Transcript TCP and UDP Performance Over A Wireless LAN

TCP and UDP Performance
Over A Wireless LAN
Professor：柯開維老師
Speaker：許家豪
Date：2004/04/16
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Our Goals
In order to ameliorate WLAN performance
problems we need a clearer understanding
of WLAN behavior and analyzing it.
Our aim was to compile a comprehensive
set of data describing the performance of
a WaveLAN system in terms of throughput
and loss under various realistic conditions.
Network Performance
Be Influenced by：(difficult to perceive)
1、Network and Host Processing
Hardware.
2、Interface Device Drivers.
3、Network Protocol Implementation
in the OS.
Methods :
We aims to extend published results in
many ways
1、System Heterogeneity
2、New Implementations
3、Bidirectional Communications
4、Error Modeling
5、Operating System
Methods Description (1)
1. We used hosts with varying processing
power and different wireless interface
implementation.
2. Use 2.4 GHz version. Also used faster
processors that could potentially achieve
higher throughputs.
3. We measured the performance of TCP, in
addition to UDP.
Methods Description (2)
4. We present additional measurements
and also analyze bidirectional traffic
effects.
5. We employed the Linux OS instead of
BSD UNIX derivatives used in previous
work.
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Hardware
Buffer of PCMCIA cards has single buffer.
Buffer of ISA cards have multiple buffers
Equipment
Digital RoamAbout 2.4GHz DSSS system,
an OEM version of the Lucent WaveLAN.
IOS (better) and MYKONOS (poor) are
desktops（PC）.
SYROS is a laptop , and its processor
operates at a lower clock frequency but has
more on_chip cache memory.
Use CAMA/CA to handle collision problems.
Software (1)
All hosts ran the Linux OS, in multiuser
mode, but with no user tasks executing,
testing time is late in the evening.
We made a minor modification to the
wireless interface drivers to record and
report detailed statistics plus histograms
of signal and noise levels.
Software (2)
Two benchmarks were used
1、TTCP：
Sends a number of packets of a specified
size to a receiver using either TCP or UDP.
2、ETTCP (For UDP test) ：
Uses packet sequence number so that
the receiver can detect and report packet
losses.
Software (3)
Use nstat to gather IP, UDP and TCP
statistics aggregated across all interfaces
to check for unexpected network activity
during the tests.
Use tcpdump to record detailed logs of all
packets set and received by the wireless
interfaces during each tests.
Location map
45
45
60
feet
Environment
A floor plan of the area (5th floor) where
the experiments took place.
All rooms are laboratories and machine
rooms containing numerous hardware
devices but no direct sources of
interference.
Hosts were kept immobile during each test
to avoid mobility induced problems.
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Testing Scenario
Testing Methods (1)
The main test parameters were transfer
direction, peer names, packet size,
protocol.
A test script first reset and dumped
interface statistics and nstat output, then
started tcpdump to record all packets
through the wireless interface, and finally
started ettcp to transfer 10000 packets.
Testing Methods (2)
All tests were performed in both directions
between the peers to reveals any
performance asymmetries.
All tests were performed for both TCP and
UDP, with four IP datagram sizes：100, 500,
1000 and 1500 bytes to show the effects of
varying amounts of overhead and packet
error probability on throughput.
Testing Methods (3)
Throughput and loss rate are
comparable across all tests since their
units are independent of packet size.
These can be used to determine the
optimal packet size where overhead and
loss are best balanced.
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Scenario 1
Two hosts were placed next to each other
to avoid signal degradation.
Goal：
It was to determine the peak performance
of ISA cards and reveal processing power
induced asymmetries.
Result (1-1)
Receiver view
Result (1-2)
When the slower host (MYKONOS) is
sending, packet loss is negligible. In
contrast, the faster host (IOS)
overwhelms a slower receiver, leading
to loss (0.3-0.6%) which grows with
packet size.
Result (2-1)
Receiver view
Mykonos
Ios
Result (2-2)
Net UDP throughput increases with
packet size since UDP/IP overhead drops.
TCP throughput is not only below UDP, it
actually drops with large packet sizes.
Slower sender makes fewer losses.
Result (3-1)
Data sent (I)
Data sent (M)
Data received (M)
Data received (I)
Result (3-2)
The gaps between sent and received
curves for both packet types show
considerable loss on the link, growing
with packet size.
Since the gaps are roughly the same,
we conclude that their magnitude
represents the number of undetected
collisions of CSMA/CA.
Result (4-1)
Result (4-2)
Result (4-3)
A collision occurs when one data and
one acknowledgment packet are shown
on the sender but not on the receiver.
Result (5-1)
Result (5-2)
Both histograms are nearly symmetric
since the peer interfaces were exactly
the same and host processing power
does not influence the radios.
Scenario 2
It employs one ISA and one PCMCIA
host, again placed next to each other,
to establish a performance baseline for
mixed interface tests.
The processing power of SYROS and
MYKONOS is roughly equivalent, thus
comparisons with the first scenario are
direct.
Result (1-1)
Result (1-2)
When the faster host is sending, implying
that both ISA and PCMCIA receivers are
overwhelmed by faster senders.
In Syros to Ios direction, the perceived
losses are due to packets never leaving
the sending interface.
Result (2-1)
Ios to Syros (UDP)
Ios to Syros (TCP)
Syros to Ios (UDP)
Syros to Ios (TCP)
Result (2-2)
In the IOS(ISA) to the SYROS(PCMCIA)
direction UDP is faster than TCP, due to
less header overhead and the absence
of TCP retransmissions and
acknowledgments.
TCP throughput in the reverse direction
is slightly lower, verifying previous
claims that PCMCIA cards are slower
senders.
Result (3-1)
Result (3-2)
Result (3-3)
Sequence numbers increase faster with an
ISA sender despite occasional
retransmissions. The PCMCIA sender
leaves short gaps between transmission
bursts due to the transmit buffer shortages
PCMCIA card has single transmit buffer,
and ISA has multiple buffers. So PCMCIA is
easy to overrun.
Scenarios 3 & 4 (1)
Scenarios 3 :
The same as 1, just indicating again
that a faster sender overruns a slower
receiver. But signal level are uniformly
lower.
Scenarios 4 :
The same as 2. But signal level are
uniformly lower.
Scenarios 3 & 4 (2)
We conclude that this distance and
obstacles do not have measurable
effects on performance in both ISA/ISA
and ISA/PCMCIA tests.
Scenarios 5
The main difference with 2 & 4 is a
nearly zero loss rate in the ISA to
PCMCIA direction. This shows that hosts
matched in processing power avoid
losses due to receiver overruns.
Scenario 6
The only difference with scenario 4 is a
slightly higher loss rate, both due to the
increased distance and obstacles. And
the signal level is lower than scenario 4.
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Conclude
Fast senders can overwhelm slower
receivers (both ISA and PCMCIA),
leading to semi-periodic packet loss.
Outline
Introduction
Experimental Setup
Testing
Analysis Of Test results
Conclude
Reference
Reference
TCP and UDP Performance over a Wireless
LAN ( George Xylomenos and George C.
Polyzos )