| 1 |
<pre>Network Working Group W. Prue |
| 2 |
Request for Comments: 1046 J. Postel |
| 3 |
ISI |
| 4 |
February 1988 |
| 5 |
|
| 6 |
|
| 7 |
<span class="h1">A Queuing Algorithm to Provide Type-of-Service for IP Links</span> |
| 8 |
|
| 9 |
Status of this Memo |
| 10 |
|
| 11 |
This memo is intended to explore how Type-of-Service might be |
| 12 |
implemented in the Internet. The proposal describes a method of |
| 13 |
queuing which can provide the different classes of service. The |
| 14 |
technique also prohibits one class of service from consuming |
| 15 |
excessive resources or excluding other classes of service. This is |
| 16 |
an "idea paper" and discussion is strongly encouraged. Distribution |
| 17 |
of this memo is unlimited. |
| 18 |
|
| 19 |
Introduction |
| 20 |
|
| 21 |
The Type-of-Service (TOS) field in IP headers allows one to chose |
| 22 |
from none to all the following service types; low delay, high |
| 23 |
throughput, and high reliability. It also has a portion allowing a |
| 24 |
priority selection from 0-7. To date, there is nothing describing |
| 25 |
what should be done with these parameters. This discussion proposes |
| 26 |
an approach to providing the different classes of service and |
| 27 |
priorities requestable in the TOS field. |
| 28 |
|
| 29 |
Desired Attributes |
| 30 |
|
| 31 |
We should first consider how we want these services to perform. We |
| 32 |
must first assume that there is a demand for service that exceeds |
| 33 |
current capabilities. If not, significant queues do not form and |
| 34 |
queuing algorithms become superfluous. |
| 35 |
|
| 36 |
The low delay class of service should have the ability to pass data |
| 37 |
through the net faster than regular data. If a request is for low |
| 38 |
delay class of service only, not high throughput or high reliability, |
| 39 |
the Internet should provide low delay for relatively less throughput, |
| 40 |
with less than high reliability. The requester is more concerned |
| 41 |
with promptness of delivery than guaranteed delivery. The Internet |
| 42 |
should provide a Maximum Guaranteed Delay (MGD) per node, or better, |
| 43 |
if the datagram successfully traverses the Internet. In the worst |
| 44 |
case, a datagram's arrival will be MGD times the number of nodes |
| 45 |
traversed. A node is any packet switching element, including IP |
| 46 |
gateways and ARPANET IMP's. The MGD bound will not be affected by |
| 47 |
the amount of traffic in the net. During non-busy hours, the delay |
| 48 |
provided should be better than the guarantee. If the delay a |
| 49 |
|
| 50 |
|
| 51 |
|
| 52 |
<span class="grey">Prue & Postel [Page 1]</span></pre> |
| 53 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-2" ></span> |
| 54 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 55 |
|
| 56 |
|
| 57 |
satellite link introduces is less than the MGD, that link should be |
| 58 |
considered in the route. If however, the MGD is less than the |
| 59 |
satellite link can provide, it should not be used. For this |
| 60 |
discussion it is assumed that delay for individual links are low |
| 61 |
enough that a sending node can provide the MGD service. |
| 62 |
|
| 63 |
Low delay class of service is not the same as low Round Trip Time |
| 64 |
(RTT). Class of service is unidirectional. The datagrams responding |
| 65 |
to low delay traffic (i.e., Acking the data) might be sent with a |
| 66 |
high reliability class of service, but not low delay. |
| 67 |
|
| 68 |
The performance of TCP might be significantly improved with an |
| 69 |
accurate estimate of the round trip time and the retransmission |
| 70 |
timeout. The TCP retransmission timeout could be set to the maximum |
| 71 |
delay for the current route (if the current route could be |
| 72 |
determined). The timeout value would have to be redetermined when |
| 73 |
the number of hops in the route changes. |
| 74 |
|
| 75 |
High throughput class of service should get a large volume of data |
| 76 |
through the Internet. Requesters of this class are less concerned |
| 77 |
with the delay the datagrams have crossing the Internet and the |
| 78 |
reliability of their delivery. This type of traffic might be served |
| 79 |
well by a satellite link, especially if the bandwidth is high. |
| 80 |
Another attribute this class might have is consistent one way |
| 81 |
traversal time for a given burst of datagrams. This class of service |
| 82 |
will have its traversal times affected by the amount of Internet |
| 83 |
load. As the Internet load goes up, the throughput for each source |
| 84 |
will go down. |
| 85 |
|
| 86 |
High reliability class of service should see most of its datagrams |
| 87 |
delivered if the Internet is not too heavily loaded. Source Quenches |
| 88 |
(SQ) should not be sent only when datagrams are discarded. SQs |
| 89 |
should be sent well before the queues become full, to advise the |
| 90 |
sender of the rate that can be currently supported. |
| 91 |
|
| 92 |
Priority service should allow data that has a higher priority to be |
| 93 |
queued ahead of other lower priority data. It is important to limit |
| 94 |
the amount of priority data. The amount of preemption a lower |
| 95 |
priority datagram suffers must also be limited. |
| 96 |
|
| 97 |
It is assumed that a queuing algorithm provides these classes of |
| 98 |
service. For one facility to be used over another, that is, making |
| 99 |
different routing decisions based upon the TOS, requires a more |
| 100 |
sophisticated routing algorithm and larger routing database. These |
| 101 |
issues are not discussed in this document. |
| 102 |
|
| 103 |
|
| 104 |
|
| 105 |
|
| 106 |
|
| 107 |
|
| 108 |
<span class="grey">Prue & Postel [Page 2]</span></pre> |
| 109 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-3" ></span> |
| 110 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 111 |
|
| 112 |
|
| 113 |
Applications for Class of Service |
| 114 |
|
| 115 |
The following are examples of how classes of service might be used. |
| 116 |
They do not necessarily represent the best choices, but are presented |
| 117 |
only to illustrate how the different classes of service might be used |
| 118 |
to advantage. |
| 119 |
|
| 120 |
Interactive timesharing access using a line-at-a-time or character- |
| 121 |
at-a-time terminal (TTY) type of access is typically low volume |
| 122 |
typing speed input with low or high volume output. Some Internet |
| 123 |
applications use echoplex or character by character echoing of user |
| 124 |
input by the destination host. PC devices also have local files that |
| 125 |
may be uploaded to remote hosts in a streaming mode. Supporting such |
| 126 |
traffic can require several types of service. User keyboard input |
| 127 |
should be forwarded with low delay. If echoplex is used, all user |
| 128 |
characters sent and echoed should be low delay to minimize the |
| 129 |
echoing delay. The computer responses should be regular or high |
| 130 |
throughput depending upon the volume of data sent and the speed of |
| 131 |
the output device. If the computer response is a single datagram of |
| 132 |
data, the user should get low delay for the response, to minimize the |
| 133 |
human/computer interaction time. If however the output takes a while |
| 134 |
to read and digest, low delay computer responses are a waste of |
| 135 |
Internet resources. When streaming input is being sent the data |
| 136 |
should be sent requesting high throughput or regular class of |
| 137 |
service. |
| 138 |
|
| 139 |
The IBM 3270 class of terminals typically have traffic volumes |
| 140 |
greater than TTY access. Echoplex is not needed. The output devices |
| 141 |
usually handle higher speed output streams and most sites do not have |
| 142 |
the ability to stream input. Input is typically a screen at a time, |
| 143 |
but some PC implementations of 3270 use a variation of the protocol |
| 144 |
to effectively stream in volumes of data. Low delay for low volume |
| 145 |
input and output is appropriate. High throughput is appropriate for |
| 146 |
the higher volume traffic. |
| 147 |
|
| 148 |
Applications that transfer high volumes of data are typically |
| 149 |
streaming in one direction only, with acks for the data, on the |
| 150 |
return path. The data transfer should be high throughput and the |
| 151 |
acks should probably be regular class of service. Transfer |
| 152 |
initiation and termination might be served best with low delay class |
| 153 |
of service. |
| 154 |
|
| 155 |
Requests to, and responses from a time service might use low delay |
| 156 |
class of service effectively. |
| 157 |
|
| 158 |
These suggestions for class of service usage implies that the |
| 159 |
application sets the service based on the knowledge it has during the |
| 160 |
session. Thus, the application should have control of this setting |
| 161 |
|
| 162 |
|
| 163 |
|
| 164 |
<span class="grey">Prue & Postel [Page 3]</span></pre> |
| 165 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-4" ></span> |
| 166 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 167 |
|
| 168 |
|
| 169 |
dynamically for each send data request, not just on a per |
| 170 |
session/conversation/transaction basis. It would be possible for the |
| 171 |
transport level protocol to guess (i.e., TCP), but it would be sub- |
| 172 |
optimal. |
| 173 |
|
| 174 |
Algorithm |
| 175 |
|
| 176 |
When we provide class of service queuing, one class may be more |
| 177 |
desirable than the others. We must limit the amount of resources |
| 178 |
each class consumes when there is contention, so the other classes |
| 179 |
may also operate effectively. To be fair, the algorithm provides the |
| 180 |
requested service by reducing the other service attributes. A |
| 181 |
request for multiple classes of service is an OR type of request not |
| 182 |
an AND request. For example, one can not get low delay and high |
| 183 |
throughput unless there is no contention for the available resources. |
| 184 |
|
| 185 |
Low Delay Queuing |
| 186 |
|
| 187 |
To support low delay, use a limited queue so requests will not wait |
| 188 |
longer than the MGD on the queue. The low delay queue should be |
| 189 |
serviced at a lower rate than other classes of service, so low delay |
| 190 |
requests will not consume excessive resources. If the number of low |
| 191 |
delay datagrams exceeds the queue limit, discard the datagrams. The |
| 192 |
service rate should be low enough so that other data can still get |
| 193 |
through. (See discussion of service rates below.) Make the queue |
| 194 |
limit small enough so that, if the datagram is queued, it will have a |
| 195 |
guaranteed transit time (MGD). It seems unlikely that Source Quench |
| 196 |
flow control mechanisms will be an effective method of flow control |
| 197 |
because of the small size of the queue. It should not be done for |
| 198 |
this class of service. Instead, datagrams should just be discarded |
| 199 |
as required. If the bandwidth or percentage allocated to low delay |
| 200 |
is such that a large queue is possible (see formula below), SQs |
| 201 |
should be reconsidered. |
| 202 |
|
| 203 |
The maximum delay a datagram with low delay class of service will |
| 204 |
experience (MGD), can be determined with the following information: |
| 205 |
|
| 206 |
N = Queue size for low delay queue |
| 207 |
P = Percentage of link resources allocated to low delay |
| 208 |
R = Link rate (in datagrams/sec.) |
| 209 |
N |
| 210 |
Max Delay = ----- |
| 211 |
P * R |
| 212 |
|
| 213 |
If Max Delay is held fixed, then as P and R go up, so does N. It is |
| 214 |
probable that low delay service datagrams will prove to be, on the |
| 215 |
average, smaller than other traffic. This means that the number of |
| 216 |
datagrams that can be sent in the allocated bandwidth can be larger. |
| 217 |
|
| 218 |
|
| 219 |
|
| 220 |
<span class="grey">Prue & Postel [Page 4]</span></pre> |
| 221 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-5" ></span> |
| 222 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 223 |
|
| 224 |
|
| 225 |
High Reliability Queuing |
| 226 |
|
| 227 |
To support high reliability class of service, use a queue that is |
| 228 |
longer than normal (longer queue means higher potential delay). Send |
| 229 |
SQ earlier (smaller percentage of max queue length) and don't discard |
| 230 |
datagrams until the queue is full. This queue should have a lower |
| 231 |
service rate than high throughput class of service. |
| 232 |
|
| 233 |
Users of this class of service should specify a Time-to-Live (TTL) |
| 234 |
which is made appropriately longer so that it will survive longer |
| 235 |
queueing times for this class of service. |
| 236 |
|
| 237 |
This queuing procedure will only be effective for Internet |
| 238 |
unreliability due to congestion. Other Internet unreliability |
| 239 |
problems such as high error rate links or reliability features such |
| 240 |
as forward error correcting modems must be dealt with by more |
| 241 |
sophisticated routing algorithms. |
| 242 |
|
| 243 |
High Throughput Queuing |
| 244 |
|
| 245 |
To support high throughput class of service have a queue that is |
| 246 |
treated like current IP queuing. It should have the highest service |
| 247 |
rate. It will experience higher average through node delay than low |
| 248 |
delay because of the larger queue size. |
| 249 |
|
| 250 |
Another thing that might be done, is to keep datagrams of the same |
| 251 |
burst together when possible. This must be done in a way that will |
| 252 |
not block other traffic. The idea is to deliver all the data to the |
| 253 |
other end in a contiguous burst. This could be an advantage by |
| 254 |
allowing piggybacking acks for the whole burst at one time. This |
| 255 |
makes some assumptions about the overlying protocol which may be |
| 256 |
inappropriate. |
| 257 |
|
| 258 |
Regular Service Queuing |
| 259 |
|
| 260 |
For datagrams which request none of the three classes of service, |
| 261 |
queue the datagrams on the queue representing the least delay between |
| 262 |
the two queues, the high throughput queue or the high reliability |
| 263 |
queue. If one queue becomes full, queue on the other. If both |
| 264 |
queues are full, follow the source quench procedure for regular class |
| 265 |
of service (see <a href="./rfc1016">RFC-1016</a>), not the procedure for the queue the |
| 266 |
datagram failed to attain. |
| 267 |
|
| 268 |
In the discussion of service rates described below, it is proposed |
| 269 |
that the high throughput queue get service three times for every two |
| 270 |
times for the high reliability queue. Therefore, the queue length of |
| 271 |
the high reliability queue should be increased by 50% (in this |
| 272 |
example) to compare the lengths of the two queues more accurately. A |
| 273 |
|
| 274 |
|
| 275 |
|
| 276 |
<span class="grey">Prue & Postel [Page 5]</span></pre> |
| 277 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-6" ></span> |
| 278 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 279 |
|
| 280 |
|
| 281 |
simplification to this method is to just queue new data on the queue |
| 282 |
that is the shortest. The slower service rate queue will quickly |
| 283 |
exceed the size of the faster service rate queue and new data will go |
| 284 |
on the proper queue. This however, would lead to more packet |
| 285 |
reordering than the first method. |
| 286 |
|
| 287 |
|
| 288 |
Service Rates |
| 289 |
|
| 290 |
In this discussion, a higher service rate means that a queue, when |
| 291 |
non-empty, will consume a larger percentage of the available |
| 292 |
bandwidth than a lower service rate queue. It will not block a lower |
| 293 |
service rate queue even if it is always full. |
| 294 |
|
| 295 |
For example, the service pattern could be; send low delay 17% of the |
| 296 |
time, high throughput 50% of the time, and high reliability 33% of |
| 297 |
the time. Throughput requires the most bandwidth and high |
| 298 |
reliability requires medium bandwidth. One could achieve this split |
| 299 |
using a pattern of L, R,R, T,T,T, where low delay is "L", high |
| 300 |
reliability is "R", and high throughput is "T'. We want to keep the |
| 301 |
high throughput datagrams together. We therefore send all of the |
| 302 |
high throughput data at one time, that is, not interspersed with the |
| 303 |
other classes of service. By keeping all of the high throughput data |
| 304 |
together, we may help higher level protocols, such as TCP, as |
| 305 |
described above. This would still be done in a way to not exceed the |
| 306 |
allowed service rate of the available bandwidth. |
| 307 |
|
| 308 |
These service rates are suggestions. Some simplifications can be |
| 309 |
considered, such as having only two routing classes; low delay, and |
| 310 |
other. |
| 311 |
|
| 312 |
Priority |
| 313 |
|
| 314 |
There is the ability to select 8 levels of priority 0-7, in addition |
| 315 |
to the class of service selected. To provide this without blocking |
| 316 |
the least priority requests, we must give preempted datagrams |
| 317 |
frustration points every time a higher priority request cuts in line |
| 318 |
in front of it. Thus if a datagram with low priority waits, it will |
| 319 |
always get through even when competing against the highest priority |
| 320 |
requests. This assumes the TTL (Time-to-Live) field does not expire. |
| 321 |
|
| 322 |
When a datagram with priority arrives at a node, the node will queue |
| 323 |
the datagram on the appropriate queue ahead of all datagrams with |
| 324 |
lower priority. Each datagram that was preempted gets its priority |
| 325 |
raised (locally). The priority data will not bump a lower priority |
| 326 |
datagram off its queue, discarding the data. If the queue is full, |
| 327 |
the newest data (priority or not) will be discarded. The priority |
| 328 |
preemption will preempt only within the class of service queue to |
| 329 |
|
| 330 |
|
| 331 |
|
| 332 |
<span class="grey">Prue & Postel [Page 6]</span></pre> |
| 333 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-7" ></span> |
| 334 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 335 |
|
| 336 |
|
| 337 |
which the priority data is targeted. A request specifying regular |
| 338 |
class of service, will contend on the queue where it is placed, high |
| 339 |
throughput or high reliability. |
| 340 |
|
| 341 |
An implementation strategy is to multiply the requested priority by 2 |
| 342 |
or 4, then store the value in a buffer overhead area. Each time the |
| 343 |
datagram is preempted, increment the value by one. Looking at an |
| 344 |
example, assume we use a multiplier of 2. A priority 6 buffer will |
| 345 |
have an initial local value of 12. A new priority 7 datagram would |
| 346 |
have a local value of 14. If 2 priority 7 datagrams arrive, |
| 347 |
preempting the priority 6 datagram, its local value is incremented to |
| 348 |
14. It can no longer be preempted. After that, it has the same |
| 349 |
local value as a priority 7 datagram and will no longer be preempted |
| 350 |
within this node. In our example, this means that a priority 0 |
| 351 |
datagram can be preempted by no more than 14 higher priority |
| 352 |
datagrams. The priority is raised only locally in the node. The |
| 353 |
datagram could again be preempted in the next node on the route. |
| 354 |
|
| 355 |
Priority queuing changes the effects we were obtaining with the low |
| 356 |
delay queuing described above. Once a buffer was queued, the delay |
| 357 |
that a datagram would see could be determined. When we accepted low |
| 358 |
delay data, we could guarantee a certain maximum delay. With this |
| 359 |
addition, if the datagram requesting low delay does not also request |
| 360 |
high priority, the guaranteed delay can vary a lot more. It could be |
| 361 |
1 up to 28 times as much as without priority queuing. |
| 362 |
|
| 363 |
Discussion and Details |
| 364 |
|
| 365 |
If a low delay queue is for a satellite link (or any high delay |
| 366 |
link), the max queue size should be reduced by the number of |
| 367 |
datagrams that can be forwarded from the queue during the one way |
| 368 |
delay for the link. That is, if the service rate for the low delay |
| 369 |
queue is L datagrams per second, the delay added by the high delay |
| 370 |
link is D seconds and M is the max delay per node allowed (MGD) in |
| 371 |
seconds, then the maximum queue size should be: |
| 372 |
|
| 373 |
Max Queue Size = L ( M - D), M > D |
| 374 |
= 0 , M <= D |
| 375 |
|
| 376 |
If the result is negative (M is less than the delay introduced by the |
| 377 |
link), then the maximum queue size should be zero because the link |
| 378 |
could never provide a delay less than the guaranteed M value. If the |
| 379 |
bandwidth is high (as in T1 links), the delay introduced by a |
| 380 |
terrestrial link and the terminating equipment could be significant |
| 381 |
and greater than the average service time for a single datagram on |
| 382 |
the low delay queue. If so, this formula should be used to reduce |
| 383 |
the queue size as well. Note that this is reducing the queue size |
| 384 |
and is not the same as the allocated bandwidth. Even though the |
| 385 |
|
| 386 |
|
| 387 |
|
| 388 |
<span class="grey">Prue & Postel [Page 7]</span></pre> |
| 389 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-8" ></span> |
| 390 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 391 |
|
| 392 |
|
| 393 |
queue size is reduced, the chit scheme described below will give low |
| 394 |
delay requesters a chance to use the allocated bandwidth. |
| 395 |
|
| 396 |
If a datagram requests multiple classes of service, only one class |
| 397 |
can be provided. For example, when both low delay and high |
| 398 |
reliability classes are requested, and if the low delay queue is |
| 399 |
full, queue the data on the high reliability queue instead. If we |
| 400 |
are able to queue the data on the low delay queue, then the datagram |
| 401 |
gets part of the high reliability service it also requested, because, |
| 402 |
once data is queued, data will not be discarded. However, the |
| 403 |
datagram will be routed as a low delay request. The same scheme is |
| 404 |
used for any other combinations of service requested. The order of |
| 405 |
selection for classes of service when more than one is requested |
| 406 |
would be low delay, high throughput, then high reliability. If a |
| 407 |
block of datagrams request multiple classes of service, it is quite |
| 408 |
possible that datagram reordering will occur. If one queue is full |
| 409 |
causing the other queue to be used for some of the data, data will be |
| 410 |
forwarded at different service rates. Requesting multiple classes of |
| 411 |
service gives the data a better chance of making it through the net |
| 412 |
because they have multiple chances of getting on a service queue. |
| 413 |
However, the datagrams pay the penalty of possible reordering and |
| 414 |
more variability in the one way transmission times. |
| 415 |
|
| 416 |
Besides total buffer consumption, individual class of service queue |
| 417 |
sizes should be used to SQ those asking for service except as noted |
| 418 |
above. |
| 419 |
|
| 420 |
A request for regular class of service is handled by queuing to the |
| 421 |
high reliability or high throughput queues evenly (proportional to |
| 422 |
the service rates of queue). The low delay queue should only receive |
| 423 |
data with the low delay service type. Its queue is too small to |
| 424 |
accept other traffic. |
| 425 |
|
| 426 |
Because of the small queue size for low delay suggested above, it is |
| 427 |
difficult for low delay service requests to consume the bandwidth |
| 428 |
allocated. To do so, low delay users must keep the small queue |
| 429 |
continuously non-empty. This is hard to do with a small queue. |
| 430 |
Traffic flow has been shown to be bursty in nature. In order for the |
| 431 |
low delay queue to be able to consume the allocated bandwidth, a |
| 432 |
count of the various types being forwarded should be kept. The |
| 433 |
service rate should increase if the actual percentage falls too low |
| 434 |
for the low delay queue. The measure of service rates would have to |
| 435 |
be smoothed over time. |
| 436 |
|
| 437 |
While this does sound complicated, a reasonably efficient way can be |
| 438 |
described. Every Q seconds, where Q is less than or equal to the |
| 439 |
MGD, each class gets N M P chits proportional to their allowed |
| 440 |
percentage. Send data for the low delay queue up to the number of |
| 441 |
|
| 442 |
|
| 443 |
|
| 444 |
<span class="grey">Prue & Postel [Page 8]</span></pre> |
| 445 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-9" ></span> |
| 446 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 447 |
|
| 448 |
|
| 449 |
chits it receives decrementing the chits as datagrams are sent. Next |
| 450 |
send from the high reliability queue as many as it has chits for. |
| 451 |
Finally, send from the high throughput queue. At this point, each |
| 452 |
queue gets N M P chits again. If the low delay queue does not |
| 453 |
consume all of its chits, when a low delay datagram arrives, before |
| 454 |
chit replenishment, send from the low delay queue immediately. This |
| 455 |
provides some smoothing of the actual bandwidth made available for |
| 456 |
low delay traffic. If operational experience shows that low delay |
| 457 |
requests are experiencing excessive congestion loss but still not |
| 458 |
consuming the classes allocated bandwidth, adjustments should be |
| 459 |
made. The service rates should be made larger and the queue sizes |
| 460 |
adjusted accordingly. This is more important on lower speed links |
| 461 |
where the above formula makes the queue small. |
| 462 |
|
| 463 |
What we should see during the Q seconds is that low delay data will |
| 464 |
be sent as soon as possible (as long as the volume is below the |
| 465 |
allowed percentage). Also, the tendency will be to send all the high |
| 466 |
throughput datagrams contiguously. This will give a more regular |
| 467 |
measured round trip time for bursts of datagrams. Classes of service |
| 468 |
will tend to be grouped together at each intermediate node in the |
| 469 |
route. If all of the queues with datagrams have consumed all of |
| 470 |
their allocated chits, but one or more classes with empty queues have |
| 471 |
unused chits then a percentage of these left over chits should be |
| 472 |
carried over. Divide the remaining chit counts by two (with round |
| 473 |
down), then add in the refresh chit counts. This allows a 50% carry |
| 474 |
over for the next interval. The carry over is self limiting to less |
| 475 |
than or equal to the refresh chit count. This prevents excessive |
| 476 |
build up. It provides some smoothing of the percentage allocation |
| 477 |
over time but will not allow an unused queue to build up chits |
| 478 |
indefinitely. No timer is required. |
| 479 |
|
| 480 |
If only a simple subset of the described algorithm is to be |
| 481 |
implemented, then low delay queuing would be the best choice. One |
| 482 |
should use a small queue. Service the queue with a high service rate |
| 483 |
but restrict the bandwidth to a small reasonable percentage of the |
| 484 |
available bandwidth. Currently, wide area networks with high traffic |
| 485 |
volumes do not provide low delay service unless low delay requests |
| 486 |
are able to preempt other traffic. |
| 487 |
|
| 488 |
Applicability |
| 489 |
|
| 490 |
When the output speed and volume match the input speed and volume, |
| 491 |
queues don't get large. If the queues never grow large enough to |
| 492 |
exceed the guaranteed low delay performance, no queuing algorithm |
| 493 |
other than first in, first out, should be used. |
| 494 |
|
| 495 |
The algorithm could be turned on when the main queue size exceeds a |
| 496 |
certain threshold. The routing node can periodically check for queue |
| 497 |
|
| 498 |
|
| 499 |
|
| 500 |
<span class="grey">Prue & Postel [Page 9]</span></pre> |
| 501 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-10" ></span> |
| 502 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 503 |
|
| 504 |
|
| 505 |
build up. This queuing algorithm can be turned on when the maximum |
| 506 |
delays will exceed the allowed nodal delay for low delay class of |
| 507 |
service. It can also be turned off when queue sizes are no longer a |
| 508 |
problem. |
| 509 |
|
| 510 |
Issues |
| 511 |
|
| 512 |
Several issues need to be addressed before type of service queuing as |
| 513 |
described should be implemented. What percentage of the bandwidth |
| 514 |
should each class of service consume assuming an infinite supply of |
| 515 |
each class of service datagrams? What maximum delay (MGD) should be |
| 516 |
guaranteed per node for low delay datagrams? |
| 517 |
|
| 518 |
It is possible to provide a more optimal route if the queue sizes for |
| 519 |
each class of service are considered in the routing decision. This, |
| 520 |
however, adds additional overhead and complexity to each routing |
| 521 |
node. This may be an unacceptable additional complexity. |
| 522 |
|
| 523 |
How are we going to limit the use of more desirable classes of |
| 524 |
service and higher priorities? The algorithm limits use of the |
| 525 |
various classes by restricting queue sizes especially the low delay |
| 526 |
queue size. This helps but it seems likely we will want to |
| 527 |
instrument the number of datagrams requesting each Type-of-Service |
| 528 |
and priority. When a datagram requests multiple classes of service, |
| 529 |
increment the instrumentation count once based upon the queue |
| 530 |
actually used, selecting, low delay, high throughput, high |
| 531 |
reliability, then regular. If instrumentation reveals an excessive |
| 532 |
imbalance, Internet operations can give this to administrators to |
| 533 |
handle. This instrumentation will show the distribution for types of |
| 534 |
service requested by the Internet users. This information can be |
| 535 |
used to tune the Internet to service the user demands. |
| 536 |
|
| 537 |
Will the routing algorithms in use today have problems when routing |
| 538 |
data with this algorithm? Simulation tests need to be done to model |
| 539 |
how the Internet will react. If, for example, an application |
| 540 |
requests multiple classes of service, round trip times may fluctuate |
| 541 |
significantly. Would TCP have to be more sophisticated in its round |
| 542 |
trip time estimator? |
| 543 |
|
| 544 |
An objection to this type of queuing algorithm is that it is making |
| 545 |
the routing and queuing more complicated. There is current interest |
| 546 |
in high speed packet switches which have very little protocol |
| 547 |
overhead when handling/routing packets. This algorithm complicates |
| 548 |
not simplifies the protocol. The bandwidth being made available is |
| 549 |
increasing. More T1 (1.5 Mbps) and higher speed links are being used |
| 550 |
all the time. However, in the history of communications, it seems |
| 551 |
that the demand for bandwidth has always exceeded the supply. When |
| 552 |
there is wide spread use of optical fiber we may temporarily |
| 553 |
|
| 554 |
|
| 555 |
|
| 556 |
<span class="grey">Prue & Postel [Page 10]</span></pre> |
| 557 |
<hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-11" ></span> |
| 558 |
<span class="grey"><a href="./rfc1046">RFC 1046</a> Type-of-Service Queuing February 1988</span> |
| 559 |
|
| 560 |
|
| 561 |
experience a glut of capacity. As soon as 1 gigabit optical fiber |
| 562 |
link becomes reasonably priced, new applications will be created to |
| 563 |
consume it all. A single full motion high resolution color image |
| 564 |
system can consume, as an upper limit, nearly a gigabit per second |
| 565 |
channel (30 fps X 24 b/pixel X 1024 X 1024 pixels). |
| 566 |
|
| 567 |
In the study of one gateway, Dave Clark discovered that the per |
| 568 |
datagram processing of the IP header constituted about 20% of the |
| 569 |
processing time. Much of the time per datagram was spent on |
| 570 |
restarting input, starting output and queuing datagrams. He thought |
| 571 |
that a small additional amount of processing to support Type-of- |
| 572 |
Service would be reasonable. He suggests that even if the code does |
| 573 |
slow the gateway down, we need to see if TOS is good for anything, so |
| 574 |
this experiment is valuable. To support the new high speed |
| 575 |
communications of the near future, Dave wants to see switches which |
| 576 |
will run one to two orders of magnitude faster. This can not be done |
| 577 |
by trimming a few instructions here or there. |
| 578 |
|
| 579 |
From a practical perspective, the problem this algorithm is trying to |
| 580 |
solve is the lack of low delay service through the Internet today. |
| 581 |
Implementing only the low delay queuing portion of this algorithm |
| 582 |
would allow the Internet to provide a class of service it otherwise |
| 583 |
could not provide. Requesters of this class of service would not get |
| 584 |
it for free. Low delay class of datagram streams get low delay at |
| 585 |
the cost of reliability and throughput. |
| 586 |
|
| 587 |
|
| 588 |
|
| 589 |
|
| 590 |
|
| 591 |
|
| 592 |
|
| 593 |
|
| 594 |
|
| 595 |
|
| 596 |
|
| 597 |
|
| 598 |
|
| 599 |
|
| 600 |
|
| 601 |
|
| 602 |
|
| 603 |
|
| 604 |
|
| 605 |
|
| 606 |
|
| 607 |
|
| 608 |
|
| 609 |
|
| 610 |
|
| 611 |
|
| 612 |
Prue & Postel [Page 11] |
| 613 |
</pre> |