| 1 |
<pre>Network Working Group S. O'Malley |
| 2 |
Request for Comments: 1263 L. Peterson |
| 3 |
University of Arizona |
| 4 |
October 1991 |
| 5 |
|
| 6 |
|
| 7 |
<span class="h1">TCP EXTENSIONS CONSIDERED HARMFUL</span> |
| 8 |
|
| 9 |
|
| 10 |
Status of this Memo |
| 11 |
|
| 12 |
This memo provides information for the Internet community. It does |
| 13 |
not specify an Internet standard. Distribution of this document is |
| 14 |
unlimited. |
| 15 |
|
| 16 |
Abstract |
| 17 |
|
| 18 |
This RFC comments on recent proposals to extend TCP. It argues that |
| 19 |
the backward compatible extensions proposed in RFC's 1072 and 1185 |
| 20 |
should not be pursued, and proposes an alternative way to evolve the |
| 21 |
Internet protocol suite. Its purpose is to stimulate discussion in |
| 22 |
the Internet community. |
| 23 |
|
| 24 |
<span class="h2"><a class="selflink" id="section-1" href="#section-1">1</a>. Introduction</span> |
| 25 |
|
| 26 |
The rapid growth of the size, capacity, and complexity of the |
| 27 |
Internet has led to the need to change the existing protocol suite. |
| 28 |
For example, the maximum TCP window size is no longer sufficient to |
| 29 |
efficiently support the high capacity links currently being planned |
| 30 |
and constructed. One is then faced with the choice of either leaving |
| 31 |
the protocol alone and accepting the fact that TCP will run no faster |
| 32 |
on high capacity links than on low capacity links, or changing TCP. |
| 33 |
This is not an isolated incident. We have counted at least eight |
| 34 |
other proposed changes to TCP (some to be taken more seriously than |
| 35 |
others), and the question is not whether to change the protocol |
| 36 |
suite, but what is the most cost effective way to change it. |
| 37 |
|
| 38 |
This RFC compares the costs and benefits of three approaches to |
| 39 |
making these changes: the creation of new protocols, backward |
| 40 |
compatible protocol extensions, and protocol evolution. The next |
| 41 |
section introduces these three approaches and enumerates the |
| 42 |
strengths and weaknesses of each. The following section describes |
| 43 |
how we believe these three approaches are best applied to the many |
| 44 |
proposed changes to TCP. Note that we have not written this RFC as an |
| 45 |
academic exercise. It is our intent to argue against acceptance of |
| 46 |
the various TCP extensions, most notably RFC's 1072 and 1185 [<a href="#ref-4" title=""TCP Extensions for Long-Delay Paths"">4</a>,<a href="#ref-5" title=""TCP Extensions for High- Speed Paths"">5</a>], |
| 47 |
by describing a more palatable alternative. |
| 48 |
|
| 49 |
|
| 50 |
|
| 51 |
|
| 52 |
<span class="grey">O'Malley & Peterson [Page 1]</span></pre> |
| 53 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-2" ></span> |
| 54 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 55 |
|
| 56 |
|
| 57 |
<span class="h2"><a class="selflink" id="section-2" href="#section-2">2</a>. Creation vs. Extension vs. Evolution</span> |
| 58 |
|
| 59 |
<span class="h3"><a class="selflink" id="section-2.1" href="#section-2.1">2.1</a>. Protocol Creation</span> |
| 60 |
|
| 61 |
Protocol creation involves the design, implementation, |
| 62 |
standardization, and distribution of an entirely new protocol. In |
| 63 |
this context, there are two basic reasons for creating a new |
| 64 |
protocol. The first is to replace an old protocol that is so outdated |
| 65 |
that it can no longer be effectively extended to perform its original |
| 66 |
function. The second is to add a new protocol because users are |
| 67 |
making demands upon the original protocol that were not envisioned by |
| 68 |
the designer and cannot be efficiently handled in terms of the |
| 69 |
original protocol. For example, TCP was designed as a reliable |
| 70 |
byte-stream protocol but is commonly used as both a reliable record- |
| 71 |
stream protocol and a reliable request-reply protocol due to the lack |
| 72 |
of such protocols in the Internet protocol suite. The performance |
| 73 |
demands placed upon a byte-stream protocol in the new Internet |
| 74 |
environment makes it difficult to extend TCP to meet these new |
| 75 |
application demands. |
| 76 |
|
| 77 |
The advantage of creating a new protocol is the ability to start with |
| 78 |
a clean sheet of paper when attempting to solve a complex network |
| 79 |
problem. The designer, free from the constraints of an existing |
| 80 |
protocol, can take maximum advantage of modern network research in |
| 81 |
the basic algorithms needed to solve the problem. Even more |
| 82 |
importantly, the implementor is free to steal from a large number of |
| 83 |
existing academic protocols that have been developed over the years. |
| 84 |
In some cases, if truly new functionality is desired, creating a new |
| 85 |
protocol is the only viable approach. |
| 86 |
|
| 87 |
The most obvious disadvantage of this approach is the high cost of |
| 88 |
standardizing and distributing an entirely new protocol. Second, |
| 89 |
there is the issue of making the new protocol reliable. Since new |
| 90 |
protocols have not undergone years of network stress testing, they |
| 91 |
often contain bugs which require backward compatible fixes, and |
| 92 |
hence, the designer is back where he or she started. A third |
| 93 |
disadvantage of introducing new protocols is that they generally have |
| 94 |
new interfaces which require significant effort on the part of the |
| 95 |
Internet community to use. This alone is often enough to kill a new |
| 96 |
protocol. |
| 97 |
|
| 98 |
Finally, there is a subtle problem introduced by the very freedom |
| 99 |
provided by this approach. Specifically, being able to introduce a |
| 100 |
new protocol often results in protocols that go far beyond the basic |
| 101 |
needs of the situation. New protocols resemble Senate appropriations |
| 102 |
bills; they tend to accumulate many amendments that have nothing to |
| 103 |
do with the original problem. A good example of this phenomena is the |
| 104 |
attempt to standardize VMTP [<a href="#ref-1" title=""VMTP: Versatile Message Transaction Protocol"">1</a>] as the Internet RPC protocol. While |
| 105 |
|
| 106 |
|
| 107 |
|
| 108 |
<span class="grey">O'Malley & Peterson [Page 2]</span></pre> |
| 109 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-3" ></span> |
| 110 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 111 |
|
| 112 |
|
| 113 |
VMTP was a large protocol to begin with, the closer it got to |
| 114 |
standardization the more features were added until it essentially |
| 115 |
collapsed under its own weight. As we argue below, new protocols |
| 116 |
should initially be minimal, and then evolve as the situation |
| 117 |
dictates. |
| 118 |
|
| 119 |
|
| 120 |
<span class="h3"><a class="selflink" id="section-2.2" href="#section-2.2">2.2</a>. Backward Compatible Extensions</span> |
| 121 |
|
| 122 |
In a backward compatible extension, the protocol is modified in such |
| 123 |
a fashion that the new version of the protocol can transparently |
| 124 |
inter-operate with existing versions of the protocol. This generally |
| 125 |
implies no changes to the protocol's header. TCP slow start [<a href="#ref-3" title=""Congestion Avoidance and Control"">3</a>] is an |
| 126 |
example of such a change. In a slightly more relaxed version of |
| 127 |
backward compatibility, no changes are made to the fixed part of a |
| 128 |
protocol's header. Instead, either some fields are added to the |
| 129 |
variable length options field found at the end of the header, or |
| 130 |
existing header fields are overloaded (i.e., used for multiple |
| 131 |
purposes). However, we can find no real advantage to this technique |
| 132 |
over simply changing the protocol. |
| 133 |
|
| 134 |
Backward compatible extensions are widely used to modify protocols |
| 135 |
because there is no need to synchronize the distribution of the new |
| 136 |
version of the protocol. The new version is essentially allowed to |
| 137 |
diffuse through the Internet at its own pace, and at least in theory, |
| 138 |
the Internet will continue to function as before. Thus, the explicit |
| 139 |
distribution costs are limited. Backward compatible extensions also |
| 140 |
avoid the bureaucratic costs of standardizing a new protocol. TCP is |
| 141 |
still TCP and the approval cost of a modification to an existing |
| 142 |
protocol is much less than that of a new protocol. Finally, the very |
| 143 |
difficulty of making such changes tends to restrict the changes to |
| 144 |
the minimal set needed to solve the current problem. Thus, it is rare |
| 145 |
to see unneeded changes made when using this technique. |
| 146 |
|
| 147 |
Unfortunately, this approach has several drawbacks. First, the time |
| 148 |
to distribute the new version of the protocol to all hosts can be |
| 149 |
quite long (forever in fact). This leaves the network in a |
| 150 |
heterogeneous state for long periods of time. If there is the |
| 151 |
slightest incompatibly between old and new versions, chaos can |
| 152 |
result. Thus, the implicit cost of this type of distribution can be |
| 153 |
quite high. Second, designing a backward compatible change to a new |
| 154 |
protocol is extremely difficult, and the implementations "tend toward |
| 155 |
complexity and ugliness" [<a href="#ref-5" title=""TCP Extensions for High- Speed Paths"">5</a>]. The need for backward compatibility |
| 156 |
ensures that no code can every really be eliminated from the |
| 157 |
protocol, and since such vestigial code is rarely executed, it is |
| 158 |
often wrong. Finally, most protocols have limits, based upon the |
| 159 |
design decisions of it inventors, that simply cannot be side-stepped |
| 160 |
in this fashion. |
| 161 |
|
| 162 |
|
| 163 |
|
| 164 |
<span class="grey">O'Malley & Peterson [Page 3]</span></pre> |
| 165 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-4" ></span> |
| 166 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 167 |
|
| 168 |
|
| 169 |
<span class="h3"><a class="selflink" id="section-2.3" href="#section-2.3">2.3</a>. Protocol Evolution</span> |
| 170 |
|
| 171 |
Protocol evolution is an approach to protocol change that attempts to |
| 172 |
escape the limits of backward compatibility without incurring all of |
| 173 |
the costs of creating new protocols. The basic idea is for the |
| 174 |
protocol designer to take an existing protocol that requires |
| 175 |
modification and make the desired changes without maintaining |
| 176 |
backward compatibility. This drastically simplifies the job of the |
| 177 |
protocol designer. For example, the limited TCP window size could be |
| 178 |
fixed by changing the definition of the window size in the header |
| 179 |
from 16-bits to 32-bits, and re-compiling the protocol. The effect of |
| 180 |
backward compatibility would be ensured by simply keeping both the |
| 181 |
new and old version of the protocol running until most machines use |
| 182 |
the new version. Since the change is small and invisible to the user |
| 183 |
interface, it is a trivial problem to dynamically select the correct |
| 184 |
TCP version at runtime. How this is done is discussed in the next |
| 185 |
section. |
| 186 |
|
| 187 |
Protocol evolution has several advantages. First, it is by far the |
| 188 |
simplest type of modification to make to a protocol, and hence, the |
| 189 |
modifications can be made faster and are less likely to contain bugs. |
| 190 |
There is no need to worry about the effects of the change on all |
| 191 |
previous versions of the protocol. Also, most of the protocol is |
| 192 |
carried over into the new version unchanged, thus avoiding the design |
| 193 |
and debugging cost of creating an entirely new protocol. Second, |
| 194 |
there is no artificial limit to the amount of change that can be made |
| 195 |
to a protocol, and as a consequence, its useful lifetime can be |
| 196 |
extended indefinitely. In a series of evolutionary steps, it is |
| 197 |
possible to make fairly radical changes to a protocol without |
| 198 |
upsetting the Internet community greatly. Specifically, it is |
| 199 |
possible to both add new features and remove features that are no |
| 200 |
longer required for the current environment. Thus, the protocol is |
| 201 |
not condemned to grow without bound. Finally, by keeping the old |
| 202 |
version of the protocol around, backward compatibility is guaranteed. |
| 203 |
The old code will work as well as it ever did. |
| 204 |
|
| 205 |
Assuming the infrastructure described in the following subsection, |
| 206 |
the only real disadvantage of protocol evolution is the amount of |
| 207 |
memory required to run several versions of the same protocol. |
| 208 |
Fortunately, memory is not the scarcest resource in modern |
| 209 |
workstations (it may, however, be at a premium in the BSD kernel and |
| 210 |
its derivatives). Since old versions may rarely if ever be executed, |
| 211 |
the old versions can be swapped out to disk with little performance |
| 212 |
loss. Finally, since this cost is explicit, there is a huge incentive |
| 213 |
to eliminate old protocol versions from the network. |
| 214 |
|
| 215 |
|
| 216 |
|
| 217 |
|
| 218 |
|
| 219 |
|
| 220 |
<span class="grey">O'Malley & Peterson [Page 4]</span></pre> |
| 221 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-5" ></span> |
| 222 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 223 |
|
| 224 |
|
| 225 |
<span class="h3"><a class="selflink" id="section-2.4" href="#section-2.4">2.4</a>. Infrastructure Support for Protocol Evolution</span> |
| 226 |
|
| 227 |
The effective use of protocol evolution implies that each protocol is |
| 228 |
considered a vector of implementations which share the same top level |
| 229 |
interface, and perhaps not much else. TCP[0] is the current |
| 230 |
implementation of TCP and exists to provide backward compatibility |
| 231 |
with all existing machines. TCP[1] is a version of TCP that is |
| 232 |
optimized for high-speed networks. TCP[0] is always present; TCP[1] |
| 233 |
may or may not be. Treating TCP as a vector of protocols requires |
| 234 |
only three changes to the way protocols are designed and implemented. |
| 235 |
|
| 236 |
First, each version of TCP is assigned a unique id, but this id is |
| 237 |
not given as an IP protocol number. (This is because IP's protocol |
| 238 |
number field is only 8 bits long and could easily be exhausted.) The |
| 239 |
"obvious" solution to this limitation is to increase IP's protocol |
| 240 |
number field to 32 bits. In this case, however, the obvious solution |
| 241 |
is wrong, not because of the difficultly of changing IP, but simply |
| 242 |
because there is a better approach. The best way to deal with this |
| 243 |
problem is to increase the IP protocol number field to 32 bits and |
| 244 |
move it to the very end of the IP header (i.e., the first four bytes |
| 245 |
of the TCP header). A backward compatible modification would be made |
| 246 |
to IP such that for all packets with a special protocol number, say |
| 247 |
77, IP would look into the four bytes following its header for its |
| 248 |
de-multiplexing information. On systems which do not support a |
| 249 |
modified IP, an actual protocol 77 would be used to perform the de- |
| 250 |
multiplexing to the correct TCP version. |
| 251 |
|
| 252 |
Second, a version control protocol, called VTCP, is used to select |
| 253 |
the appropriate version of TCP for a particular connection. VTCP is |
| 254 |
an example of a virtual protocol as introduced in [<a href="#ref-2" title=""RPC in the x-Kernel: Evaluating New Design Techniques"">2</a>]. Application |
| 255 |
programs access the various versions of TCP through VTCP. When a TCP |
| 256 |
connection is opened to a specific machine, VTCP checks its local |
| 257 |
cache to determine the highest common version shared by the two |
| 258 |
machines. If the target machine is in the cache, it opens that |
| 259 |
version of TCP and returns the connection to the protocol above and |
| 260 |
does not effect performance. If the target machine is not found in |
| 261 |
the cache, VTCP sends a UDP packet to the other machine asking what |
| 262 |
versions of TCP that machine supports. If it receives a response, it |
| 263 |
uses that information to select a version and puts the information in |
| 264 |
the cache. If no reply is forthcoming, it assumes that the other |
| 265 |
machine does not support VTCP and attempts to open a TCP[0] |
| 266 |
connection. VTCP's cache is flushed occasionally to ensure that its |
| 267 |
information is current. |
| 268 |
|
| 269 |
Note that this is only one possible way for VTCP to decide the right |
| 270 |
version of TCP to use. Another possibility is for VTCP to learn the |
| 271 |
right version for a particular host when it resolves the host's name. |
| 272 |
That is, version information could be stored in the Domain Name |
| 273 |
|
| 274 |
|
| 275 |
|
| 276 |
<span class="grey">O'Malley & Peterson [Page 5]</span></pre> |
| 277 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-6" ></span> |
| 278 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 279 |
|
| 280 |
|
| 281 |
System. It is also possible that VTCP might take the performance |
| 282 |
characteristics of the network into consideration when selecting a |
| 283 |
version; TCP[0] may in fact turn out to be the correct choice for a |
| 284 |
low-bandwidth network. |
| 285 |
|
| 286 |
Third, because our proposal would lead to a more dynamically changing |
| 287 |
network architecture, a mechanism for distributing new versions will |
| 288 |
need to be developed. This is clearly the hardest requirement of the |
| 289 |
infrastructure, but we believe that it can be addressed in stages. |
| 290 |
More importantly, we believe this problem can be addressed after the |
| 291 |
decision has been made to go the protocol evolution route. In the |
| 292 |
short term, we are considering only a single new version of TCP--- |
| 293 |
TCP[1]. This version can be distributed in the same ad hoc way, and |
| 294 |
at exactly the same cost, as the backward compatible changes |
| 295 |
suggested in RFC's 1072 and 1185. |
| 296 |
|
| 297 |
In the medium term, we envision the IAB approving new versions of TCP |
| 298 |
every year or so. Given this scenario, a simple distribution |
| 299 |
mechanism can be designed based on software distribution mechanisms |
| 300 |
that have be developed for other environments; e.g., Unix RDIST and |
| 301 |
Mach SUP. Such a mechanism need not be available on all hosts. |
| 302 |
Instead, hosts will be divided into two sets, those that can quickly |
| 303 |
be updated with new protocols and those that cannot. High |
| 304 |
performance machines that can use high performance networks will need |
| 305 |
the most current version of TCP as soon as it is available, thus they |
| 306 |
have incentive to change. Old machines which are too slow to drive a |
| 307 |
high capacity lines can be ignored, and probably should be ignored. |
| 308 |
|
| 309 |
In the long term, we envision protocols being designed on an |
| 310 |
application by application basis, without the need for central |
| 311 |
approval. In such a world, a common protocol implementation |
| 312 |
environment---a protocol backplane---is the right way to go. Given |
| 313 |
such a backplane, protocols can be automatically installed over the |
| 314 |
network. While we claim to know how to build such an environment, |
| 315 |
such a discussion is beyond the scope of this paper. |
| 316 |
|
| 317 |
|
| 318 |
<span class="h3"><a class="selflink" id="section-2.5" href="#section-2.5">2.5</a>. Remarks</span> |
| 319 |
|
| 320 |
Each of these three methods has its advantages. When used in |
| 321 |
combination, the result is better protocols at a lower overall cost. |
| 322 |
Backward compatible changes are best reserved for changes that do not |
| 323 |
affect the protocol's header, and do not require that the instance |
| 324 |
running on the other end of the connection also be changed. Protocol |
| 325 |
evolution should be the primary way of dealing with header fields |
| 326 |
that are no longer large enough, or when one algorithm is substituted |
| 327 |
directly for another. New protocols should be written to off load |
| 328 |
unexpected user demands on existing protocols, or better yet, to |
| 329 |
|
| 330 |
|
| 331 |
|
| 332 |
<span class="grey">O'Malley & Peterson [Page 6]</span></pre> |
| 333 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-7" ></span> |
| 334 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 335 |
|
| 336 |
|
| 337 |
catch them before they start. |
| 338 |
|
| 339 |
There are also synergistic effects. First, since we know it is |
| 340 |
possible to evolve a newly created protocol once it has been put in |
| 341 |
place, the pressure to add unnecessary features should be reduced. |
| 342 |
Second, the ability to create new protocols removes the pressure to |
| 343 |
overextend a given protocol. Finally, the ability to evolve a |
| 344 |
protocol removes the pressure to maintain backward compatibility |
| 345 |
where it is really not possible. |
| 346 |
|
| 347 |
|
| 348 |
<span class="h2"><a class="selflink" id="section-3" href="#section-3">3</a>. TCP Extensions: A Case Study</span> |
| 349 |
|
| 350 |
This section examines the effects of using our proposed methodology |
| 351 |
to implement changes to TCP. We will begin by analyzing the backward |
| 352 |
compatible extensions defined in RFC's 1072 and 1185, and proposing a |
| 353 |
set of much simpler evolutionary modifications. We also analyze |
| 354 |
several more problematical extensions to TCP, such as Transactional |
| 355 |
TCP. Finally, we point our some areas of TCP which may require |
| 356 |
changes in the future. |
| 357 |
|
| 358 |
The evolutionary modification to TCP that we propose includes all of |
| 359 |
the functionality described in RFC's 1072 and 1185, but does not |
| 360 |
preserve the header format. At the risk of being misunderstood as |
| 361 |
believing backward compatibility is a good idea, we also show how our |
| 362 |
proposed changes to TCP can be folded into a backward compatible |
| 363 |
implementation of TCP. We do this as a courtesy for those readers |
| 364 |
that cannot accept the possibility of multiple versions of TCP. |
| 365 |
|
| 366 |
|
| 367 |
<span class="h3"><a class="selflink" id="section-3.1" href="#section-3.1">3.1</a>. RFC's 1072 and 1185</span> |
| 368 |
|
| 369 |
3.1.1. Round Trip Timing |
| 370 |
|
| 371 |
In <a href="./rfc1072">RFC 1072</a>, a new ECHO option is proposed that allows each TCP |
| 372 |
packet to carry a timestamp in its header. This timestamp is used to |
| 373 |
keep a more accurate estimate of the RTT (round trip time) used to |
| 374 |
decide when to re-transmit segments. In the original TCP algorithm, |
| 375 |
the sender manually times a small number of sends. The resulting |
| 376 |
algorithm was quite complex and does not produce an accurate enough |
| 377 |
RTT for high capacity networks. The inclusion of a timestamp in every |
| 378 |
header both simplifies the code needed to calculate the RTT and |
| 379 |
improves the accuracy and robustness of the algorithm. |
| 380 |
|
| 381 |
The new algorithm as proposed in <a href="./rfc1072">RFC 1072</a> does not appear to have any |
| 382 |
serious problems. However, the authors of <a href="./rfc1072">RFC 1072</a> go to great |
| 383 |
lengths in an attempt to keep this modification backward compatible |
| 384 |
with the previous version of TCP. They place an ECHO option in the |
| 385 |
|
| 386 |
|
| 387 |
|
| 388 |
<span class="grey">O'Malley & Peterson [Page 7]</span></pre> |
| 389 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-8" ></span> |
| 390 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 391 |
|
| 392 |
|
| 393 |
SYN segment and state, "It is likely that most implementations will |
| 394 |
properly ignore any options in the SYN segment that they do not |
| 395 |
understand, so new initial options should not cause problems" [<a href="#ref-4" title=""TCP Extensions for Long-Delay Paths"">4</a>]. |
| 396 |
This statement does not exactly inspire confidence, and we consider |
| 397 |
the addition of an optional field to any protocol to be a de-facto, |
| 398 |
if not a de-jure, example of an evolutionary change. Optional fields |
| 399 |
simply attempt to hide the basic incompatibility inside the protocol, |
| 400 |
it does not eliminate it. Therefore, since we are making an |
| 401 |
evolutionary change anyway, the only modification to the proposed |
| 402 |
algorithm is to move the fields into the header proper. Thus, each |
| 403 |
header will contain 32-bit echo and echo reply fields. Two fields are |
| 404 |
needed to handle bi-directional data streams. |
| 405 |
|
| 406 |
|
| 407 |
3.1.2. Window Size and Sequence Number Space |
| 408 |
|
| 409 |
Long Fat Networks (LFN's), networks which contain very high capacity |
| 410 |
lines with very high latency, introduce the possibility that the |
| 411 |
number of bits in transit (the bandwidth-delay product) could exceed |
| 412 |
the TCP window size, thus making TCP the limiting factor in network |
| 413 |
performance. Worse yet, the time it takes the sequence numbers to |
| 414 |
wrap around could be reduced to a point below the MSL (maximum |
| 415 |
segment lifetime), introducing the possibility of old packets being |
| 416 |
mistakenly accepted as new. |
| 417 |
|
| 418 |
<a href="./rfc1072">RFC 1072</a> extends the window size through the use of an implicit |
| 419 |
constant scaling factor. The window size in the TCP header is |
| 420 |
multiplied by this factor to get the true window size. This |
| 421 |
algorithm has three problems. First, one must prove that at all times |
| 422 |
the implicit scaling factor used by the sender is the same as the |
| 423 |
receiver. The proposed algorithm appears to do so, but the |
| 424 |
complexity of the algorithm creates the opportunity for poor |
| 425 |
implementations to affect the correctness of TCP. Second, the use of |
| 426 |
a scaling factor complicates the TCP implementation in general, and |
| 427 |
can have serious effects on other parts of the protocol. |
| 428 |
|
| 429 |
A final problem is what we characterize as the "quantum window |
| 430 |
sizing" problem. Assuming that the scaling factors will be powers of |
| 431 |
two, the algorithm right shifts the receiver's window before sending |
| 432 |
it. This effectively rounds the window size down to the nearest |
| 433 |
multiple of the scaling factor. For large scaling factors, say 64k, |
| 434 |
this implies that window values are all multiples of 64k and the |
| 435 |
minimum window size is 64k; advertising a smaller window is |
| 436 |
impossible. While this is not necessarily a problem (and it seems to |
| 437 |
be an extreme solution to the silly window syndrome) what effect this |
| 438 |
will have on the performance of high-speed network links is anyone's |
| 439 |
guess. We can imagine this extension leading to future papers |
| 440 |
entitled "A Quantum Mechanical Approach to Network Performance". |
| 441 |
|
| 442 |
|
| 443 |
|
| 444 |
<span class="grey">O'Malley & Peterson [Page 8]</span></pre> |
| 445 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-9" ></span> |
| 446 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 447 |
|
| 448 |
|
| 449 |
<a href="./rfc1185">RFC 1185</a> is an attempt to get around the problem of the window |
| 450 |
wrapping too quickly without explicitly increasing the sequence |
| 451 |
number space. Instead, the RFC proposes to use the timestamp used in |
| 452 |
the ECHO option to weed out old duplicate messages. The algorithm |
| 453 |
presented in <a href="./rfc1185">RFC 1185</a> is complex and has been shown to be seriously |
| 454 |
flawed at a recent End-to-End Research Group meeting. Attempts are |
| 455 |
currently underway to fix the algorithm presented in the RFC. We |
| 456 |
believe that this is a serious mistake. |
| 457 |
|
| 458 |
We see two problems with this approach on a very fundamental level. |
| 459 |
First, we believe that making TCP depend on accurate clocks for |
| 460 |
correctness to be a mistake. The Internet community has NO experience |
| 461 |
with transport protocols that depend on clocks for correctness. |
| 462 |
Second, the proposal uses two distinct schemes to deal with old |
| 463 |
duplicate packets: the sliding window algorithm takes care of "new" |
| 464 |
old packets (packets from the current sequence number epoch) and the |
| 465 |
timestamp algorithm deals with "old" old packets (packets from |
| 466 |
previous sequence number epochs). It is hard enough getting one of |
| 467 |
these schemes to work much less to get two to work and ensure that |
| 468 |
they do not interfere with one another. |
| 469 |
|
| 470 |
In <a href="./rfc1185">RFC 1185</a>, the statement is made that "An obvious fix for the |
| 471 |
problem of cycling the sequence number space is to increase the size |
| 472 |
of the TCP sequence number field." Using protocol evolution, the |
| 473 |
obvious fix is also the correct one. The window size can be increased |
| 474 |
to 32 bits by simply changing a short to a long in the definition of |
| 475 |
the TCP header. At the same time, the sequence number and |
| 476 |
acknowledgment fields can be increased to 64 bits. This change is |
| 477 |
the minimum complexity modification to get the job done and requires |
| 478 |
little or no analysis to be shown to work correctly. |
| 479 |
|
| 480 |
On machines that do not support 64-bit integers, increasing the |
| 481 |
sequence number size is not as trivial as increasing the window size. |
| 482 |
However, it is identical in cost to the modification proposed in <a href="./rfc1185">RFC</a> |
| 483 |
<a href="./rfc1185">1185</a>; the high order bits can be thought of as an optimal clock that |
| 484 |
ticks only when it has to. Also, because we are not dealing with |
| 485 |
real time, the problems with unreliable system clocks is avoided. On |
| 486 |
machines that support 64-bit integers, the original TCP code may be |
| 487 |
reused. Since only very high performance machines can hope to drive |
| 488 |
a communications network at the rates this modification is designed |
| 489 |
to support, and the new generation of RISC microprocessors (e.g., |
| 490 |
MIPS R4000 and PA-RISC) do support 64-bit integers, the assumption of |
| 491 |
64-bit arithmetic may be more of an advantage than a liability. |
| 492 |
|
| 493 |
|
| 494 |
|
| 495 |
|
| 496 |
|
| 497 |
|
| 498 |
|
| 499 |
|
| 500 |
<span class="grey">O'Malley & Peterson [Page 9]</span></pre> |
| 501 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-10" ></span> |
| 502 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 503 |
|
| 504 |
|
| 505 |
3.1.3. Selective Retransmission |
| 506 |
|
| 507 |
Another problem with TCP's support for LFN's is that the sliding |
| 508 |
window algorithm used by TCP does not support any form of selective |
| 509 |
acknowledgment. Thus, if a segment is lost, the total amount of data |
| 510 |
that must be re-transmitted is some constant times the bandwidth- |
| 511 |
delay product, despite the fact that most of the segments have in |
| 512 |
fact arrived at the receiver. <a href="./rfc1072">RFC 1072</a> proposes to extend TCP to |
| 513 |
allow the receiver to return partial acknowledgments to the sender in |
| 514 |
the hope that the sender will use that information to avoid |
| 515 |
unnecessary re-transmissions. |
| 516 |
|
| 517 |
It has been our experience on predictable local area networks that |
| 518 |
the performance of partial re-transmission strategies is highly non- |
| 519 |
obvious, and it generally requires more than one iteration to find a |
| 520 |
decent algorithm. It is therefore not surprising that the algorithm |
| 521 |
proposed in <a href="./rfc1072">RFC 1072</a> has some problems. The proposed TCP extension |
| 522 |
allows the receiver to include a short list of received fragments |
| 523 |
with every ACK. The idea being that when the receiver sends back a |
| 524 |
normal ACK, it checks its queue of segments that have been received |
| 525 |
out of order and sends the relative sequence numbers of contiguous |
| 526 |
blocks of segments back to the sender. The sender then uses this |
| 527 |
information to re-transmit the segments transmitted but not listed in |
| 528 |
the ACK. |
| 529 |
|
| 530 |
As specified, this algorithm has two related problems: (1) it ignores |
| 531 |
the relative frequencies of delivered and dropped packets, and (2) |
| 532 |
the list provided in the option field is probably too short to do |
| 533 |
much good on networks with large bandwidth-delay products. In every |
| 534 |
model of high bandwidth networks that we have seen, the packet loss |
| 535 |
rate is very low, and thus, the ratio of dropped packets to delivered |
| 536 |
packets is very low. An algorithm that returns ACKs as proposed is |
| 537 |
simply going to have to send more information than one in which the |
| 538 |
receiver returns NAKs. |
| 539 |
|
| 540 |
This problem is compounded by the short size of the TCP option field |
| 541 |
(44 bytes). In theory, since we are only worried about high bandwidth |
| 542 |
networks, returning ACKs instead of NAKs is not really a problem; the |
| 543 |
bandwidth is available to send any information that's needed. The |
| 544 |
problem comes when trying to compress the ACK information into the 44 |
| 545 |
bytes allowed. The proposed extensions effectively compresses the |
| 546 |
ACK information by allowing the receiver to ACK byte ranges rather |
| 547 |
than segments, and scaling the relative sequence numbers of the re- |
| 548 |
transmitted segments. This makes it much more difficult for the |
| 549 |
sender to tell which segments should be re-transmitted, and |
| 550 |
complicates the re-transmission code. More importantly, one should |
| 551 |
never compress small amounts of data being sent over a high bandwidth |
| 552 |
network; it trades a scarce resource for an abundant resource. On |
| 553 |
|
| 554 |
|
| 555 |
|
| 556 |
<span class="grey">O'Malley & Peterson [Page 10]</span></pre> |
| 557 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-11" ></span> |
| 558 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 559 |
|
| 560 |
|
| 561 |
low bandwidth networks, selective retransmission is not needed and |
| 562 |
the SACK option should be disabled. |
| 563 |
|
| 564 |
We propose two solutions to this problem. First, the receiver can |
| 565 |
examine its list of out-of-order packets and guess which segments |
| 566 |
have been dropped, and NAK those segments back to the sender. The |
| 567 |
number of NAKs should be low enough that one per TCP packet should be |
| 568 |
sufficient. Note that the receiver has just as much information as |
| 569 |
the sender about what packets should be retransmitted, and in any |
| 570 |
case, the NAKs are simply suggestions which have no effect on |
| 571 |
correctness. |
| 572 |
|
| 573 |
Our second proposed modification is to increase the offset field in |
| 574 |
the TCP header from 4 bits to 16 bits. This allows 64k-bytes of TCP |
| 575 |
header, which allows us to radically simplify the selective re- |
| 576 |
transmission algorithm proposed in <a href="./rfc1072">RFC 1072</a>. The receiver can now |
| 577 |
simply send a list of 64-bit sequence numbers for the out-of-order |
| 578 |
segments to the sender. The sender can then use this information to |
| 579 |
do a partial retransmission without needing an ouji board to |
| 580 |
translate ACKs into segments. With the new header size, it may be |
| 581 |
faster for the receiver to send a large list than to attempt to |
| 582 |
aggregate segments into larger blocks. |
| 583 |
|
| 584 |
|
| 585 |
3.1.4. Header Modifications |
| 586 |
|
| 587 |
The modifications proposed above drastically change the size and |
| 588 |
structure of the TCP header. This makes it a good time to re-think |
| 589 |
the structure of the proposed TCP header. The primary goal of the |
| 590 |
current TCP header is to save bits in the output stream. When TCP was |
| 591 |
developed, a high bandwidth network was 56kbps, and the key use for |
| 592 |
TCP was terminal I/O. In both situations, minimal header size was |
| 593 |
important. Unfortunately, while the network has drastically |
| 594 |
increased in performance and the usage pattern of the network is now |
| 595 |
vastly different, most protocol designers still consider saving a few |
| 596 |
bits in the header to be worth almost any price. Our basic goal is |
| 597 |
different: to improve performance by eliminating the need to extract |
| 598 |
information packed into odd length bit fields in the header. Below |
| 599 |
is our first cut at such a modification. |
| 600 |
|
| 601 |
The protocol id field is there to make further evolutionary |
| 602 |
modifications to TCP easier. This field basically subsumes the |
| 603 |
protocol number field contained in the IP header with a version |
| 604 |
number. Each distinct TCP version has a different protocol id and |
| 605 |
this field ensures that the right code is looking at the right |
| 606 |
header. The offset field has been increased to 16 bits to support |
| 607 |
the larger header size required, and to simplify header processing. |
| 608 |
The code field has been extended to 16 bits to support more options. |
| 609 |
|
| 610 |
|
| 611 |
|
| 612 |
<span class="grey">O'Malley & Peterson [Page 11]</span></pre> |
| 613 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-12" ></span> |
| 614 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 615 |
|
| 616 |
|
| 617 |
The source port and destination port are unchanged. The size of both |
| 618 |
the sequence number and ACK fields have been increased to 64 bits. |
| 619 |
The open window field has been increased to 32 bits. The checksum and |
| 620 |
urgent data pointer fields are unchanged. The echo and echo reply |
| 621 |
fields are added. The option field remains but can be much larger |
| 622 |
than in the old TCP. All headers are padded out to 32 bit |
| 623 |
boundaries. Note that these changes increase the minimum header size |
| 624 |
from 24 bytes (actually 36 bytes if the ECHO and ECHO reply options |
| 625 |
defined in <a href="./rfc1072">RFC 1072</a> are included on every packet) to 48 bytes. The |
| 626 |
maximum header size has been increased to the maximum segment size. |
| 627 |
We do not believe that the the increased header size will have a |
| 628 |
measurable effect on protocol performance. |
| 629 |
|
| 630 |
0 1 2 3 |
| 631 |
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 |
| 632 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 633 |
| Protocol ID | |
| 634 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 635 |
| Offset | Code | |
| 636 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 637 |
| Source | Dest | |
| 638 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 639 |
| Seq | |
| 640 |
| | |
| 641 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 642 |
| Ack | |
| 643 |
| | |
| 644 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 645 |
| Window | |
| 646 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 647 |
| Checksum | Urgent | |
| 648 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 649 |
| Echo | |
| 650 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 651 |
| Echo Reply | |
| 652 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 653 |
| Options | Pad | |
| 654 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
| 655 |
|
| 656 |
|
| 657 |
3.1.5. Backward Compatibility |
| 658 |
|
| 659 |
The most likely objection to the proposed TCP extension is that it is |
| 660 |
not backward compatible with the current version of TCP, and most |
| 661 |
importantly, TCP's header. In this section we will present three |
| 662 |
versions of the proposed extension with increasing degrees of |
| 663 |
backward compatibility. The final version will combine the same |
| 664 |
degree of backward compatibility found in the protocol described in |
| 665 |
|
| 666 |
|
| 667 |
|
| 668 |
<span class="grey">O'Malley & Peterson [Page 12]</span></pre> |
| 669 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-13" ></span> |
| 670 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 671 |
|
| 672 |
|
| 673 |
RFC's 1072/1185, with the much simpler semantics described in this |
| 674 |
RFC. |
| 675 |
|
| 676 |
We believe that the best way to preserve backward compatibility is to |
| 677 |
leave all of TCP alone and support the transparent use of a new |
| 678 |
protocol when and where it is needed. The basic scheme is the one |
| 679 |
described in <a href="#section-2.4">section 2.4</a>. Those machines and operating systems that |
| 680 |
need to support high speed connections should implement some general |
| 681 |
protocol infrastructure that allows them to rapidly evolve protocols. |
| 682 |
Machines that do not require such service simply keep using the |
| 683 |
existing version of TCP. A virtual protocol is used to manage the use |
| 684 |
of multiple TCP versions. |
| 685 |
|
| 686 |
This approach has several advantages. First, it guarantees backward |
| 687 |
compatibility with ALL existing TCP versions because such |
| 688 |
implementations will never see strange packets with new options. |
| 689 |
Second, it supports further modification of TCP with little |
| 690 |
additional costs. Finally, since our version of TCP will more closely |
| 691 |
resemble the existing TCP protocol than that proposed in RFC's |
| 692 |
1072/1185, the cost of maintaining two simple protocols will probably |
| 693 |
be lower than maintaining one complex protocol. (Note that with high |
| 694 |
probability you still have to maintain two versions of TCP in any |
| 695 |
case.) The only additional cost is the memory required for keeping |
| 696 |
around two copies of TCP. |
| 697 |
|
| 698 |
For those that insist that the only efficient way to implement TCP |
| 699 |
modifications is in a single monolithic protocol, or those that |
| 700 |
believe that the space requirements of two protocols would be too |
| 701 |
great, we simply migrate the virtual protocol into TCP. TCP is |
| 702 |
modified so that when opening a connection, the sender uses the TCP |
| 703 |
VERSION option attached to the SYN packet to request using the new |
| 704 |
version. The receiver responds with a TCP VERSION ACK in the SYN ACK |
| 705 |
packet, after which point, the new header format described in <a href="#section-3.1.4">Section</a> |
| 706 |
<a href="#section-3.1.4">3.1.4</a> is used. Thus, there is only one version of TCP, but that |
| 707 |
version supports multiple header formats. The complexity of such a |
| 708 |
protocol would be no worse than the protocol described in <a href="./rfc1072">RFC</a> |
| 709 |
<a href="./rfc1072">1072</a>/1185. It does, however, make it more difficult to make |
| 710 |
additional changes to TCP. |
| 711 |
|
| 712 |
Finally, for those that believe that the preservation of the TCP's |
| 713 |
header format has any intrinsic value (e.g., for those that don't |
| 714 |
want to re-program their ethernet monitors), a header compatible |
| 715 |
version of our proposal is possible. One simply takes all of the |
| 716 |
additional information contained in the header given in <a href="#section-3.1.4">Section 3.1.4</a> |
| 717 |
and places it into a single optional field. Thus, one could define a |
| 718 |
new TCP option which consists of the top 32 bits of the sequence and |
| 719 |
ack fields, the echo and echo_reply fields, and the top 16 bits of |
| 720 |
the window field. This modification makes it more difficult to take |
| 721 |
|
| 722 |
|
| 723 |
|
| 724 |
<span class="grey">O'Malley & Peterson [Page 13]</span></pre> |
| 725 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-14" ></span> |
| 726 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 727 |
|
| 728 |
|
| 729 |
advantage of machines with 64-bit address spaces, but at a minimum |
| 730 |
will be just as easy to process as the protocol described in <a href="./rfc1072">RFC</a> |
| 731 |
<a href="./rfc1072">1072</a>/1185. The only restriction is that the size of the header |
| 732 |
option field is still limited to 44 bytes, and thus, selective |
| 733 |
retransmission using NAKs rather than ACKs will probably be required. |
| 734 |
|
| 735 |
The key observation is that one should make a protocol extension |
| 736 |
correct and simple before trying to make it backward compatible. As |
| 737 |
far as we can tell, the only advantages possessed by the protocol |
| 738 |
described in <a href="./rfc1072">RFC 1072</a>/1185 is that its typical header, size including |
| 739 |
options, is 8 to 10 bytes shorter. The price for this "advantage" is |
| 740 |
a protocol of such complexity that it may prove impossible for normal |
| 741 |
humans to implement. Trying to maintain backward compatibility at |
| 742 |
every stage of the protocol design process is a serious mistake. |
| 743 |
|
| 744 |
|
| 745 |
<span class="h3"><a class="selflink" id="section-3.2" href="#section-3.2">3.2</a>. TCP Over Extension</span> |
| 746 |
|
| 747 |
Another potential problem with TCP that has been discussed recently, |
| 748 |
but has not yet resulted in the generation of an RFC, is the |
| 749 |
potential for TCP to grab and hold all 2**16 port numbers on a given |
| 750 |
machine. This problem is caused by short port numbers, long MSLs, |
| 751 |
and the misuse of TCP as a request-reply protocol. TCP must hold onto |
| 752 |
each port after a close until all possible messages to that port have |
| 753 |
died, about 240 seconds. Even worse, this time is not decreasing with |
| 754 |
increase network performance. With new fast hardware, it is possible |
| 755 |
for an application to open a TCP connection, send data, get a reply, |
| 756 |
and close the connection at a rate fast enough to use up all the |
| 757 |
ports in less than 240 seconds. This usage pattern is generated by |
| 758 |
people using TCP for something it was never intended to do--- |
| 759 |
guaranteeing at-most-once semantics for remote procedure calls. |
| 760 |
|
| 761 |
The proposed solution is to embed an RPC protocol into TCP while |
| 762 |
preserving backward compatibility. This is done by piggybacking the |
| 763 |
request message on the SYN packet and the reply message on the SYN- |
| 764 |
ACK packet. This approach suffers from one key problem: it reduces |
| 765 |
the probability of a correct TCP implementation to near 0. The basic |
| 766 |
problem has nothing to do with TCP, rather it is the lack of an |
| 767 |
Internet request-reply protocol that guarantees at-most-once |
| 768 |
semantics. |
| 769 |
|
| 770 |
We propose to solve this problem by the creation of a new protocol. |
| 771 |
This has already been attempted with VMTP, but the size and |
| 772 |
complexity of VMTP, coupled with the process currently required to |
| 773 |
standardize a new protocol doomed it from the start. Instead of |
| 774 |
solving the general problem, we propose to use Sprite RPC [<a href="#ref-7" title=""The Sprite Remote Procedure Call System"">7</a>], a much |
| 775 |
simpler protocol, as a means of off-loading inappropriate users from |
| 776 |
TCP. |
| 777 |
|
| 778 |
|
| 779 |
|
| 780 |
<span class="grey">O'Malley & Peterson [Page 14]</span></pre> |
| 781 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-15" ></span> |
| 782 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 783 |
|
| 784 |
|
| 785 |
The basic design would attempt to preserve as much of the TCP |
| 786 |
interface as possible in order that current TCP (mis)users could be |
| 787 |
switched to Sprite RPC without requiring code modification on their |
| 788 |
part. A virtual protocol could be used to select the correct protocol |
| 789 |
TCP or Sprite RPC if it exists on the other machine. A backward |
| 790 |
compatible modification to TCP could be made which would simply |
| 791 |
prevent it from grabbing all of the ports by refusing connections. |
| 792 |
This would encourage TCP abusers to use the new protocol. |
| 793 |
|
| 794 |
Sprite RPC, which is designed for a local area network, has two |
| 795 |
problems when extended into the Internet. First, it does not have a |
| 796 |
usefully flow control algorithm. Second, it lacks the necessary |
| 797 |
semantics to reliably tear down connections. The lack of a tear down |
| 798 |
mechanism needs to be solved, but the flow control problem could be |
| 799 |
dealt with in later iterations of the protocol as Internet blast |
| 800 |
protocols are not yet well understood; for now, we could simple limit |
| 801 |
the size of each message to 16k or 32k bytes. This might also be a |
| 802 |
good place to use a decomposed version of Sprite RPC [<a href="#ref-2" title=""RPC in the x-Kernel: Evaluating New Design Techniques"">2</a>], which |
| 803 |
exposes each of these features as separate protocols. This would |
| 804 |
permit the quick change of algorithms, and once the protocol had |
| 805 |
stabilized, a monolithic version could be constructed and distributed |
| 806 |
to replace the decomposed version. |
| 807 |
|
| 808 |
In other words, the basic strategy is to introduce as simple of RPC |
| 809 |
protocol as possible today, and later evolve this protocol to address |
| 810 |
the known limitations. |
| 811 |
|
| 812 |
|
| 813 |
<span class="h3"><a class="selflink" id="section-3.3" href="#section-3.3">3.3</a>. Future Modifications</span> |
| 814 |
|
| 815 |
The header prediction algorithm should be generalized so as to be |
| 816 |
less sensitive to changes in the protocols header and algorithm. |
| 817 |
There almost seems to be as much effort to make all modifications to |
| 818 |
TCP backward compatible with header prediction as there is to make |
| 819 |
them backward compatible with TCP. The question that needs to be |
| 820 |
answered is: are there any changes we can made to TCP to make header |
| 821 |
prediction easier, including the addition of information into the |
| 822 |
header. In [<a href="#ref-6" title=""A Tran- sparent Blast Facility"">6</a>], the authors showed how one might generalize |
| 823 |
optimistic blast from VMTP to almost any protocol that performs |
| 824 |
fragmentation and reassembly. Generalizing header prediction so that |
| 825 |
it scales with TCP modification would be step in the right direction. |
| 826 |
|
| 827 |
It is clear that an evolutionary change to increase the size of the |
| 828 |
source and destination ports in the TCP header will eventually be |
| 829 |
necessary. We also believe that TCP could be made significantly |
| 830 |
simpler and more flexible through the elimination of the pseudo- |
| 831 |
header. The solution to this problem is to simply add a length field |
| 832 |
and the IP address of the destination to the TCP header. It has also |
| 833 |
|
| 834 |
|
| 835 |
|
| 836 |
<span class="grey">O'Malley & Peterson [Page 15]</span></pre> |
| 837 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-16" ></span> |
| 838 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 839 |
|
| 840 |
|
| 841 |
been mentioned that better and simpler TCP connection establishment |
| 842 |
algorithms would be useful. Some form of reliable record stream |
| 843 |
protocol should be developed. Performing sliding window and flow |
| 844 |
control over records rather than bytes would provide numerous |
| 845 |
opportunities for optimizations and allow TCP to return to its |
| 846 |
original purpose as a byte-stream protocol. Finally, it has become |
| 847 |
clear to us that the current Internet congestion control strategy is |
| 848 |
to use TCP for everything since it is the only protocol that supports |
| 849 |
congestion control. One of the primary reasons many "new protocols" |
| 850 |
are proposed as TCP options is that it is the only way to get at |
| 851 |
TCP's congestion control. At some point, a TCP-independent congestion |
| 852 |
control scheme must be implemented and one might then be able to |
| 853 |
remove the existing congestion control from TCP and radically |
| 854 |
simplify the protocol. |
| 855 |
|
| 856 |
|
| 857 |
<span class="h2"><a class="selflink" id="section-4" href="#section-4">4</a>. Discussion</span> |
| 858 |
|
| 859 |
One obvious side effect of the changes we propose is to increase the |
| 860 |
size of the TCP header. In some sense, this is inevitable; just about |
| 861 |
every field in the header has been pushed to its limit by the radical |
| 862 |
growth of the network. However, we have made very little effort to |
| 863 |
make the minimal changes to solve the current problem. In fact, we |
| 864 |
have tended to sacrifice header size in order to defer future changes |
| 865 |
as long as possible. The problem with this is that one of TCP's |
| 866 |
claims to fame is its efficiency at sending small one byte packets |
| 867 |
over slow networks. Increasing the size of the TCP header will |
| 868 |
inevitably result in some increase in overhead on small packets on |
| 869 |
slow networks. Clark among others have stated that they see no |
| 870 |
fundamental performance limitations that would prevent TCP from |
| 871 |
supporting very high speed networks. This is true as far as it goes; |
| 872 |
there seems to be a direct trade-off between TCP performance on high |
| 873 |
speed networks and TCP performance on slow speed networks. The |
| 874 |
dynamic range is simply too great to be optimally supported by one |
| 875 |
protocol. Hence, in keeping around the old version of TCP we have |
| 876 |
effectively split TCP into two protocols, one for high bandwidth |
| 877 |
lines and the other for low bandwidth lines. |
| 878 |
|
| 879 |
Another potential argument is that all of the changes mentioned above |
| 880 |
should be packaged together as a new version of TCP. This version |
| 881 |
could be standardized and we could all go back to the status quo of |
| 882 |
stable unchanging protocols. While to a certain extent this is |
| 883 |
inevitable---there is a backlog of necessary TCP changes because of |
| 884 |
the current logistical problems in modifying protocols---it is only |
| 885 |
begs the question. The status quo is simply unacceptably static; |
| 886 |
there will always be future changes to TCP. Evolutionary change will |
| 887 |
also result in a better and more reliable TCP. Making small changes |
| 888 |
and distributing them at regular intervals ensures that one change |
| 889 |
|
| 890 |
|
| 891 |
|
| 892 |
<span class="grey">O'Malley & Peterson [Page 16]</span></pre> |
| 893 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-17" ></span> |
| 894 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 895 |
|
| 896 |
|
| 897 |
has actually been stabilized before the next has been made. It also |
| 898 |
presents a more balanced workload to the protocol designer; rather |
| 899 |
than designing one new protocol every 10 years he makes annual |
| 900 |
protocol extensions. It will also eventually make protocol |
| 901 |
distribution easier: the basic problem with protocol distribution now |
| 902 |
is that it is done so rarely that no one knows how to do it and there |
| 903 |
is no incentive to develop the infrastructure needed to perform the |
| 904 |
task efficiently. While the first protocol distribution is almost |
| 905 |
guaranteed to be a disaster, the problem will get easier with each |
| 906 |
additional one. Finally, such a new TCP would have the same problems |
| 907 |
as VMTP did; a radically new protocol presents a bigger target. |
| 908 |
|
| 909 |
The violation of backward compatibility in systems as complex as the |
| 910 |
Internet is always a serious step. However, backward compatibility is |
| 911 |
a technique, not a religion. Two facts are often overlooked when |
| 912 |
backward compatibility gets out of hand. First, violating backward |
| 913 |
compatibility is always a big win when you can get away with it. One |
| 914 |
of the key advantages of RISC chips over CISC chips is simply that |
| 915 |
they were not backward compatible with anything. Thus, they were not |
| 916 |
bound by design decisions made when compilers were stupid and real |
| 917 |
men programmed in assembler. Second, one is going to have to break |
| 918 |
backward compatibility at some point anyway. Every system has some |
| 919 |
headroom limitations which result in either stagnation (IBM mainframe |
| 920 |
software) or even worse, accidental violations of backward |
| 921 |
compatibility. |
| 922 |
|
| 923 |
Of course, the biggest problem with our approach is that it is not |
| 924 |
compatible with the existing standardization process. We hope to be |
| 925 |
able to design and distribute protocols in less time than it takes a |
| 926 |
standards committee to agree on an acceptable meeting time. This is |
| 927 |
inevitable because the basic problem with networking is the |
| 928 |
standardization process. Over the last several years, there has been |
| 929 |
a push in the research community for lightweight protocols, when in |
| 930 |
fact what is needed are lightweight standards. Also note that we |
| 931 |
have not proposed to implement some entirely new set of "superior" |
| 932 |
communications protocols, we have simply proposed a system for making |
| 933 |
necessary changes to the existing protocol suites fast enough to keep |
| 934 |
up with the underlying change in the network. In fact, the first |
| 935 |
standards organization that realizes that the primary impediment to |
| 936 |
standardization is poor logistical support will probably win. |
| 937 |
|
| 938 |
|
| 939 |
<span class="h2"><a class="selflink" id="section-5" href="#section-5">5</a>. Conclusions</span> |
| 940 |
|
| 941 |
The most important conclusion of this RFC is that protocol change |
| 942 |
happens and is currently happening at a very respectable clip. While |
| 943 |
all of the changes given as example in this document are from TCP, |
| 944 |
there are many other protocols that require modification. In a more |
| 945 |
|
| 946 |
|
| 947 |
|
| 948 |
<span class="grey">O'Malley & Peterson [Page 17]</span></pre> |
| 949 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-18" ></span> |
| 950 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 951 |
|
| 952 |
|
| 953 |
prosaic domain, the telephone company is running out of phone |
| 954 |
numbers; they are being overrun by fax machines, modems, and cars. |
| 955 |
The underlying cause of these problems seems to be an consistent |
| 956 |
exponential increase almost all network metrics: number of hosts, |
| 957 |
bandwidth, host performance, applications, and so on, combined with |
| 958 |
an attempt to run the network with a static set of unchanging network |
| 959 |
protocols. This has been shown to be impossible and one can almost |
| 960 |
feel the pressure for protocol change building. We simply propose to |
| 961 |
explicitly deal with the changes rather keep trying to hold back the |
| 962 |
flood. |
| 963 |
|
| 964 |
Of almost equal importance is the observation that TCP is a protocol |
| 965 |
and not a platform for implementing other protocols. Because of a |
| 966 |
lack of any alternatives, TCP has become a de-facto platform for |
| 967 |
implementing other protocols. It provides a vague standard interface |
| 968 |
with the kernel, it runs on many machines, and has a well defined |
| 969 |
distribution path. Otherwise sane people have proposed Bounded Time |
| 970 |
TCP (an unreliable byte stream protocol), Simplex TCP (which supports |
| 971 |
data in only one direction) and Multi-cast TCP (too horrible to even |
| 972 |
consider). All of these protocols probably have their uses, but not |
| 973 |
as TCP options. The fact that a large number of people are willing to |
| 974 |
use TCP as a protocol implementation platform points to the desperate |
| 975 |
need for a protocol independent platform. |
| 976 |
|
| 977 |
Finally, we point out that in our research we have found very little |
| 978 |
difference in the actual technical work involved with the three |
| 979 |
proposed methods of protocol modification. The amount of work |
| 980 |
involved in a backward compatible change is often more than that |
| 981 |
required for an evolutionary change or the creation of a new |
| 982 |
protocol. Even the distribution costs seem to be identical. The |
| 983 |
primary cost difference between the three approaches is the cost of |
| 984 |
getting the modification approved. A protocol modification, no matter |
| 985 |
how extensive or bizarre, seems to incur much less cost and risk. It |
| 986 |
is time to stop changing the protocols to fit our current way of |
| 987 |
thinking, and start changing our way of thinking to fit the |
| 988 |
protocols. |
| 989 |
|
| 990 |
|
| 991 |
<span class="h2"><a class="selflink" id="section-6" href="#section-6">6</a>. References</span> |
| 992 |
|
| 993 |
|
| 994 |
[<a id="ref-1">1</a>] Cheriton D., "VMTP: Versatile Message Transaction Protocol", <a href="./rfc1045">RFC</a> |
| 995 |
<a href="./rfc1045">1045</a>, Stanford University, February 1988. |
| 996 |
|
| 997 |
|
| 998 |
[<a id="ref-2">2</a>] Hutchinson, N., Peterson, L., Abbott, M., and S. O'Malley, "RPC in |
| 999 |
the x-Kernel: Evaluating New Design Techniques", Proceedings of the |
| 1000 |
12th Symposium on Operating System Principles, Pgs. 91-101, |
| 1001 |
|
| 1002 |
|
| 1003 |
|
| 1004 |
<span class="grey">O'Malley & Peterson [Page 18]</span></pre> |
| 1005 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-19" ></span> |
| 1006 |
<span class="grey"><a href="./rfc1263">RFC 1263</a> TCP Extensions Considered Harmful October 1991</span> |
| 1007 |
|
| 1008 |
|
| 1009 |
December 1989. |
| 1010 |
|
| 1011 |
|
| 1012 |
[<a id="ref-3">3</a>] Jacobson, V., "Congestion Avoidance and Control", SIGCOMM '88, |
| 1013 |
August 1988. |
| 1014 |
|
| 1015 |
|
| 1016 |
[<a id="ref-4">4</a>] Jacobson, V., and R. Braden, "TCP Extensions for Long-Delay Paths", |
| 1017 |
<a href="./rfc1072">RFC 1072</a>, LBL, ISI, October 1988. |
| 1018 |
|
| 1019 |
|
| 1020 |
[<a id="ref-5">5</a>] Jacobson, V., Braden, R., and L. Zhang, "TCP Extensions for High- |
| 1021 |
Speed Paths", <a href="./rfc1185">RFC 1185</a>, LBL, ISI, PARC, October 1990. |
| 1022 |
|
| 1023 |
|
| 1024 |
[<a id="ref-6">6</a>] O'Malley, S., Abbott, M., Hutchinson, N., and L. Peterson, "A Tran- |
| 1025 |
sparent Blast Facility", Journal of Internetworking, Vol. 1, No. |
| 1026 |
2, Pgs. 57-75, December 1990. |
| 1027 |
|
| 1028 |
|
| 1029 |
[<a id="ref-7">7</a>] Welch, B., "The Sprite Remote Procedure Call System", UCB/CSD |
| 1030 |
86/302, University of California at Berkeley, June 1988. |
| 1031 |
|
| 1032 |
<span class="h2"><a class="selflink" id="section-7" href="#section-7">7</a>. Security Considerations</span> |
| 1033 |
|
| 1034 |
Security issues are not discussed in this memo. |
| 1035 |
|
| 1036 |
|
| 1037 |
<span class="h2"><a class="selflink" id="section-8" href="#section-8">8</a>. Authors' Addresses</span> |
| 1038 |
|
| 1039 |
Larry L. Peterson |
| 1040 |
University of Arizona |
| 1041 |
Department of Computer Sciences |
| 1042 |
Tucson, AZ 85721 |
| 1043 |
|
| 1044 |
Phone: (602) 621-4231 |
| 1045 |
EMail: [email protected] |
| 1046 |
|
| 1047 |
|
| 1048 |
Sean O'Malley |
| 1049 |
University of Arizona |
| 1050 |
Department of Computer Sciences |
| 1051 |
Tucson, AZ 85721 |
| 1052 |
|
| 1053 |
Phone: 602-621-8373 |
| 1054 |
EMail: [email protected] |
| 1055 |
|
| 1056 |
|
| 1057 |
|
| 1058 |
|
| 1059 |
|
| 1060 |
O'Malley & Peterson [Page 19] |
| 1061 |
</pre> |