- Ultimate Goal:
- We want to transmit (transfer) a (large) file from point A to point B as FAST as possible
- How to achieve the goal:
- Minimize the
number of retransmission
used by the sender
in the act of data transfer
- This is achieved by:
- Avoid buffer overflow condition in routers
- Minimize the
number of retransmission
used by the sender
in the act of data transfer
- Congestion control
is avoid causing buffer overflow
in routers
while transmitting as fast as possible
Congestion:
- When an output buffer is at or near capacity (i.e., it can overflow anytime), the state is called congested
Congestion control is also known as congestion avoidance
- Trivial congestion avoidance:
- It is trivially simple
to avoid buffer overflow condition
by transmitting very slowly
- That's why there is the condition that the sender must:
- transmit as fast as possible without causing congestion
- It is trivially simple
to avoid buffer overflow condition
by transmitting very slowly
- The basic principles
of congestion control is simple:
- Decrease
your transmission rate
if you detect
congestion
- Increase your transmission rate if you detect no congestion
- Decrease
your transmission rate
if you detect
congestion
- A couple of $64,000 questions:
- How do you
detect congestion ?
- How fast
should you decrease
your transmission rate
if you detect
congestion
- How fast should you increase your transmission rate if you detect no congestion
- How do you
detect congestion ?
- You need to
solve 2 problems
in order to
design
a
congestion control method:
- How to detect
an onset or presence of
congestion
- How fast
do you:
- decrease transmission rate when you do detect congestion... and
- increase transmission rate when no congestion is detected
- How to detect
an onset or presence of
congestion
- Network congestion
is detected by using
one or more "signals"
from the network...
- Type of signals
used to detect congestion:
- Explicit signals:
- An explicit signal
is a value that is sent
inside a packet
- Example:
-
In the DECBit method,
the packet header contains a
congestion bit
When the packet is sent by the sender, this bit is reset
When the packet passes through a congested router, (the queue length exceeds a certain threshold) the router will set the congestion bit
NOTE: Explicit signals requires the assistence of routers and therefore, it is NOT useful practically because present day Internet routers do not flip bits in packets...
- An explicit signal
is a value that is sent
inside a packet
- Implicit signals:
- An implicit signal
is a value that is
derived (concluded)
from
events.
The signal is not convey through packets
- Examples:
- Packet loss
-
Congestion is detected when the sender detects that
some packet was lossed
Packet loss can be detected through timeout:
Or by duplicate ACKs:
- Packet delay
-
The sender measures the
time for the ACK of packets to
return back.
When the time exceeds a certain threshold, the source assumes congestion.
- Packet loss
- An implicit signal
is a value that is
derived (concluded)
from
events.
- Explicit signals:
- Due to the fact that routers in the Internet do NOT provide explicit signals, we will only consider congestion control using implicit signals
- Fact:
- Today's network links are
extremely reliable
(fiber optics)
- When a packet is lossed,
the usual cause is
packet drop due to buffer overflow.
- In other words:
- Packet loss is a very reliable signal that the network path is congested
- Today's network links are
extremely reliable
(fiber optics)
- However:
- The packet loss signal is the worst possible feedback signal that you can use !!!
Reason:
- The
purpose
of
congestion avoidance
is to
avoid congestion
- The loss of a packet
is in fact an indication
that the network
is already congested
I.e., it's too late !!!
Because a buffer at some router has overflowed !!!
What we actually want to achieve is to avoid the buffer being filled up completely...
- Congestion control using
packet drop indication
is somewhat akin to a
medical situation
where the
patient is having a heart attack:
- What we really want
is to prevent a heart attack
- What we are really doing is: fix the patient when the heart attack is detected :-)...
- What we really want
is to prevent a heart attack
- NOTE:
- Because "it was too late", the congestion control measure taken by the sender when packet loss is detected will have to be very drastic (discussed later)
- The End-to-end delay
is a far better indication
than packet loss
for
congestion avoidance:
By measuring the end-to-end delay, the source can detect an impeding collapse and take measure to reduce the traffic before the collapse happens:
- However:
- the End-to-end delay signal
is
harder to use
It is much more difficult to determine if the buffer of a router is about to overflow by observing the end-to-end delay...
- The obvious algorithm:
- When the delay
increases, there is
potential network congestion
Problem: you cannot accurately tell from the increase in the delay how closely a buffer is from overflowing.
- When the delay
increases, there is
potential network congestion
Example:
- When the drain rate
is high (1 Gbps),
a small increase in delay can correspond to
a buffer overflow:
- The additional buffer space occupied is
5 Mbits
- The additional delay inncured is only 0.005 sec
- The additional buffer space occupied is
5 Mbits
- Now consider a
lower drain rate:
- The additional buffer space occupied is ONLY
2 Mbits
- But the additional delay inncured is 0.02 sec
- The additional buffer space occupied is ONLY
2 Mbits
Conclusion:
- The increase in delay alone is not sufficient to make any conclusion on congestion....
- the End-to-end delay signal
is
harder to use
- There are many different
TCP protocols proposed.
Most common TCP protocols:
- TCP Tahoe (1982, Van Jacobson)
- TCP Reno - Tahoe with fast recovery (most commonly used version today)
- TCP Vegas - uses delay as congestion indication
- I will discussed TCP Tahoe and
TCP Reno in this course
These methods use packet loss as congestion indication
- I will not discuss delay based
congestion methods in this course.
If you want to explore this topic, here is a great paper on delayed based congestion control by the poineer in congestion avoidance:
- Raj Jain (he was the poineering expert in congestion research) paper: click here
- TCP Vegas uses delay as congestion indication