2. Each point in these
figures represents an independent TCP simulation for the given
maximum congestion window. The use of the line segments
to connect the points represents the expected continuity when
simulations are run at intermediate points (which we validated
by several tests). For the 1000Mbps simulation, the value is ten
Fig. 2. Static Flow Only
The results in Figure 2 clearly illustrate the inadequacy
of static flow-control windows in a WAN. In this case, the
static flow throttles the bandwidth to 145,527 _ 8.7% and
1,66,s6 0.55% of what is available over the 100Mbps and
1000Mbps topologies, respectively! This is obviously atrocious
and motivates the current laborious practice of hand-tuning con-
nections. Using dynamic right-sizing would automatically in-
crease f wnd, and hence, ewnd appropriately, without the waste
of memory that manual tuning would incur.
The above behavior is clarified by the following analysis.
With a fixed 64KB fwnd (i.e., 44 standard Ethernet packets),
ewnd cannot grow beyond 64KB even though cwnd can. Thus,
the performance of the static flow is uniform no matter how
large cwnd grows beyond 64KB. And with the bandwidth-delay
products of the 100Mbps and 1000Mbps networks being at least
500 and 5000 packets, respectively (more with realistic queue-
ing delays), the static fwnd throttles bandwidth to at most
0 ^ 8.8% and 5 "' 0.88% of what is available over the
100Mbps and 1000Mbps topologies, respectively. These per-
centages decrease further with increased bandwidth-delay prod-
ucts caused by realistic queueing delays.
B. One Static vs. One Dynamic
Figure 3 presents results for one flow of each type and queue
sizes of 100 and 500 packets over the 100Mbps network. We
vary fwnd from 50 packets to 700 packets, sizes correspond-
ing to window sizes of 73KB (i.e., just larger than the static
flow window size of 64 KB) up to 1 MB. This case most clearly
shows the interaction between flows using static and dynamic
0 100 200 300
400 500 600 700
Fig. 3. Static vs. Dynamic Flow
The upper two lines are dynamic flows while the lower two
are static flows. When the rate limiting maxwnd is relaxed,
the dynamic flows deliver over 90Mbps throughput, while the
static flows remain unaffected and deliver a paltry 8.7 Mbps of
No differences based on queue size are evident until maxwnd
reaches 650 packets. There, for a queue size of 100 packets, the
dynamic flow overflows the router's buffer and induces massive
packet loss. The dynamic flow cuts its sending rate due to these
0 100 200 300 400 500
Maximum Window (packets)
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/m1/4/: accessed April 19, 2019), University of North Texas Libraries, Digital Library, https://digital.library.unt.edu; crediting UNT Libraries Government Documents Department.