The Time Delay of TCP/IP and Related Protocols - mechatronics

The Time Delay of TCP/IP and
Related Protocols

Zhen Song

Motivation: why study TCP/IP


It is popular and reliable.


Low cost. (Both software and hardware)


Scalable. (WAN + LAN)


Platform independent. (Wire/Wireless,
Connection/Connectionless, Different OS,
Various of hardware)


Potential applications: LonWorks,
Teleoperation.


TCP/IP stack layout


Lan adapters

Lan drivers

IP

ICMP

Routing

ARP

TCP

UDP

Other protocols

applications

applications

applications

Wire/fiber

Other protocols

From http://netlab1.usu.edu/pub/bsuk99/tcp.ppt

TCP/IP vs. OSI

OSI: Open System Interconnection

Related Protocols

Figure from A.S. Tanenbaum “Computer Networks” 3ed Edition

For most PC systems


Data link: Ethernet, ISDN/ADSL/Telephone/Internet.


Network: IP, ARP.


Transport: TCP, UDP.


Application: WWW, TELNET, FTP, SMTP, DNS.



Delay of The Ethernet
(IEEE 802.3)


Data Link layer, Medium Access sublayer.


Performance


If there are k stations ready for transmit and the probability for each
station transmits during a contention slot is p. The probability A that some
station acquires the channel in that slot is A=k p(1
-
p)
k
-
1
.


A is maximized when p=1/k, with A
-
> 1/e as k
-
>Inf, so the mean number
of slots per contention is



Each slot has a duration 2t, the mean contention interval, w, is 2t/A.
Assuming optimal p, the mean number of contention slots is never more than
e, so w is at most .


Cont.


If the mean frame takes P sec to transmit, when many stations have frames to
send,

Channel efficiency =



Larger frames are more efficient, but others have
to wait longer before they can get access to the
channel. 1024 bytes is a round off between
performance and “fairness”.


The trouble maker for controller: the “mean” not
“worst” case.


Channel Efficiency

1024 byte frames

512 byte frames

256 byte frames

128 byte frames

64 byte frames

Number of stations trying to send

Channel efficiency

Figure from A.S.
Tanenbaum
“Computer Networks”
3ed Edition

Collision of The Ethernet


For a mean interval of 2t/A, collision happens
and Binary Expenential Backoff (BEB) procedure
starts.


BEB: after i collisions, a random number between
0 and 2
i
-
1 is chosen as the number of slots to
skip. The maximum number of the skipped slots
is 1023.


The possibility of collision after the last round of
BEB is negligible for a LAN, but significant
delay might be introduced.

Go to WAN


ARP: Ethernet address (MAC) to Internet address (IP).
(Defined in RFC826)


Why: Every computer on the Internet has a unique IP address, which is not the address
of Ethernet. So there must be a mechanism to translate between them.

How: Broadcast “Who owns the IP address 129.123.85.29”


Acknowledge “I have, and my MAC address is …..”

Note: ARP is not routable.


The commercial implementations are optimized by
various way, including ARP cache. So the delay of ARP
introduced only once.


ARP is not mandatory. A system with static database does
not need ARP. ARP was designed to reduce the work load
of people.


IP Routing


Simple IP routing

The Netmask can judge if a specific IP is in the associate
subnet.


If (((their_IP ^ my_IP) & netmask) )!=0




use_gateway( ); /* not in this subnet */


else


go_direct( ); /* on the same wire */


This job would be done by the router, with
unknown delay. Depends on the CPU, software,
load, etc of the router.



More complex routing


The interior gateway routing protocol:OSPF

RFC1247.

OSPF represents Open Shortest Path First.

Original Internet interior gateway protocol was a distance vector protocol
(RIP) based on the Bellman
-
Ford algorithm. It worked well in small systems,
but less well on larger ones. OSPF became a standard in 1990, and it will
become the major interior gateway routing protocol soon.

Unknown delay


The exterior gateway routing protocol:BGP

RFC1265

BGP: Border Gateway Protocol

Not just the shortest path, It considers more about politics.


Traffic starting or ending at Netscape should not transit Microsoft.

Congestion Control in The
Network Layer


John Nagle mode.
“
Congestion Control in IP/TCP
” RFC896


If there are infinite memory on the router, or the
CPU of router is too slow, congestion may
happen.


General principle of congestion control: using
“control theory point of view”, “close loop”.


There are many algorithms, such as

Nagle mode, Leaky bucket, Token bucket, etc.

DNS



From domain name to IP address

www.yahoo.com

-
> 126.115.102.75


-
> 126.115.102.76

www.sohu.com

-
> 61.135.131.13

www.virtualhost.com

-
> 159.226.35.24

www.anothervirtual.com

-
>
159.226.35.24



TCP congestion control



Jacobson algorithm

TCP updates an estimate of the average
RTT (round trip time) a by

,where g is a ‘gain’ (0<g<1) that
should be related to the signal
-
to
-
noise ratio of m.

Measuring The Delay Is Simple

Just Ping

Conclusion


Consider the worst case for either LAN, such as
Ethernet, or WAN, is very difficult.


There is possibility mathematic model for the
collision of the Ethernet. But some parameters for
the formula cannot be calculated, rather, they are
measured on the real system. Not surprisingly,
different systems might have quite different
delay.

Cont.


The delay is very easy to measure but hard to
predict.


The more you know about the delay about other
computers, the easier to predict the delay of your
system.


Though it has not been proved yet, the delay,
even of LAN, must be uncontrollable, if other
computer involved in the same network.


High performance indicates small delay. For a
reliable medium like twist par 100M Ethernet,
large frames will have better performance.

Cont.


Unfortunately, control systems prefer small
frame. So we should not put heavy load on the
network. Efficiency theory gives us a reference
on selection the working load.


At the beginning, researchers assumes the traffic
of the Ethernet is Poisson, but it now appears the
network traffic is
self
-
similar
(
Paxson and Floyd
1994; and
Willinger et al
., 1997)


Internet tracking software.

http://www.sane.com/

http://www.deepmetrix.com/

References


Classics

352 citations
Wide
-
Area Traffic: The Failure of Poisson Modeling

254 citations
Self Similarity Through High
-
Variability: Statistical Analysis of
Ethernet LAN Traffic at the Source Level

148 citations
End
-
to
-
End Routing Behavior in the Internet


A good presentation

Why is the internet traffic self
-
similar?


30 citations
Forecasting network performance to support dynamic scheduling
using the network weather service



Papers on 24th Conference on Local Computer Networks

17
-

20 October, 1999 Lowell, Massachusetts

Time Series Models for Internet Data Traffic

C. You and K. Chandra

On the Distribution of Round
-
Trip Delays in TCP/IP Networks

T. Élteto and S. Molnár

Efficient and Accurate Ethernet Simulation

J. Wang and S. Keshav

The Worst
-
Case Scenario for Transmission of Synchronous Traffic in an FDDI Network

S. Zhang and E. Lee


Cont.


Papers on frequency domain.


Modeling Spectral Features in TCP traffic

G. Olowoyeye, B. Kim and K. Chandra 1998


Network traffic modeling using a multifractal wavelets model


Non
-
linear time
-
series models for Ethernet traffic


K. Chandra, C. You, G. Olowoyeye and C. Thompson , June 1998


Analyzing Stability in Wide
-
Area Network Performance


Measuring link bandwidths using a deterministic model of packet delay,


Characteristics of wide
-
area TCP/IP conversations,