1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
|
Network Working Group A. McKenzie
Request for Comments: 1110 BBN STC
August 1989
A Problem with the TCP Big Window Option
Status of this Memo
This memo comments on the TCP Big Window option described in RFC
1106. Distribution of this memo is unlimited.
Abstract
The TCP Big Window option discussed in RFC 1106 will not work
properly in an Internet environment which has both a high bandwidth *
delay product and the possibility of disordering and duplicating
packets. In such networks, the window size must not be increased
without a similar increase in the sequence number space. Therefore,
a different approach to big windows should be taken in the Internet.
Discussion
TCP was designed to work in a packet store-and-forward environment
characterized by the possibility of packet loss, packet disordering,
and packet duplication. Packet loss can occur, for example, by a
congested network element discarding a packet. Packet disordering
can occur, for example, by packets of a TCP connection being
arbitrarily transmitted partially over a low bandwidth terrestrial
path and partially over a high bandwidth satellite path. Packet
duplication can occur, for example, when two directly-connected
network elements use a reliable link protocol and the link goes down
after the receiver correctly receives a packet but before the
transmitter receives an acknowledgement for the packet; the
transmitter and receiver now each take responsibility for attempting
to deliver the same packet to its ultimate destination.
TCP has the task of recreating at the destination an exact copy of
the data stream generated at the source, in the same order and with
no gaps or duplicates. The mechanism used to accomplish this task is
to assign a "unique" sequence number to each byte of data at its
source, and to sort the bytes at the destination according to the
sequence number. The sorting operation corrects any disordering. An
acknowledgement, timeout, and retransmission scheme corrects for data
loss. The uniqueness of the sequence number corrects for data
duplication.
As a practical matter, however, the sequence number is not unique; it
McKenzie [Page 1]
^L
RFC 1110 Comments on TCP Big Window Option August 1989
is contained in a 32-bit field and therefore "wraps around" after the
transmission of 2**32 bytes of data. Two additional mechanisms are
used to insure the effective uniqueness of sequence numbers; these
are the TCP transmission window and bounds on packet lifetime within
the Internet, including the IP Time-to-Live (TTL). The transmission
window specifies the maximum number of bytes which may be sent by the
source in one source-destination roundtrip time. Since the TCP
transmission window is specified by 16 bits, which is 1/65536 of the
sequence number space, a sequence number will not be reused (used to
number another byte) for 65,536 roundtrip times. So long as the
combination of gateway action on the IP TTL and holding times within
the individual networks which interconnect the gateways do not allow
a packet's lifetime to exceed 65,536 roundtrip times, each sequence
number is effectively unique. It was believed by the TCP designers
that the networks and gateways forming the internet would meet this
constraint, and such has been the case.
The proposed TCP Big Window option, as described in RFC 1106, expands
the size of the window specification to 30 bits, while leaving the
sequence number space unchanged. Thus, a sequence number can be
reused after 4 roundtrip times. Further, the Nak option allows a
packet to be retransmitted (i.e., potentially duplicated) by the
source after only one roundtrip time. Thus, if a packet becomes
"lost" in the Internet for only about 5 roundtrip times it may be
delivered when its sequence number again lies within the window,
albeit a later cycle of the window. In this case, TCP will not
necessarily recreate at the destination an exact copy of the data
stream generated at the source; it may replace some data with earlier
data.
Of course, the problem described above results from the storage of
the "lost" packet within the net, and its subsequent out-of-order
delivery. RFC 1106 seems to describe use of the proposed options in
an isolated satellite network. We may hypothesize that this network
is memoryless, and thus cannot deliver packets out of order; it
either delivers a packet in order or loses it. If this is the case,
then there is no problem with the proposed options. The Internet,
however, can deliver packets out of order, and this will likely
continue to be true even if gigabit links become part of the
Internet. Therefore, the approach described in RFC 1106 cannot be
adopted for general Internet use.
McKenzie [Page 2]
^L
RFC 1110 Comments on TCP Big Window Option August 1989
Author's Address
Alex McKenzie
Bolt Beranek and Newman Inc.
10 Moulton Street
Cambridge, MA 02238
Phone: (617) 873-2962
EMail: MCKENZIE@BBN.COM
McKenzie [Page 3]
^L
|