|
Previous
|
Content
|
Next
|
|
|
10.0.- On Traffic Phase
Effects in Packet-Switched Gateways |
|
|
|
|
Sally Floyd and Van Jacobson wrote "On Traffic Phase
Effects in Packet-Switched Gateways" in 1991. The old original
paper is a very large document trying to explain some alien effect that
occur in packet-switched networks transporting TCP/IP when
using DropTail gateways or routers. The alien effect causes
some TCP connection being starved by some other, by a
timely effect created in the router's queues. It is very improbable that
this effect is present in the actual Internet, being this a highly
random network, but because the effect can be presented in some simulation
configurations, it is good enough to know about this and how it can affect
the TCP behavior when making tests or investigations based on
simulation. |
|
Introduction |
|
Gateway algorithms for congestion control and avoidance are
frequently developed assuming that incoming traffic is "random"
(according to some probability distribution). However, much real network
traffic, such as bulk data transfer shown in Figure 1, has a strongly
periodic structure. For a particular connection the number of
outstanding packets is controlled by the current window. When the sink
receives a data packet it immediately sends an acknowledgment (ACK)
packet in response. When the source receives an ACK it immediately
transmits another data packet. Thus the roundtrip time (including
queueing delays) of the connection is the traffic "period". |
|
|
|
Most current network traffic is either bulk data transfer (i.e., the
total amount of data is large compared to the bandwidth-delay product
and throughput is limited by network bandwidth), or interactive
(i.e., transfers are small compared to bandwidth-delay product and/or
infrequent relative to the roundtrip time). We refer to the former as
FTP traffic having a periodic structure and the latter as Telnet
traffic having a Poisson sources to model it. By random
traffic we mean traffic sent at a random time from a telnet
source. |
|
Consider FTP traffic with a single bottleneck gateway and a
backlog at the bottleneck. When all of the packets in one direction are
the same size, output packet completions occur at a fixed frequency,
determined by the time to transmit a packet on the output line. For example,
the following is a schematic of the packet flow in figure 1: |
|
|
| Packets leaving the bottleneck gateway are all the same size and have a
transmission time of b seconds. The source-sink-source "pipe"
is completely full (i.e., if the roundtrip time including queueing
delay is r, there are r/b packets in transit). A packet that
departs the gateway at time D results in a new packet arrival at time
D+r (the time to take one trip around the loop). The queue length is
decremented at packet departures and incremented at packet arrivals. There
will be a gap of Ø = r mod b
between the departure of a packet from the gateway queue and the arrival of
the next packet at the queue. We call this gap the phase of the
conversation relative to this gateway. |
|
|
|
Explanation
Suppose bulk data 1000-byte packets flowing in a 2 Mbps link having a total
roundtrip time of 60 ms (including router's queue). In this case we have:
| |
r = 60 ms
1000 x 8 x 1000
b = --------------- = 3.8147 ms
2 x 1024 x 1024
n = 60 ms / 3.8147 ms = 15 packets
Ø = 60 ms - 15 x 3.8147 ms = 2.7795 ms |
|
where n is the number of packets flowing in the link and
Ø is the phase of the conversation.
L.B. |
|
Simulations of phase effects |
|
Simulations showing the discriminatory behavior of phase effect in a
network with Drop Tail gateways and TCP congestion control
will be based in the network in Figure 3, with two FTP connections, a
Drop Tail gateway and a shared sink. |
|
|
|
To show the effect, the roundtrip time for node 2 packets
is changed slightly for each new simulation, while the roundtrip time
for node 1 packets is kept constant. In simulations
where the two connections have the same roundtrip time, the
two connections receive equal throughput. However, when the two
roundtrip times differ, the network preferentially drops packets from
one of the two connections and its throughput suffers. This behavior is
a function of the relative phase of the two connections and changes
with small changes to the propagation time of any link. |
|
In this simulation the gateways use FIFO queueing and
Drop Tail on queue overflow. FTP sources always have a packet
to send and always send a maximalsized packet as soon as the window
allows them to do so. A sink immediately sends an ACK packet when it
receives a data packet. Source and sink nodes implement a congestion
control algorithm similar to that in 4.3tahoe
BSD TCP. |
|
The essential characteristic of the network in Figure 3 is that two fast
lines are feeding into one slower line. Our simulations use 1000byte
FTP packets and 40-byte ACK packets. The gateway buffer has a
capacity of 15 packets. With 100ms of propagation delay in the
link 3-4, packets from node 1 have a roundtrip time of
221.44ms in the absence of queues. The gateway takes 10ms to
transmit an FTP packet on the slow line, so a window of 23 packets
is sufficient to "fill the pipe". (This means that when a connection
has a window greater than 23 packets, there must be at least one
packet in the gateway queue.) This small network is not intended to model
realistic network traffic, but is intended as a simple model exhibiting
traffic phase effects. |
|
|
|
|
Previous
|
Content
|
Next
|
