Debian Codegraph @

rfc1633.html

raw
1 <pre>Network Working Group R. Braden
2 Request for Comments: 1633 ISI
3 Category: Informational D. Clark
4 MIT
5 S. Shenker
6 Xerox PARC
7 June 1994
8
9
10 <span class="h1">Integrated Services in the Internet Architecture: an Overview</span>
11
12 Status of this Memo
13
14 This memo provides information for the Internet community. This memo
15 does not specify an Internet standard of any kind. Distribution of
16 this memo is unlimited.
17
18 Abstract
19
20 This memo discusses a proposed extension to the Internet architecture
21 and protocols to provide integrated services, i.e., to support real-
22 time as well as the current non-real-time service of IP. This
23 extension is necessary to meet the growing need for real-time service
24 for a variety of new applications, including teleconferencing, remote
25 seminars, telescience, and distributed simulation.
26
27 This memo represents the direct product of recent work by Dave Clark,
28 Scott Shenker, Lixia Zhang, Deborah Estrin, Sugih Jamin, John
29 Wroclawski, Shai Herzog, and Bob Braden, and indirectly draws upon
30 the work of many others.
31
32 Table of Contents
33
34 <a href="#section-1">1</a>. Introduction ...................................................<a href="#page-2">2</a>
35 <a href="#section-2">2</a>. Elements of the Architecture ...................................<a href="#page-3">3</a>
36 <a href="#section-2.1">2.1</a> Integrated Services Model ..................................<a href="#page-3">3</a>
37 <a href="#section-2.2">2.2</a> Reference Implementation Framework .........................<a href="#page-6">6</a>
38 <a href="#section-3">3</a>. Integrated Services Model ......................................<a href="#page-11">11</a>
39 <a href="#section-3.1">3.1</a> Quality of Service Requirements ............................<a href="#page-12">12</a>
40 <a href="#section-3.2">3.2</a> Resource-Sharing Requirements and Service Models ...........<a href="#page-16">16</a>
41 <a href="#section-3.3">3.3</a> Packet Dropping ............................................<a href="#page-18">18</a>
42 <a href="#section-3.4">3.4</a> Usage Feedback .............................................<a href="#page-19">19</a>
43 <a href="#section-3.5">3.5</a> Reservation Model ..........................................<a href="#page-19">19</a>
44 <a href="#section-4">4</a>. Traffic Control Mechanisms .....................................<a href="#page-20">20</a>
45 <a href="#section-4.1">4.1</a> Basic Functions ............................................<a href="#page-20">20</a>
46 <a href="#section-4.2">4.2</a> Applying the Mechanisms ....................................<a href="#page-23">23</a>
47 <a href="#section-4.3">4.3</a> An example .................................................<a href="#page-24">24</a>
48 <a href="#section-5">5</a>. Reservation Setup Protocol .....................................<a href="#page-25">25</a>
49
50
51
52 <span class="grey">Braden, Clark &amp; Shenker [Page 1]</span></pre>
53 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-2" ></span>
54 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
55
56
57 <a href="#section-5.1">5.1</a> RSVP Overview ..............................................<a href="#page-25">25</a>
58 <a href="#section-5.2">5.2</a> Routing and Reservations ...................................<a href="#page-28">28</a>
59 <a href="#section-6">6</a>. Acknowledgments ................................................<a href="#page-30">30</a>
60 References ........................................................<a href="#page-31">31</a>
61 Security Considerations ...........................................<a href="#page-32">32</a>
62 Authors' Addresses ................................................<a href="#page-33">33</a>
63
64 <span class="h2"><a class="selflink" id="section-1" href="#section-1">1</a>. Introduction</span>
65
66 The multicasts of IETF meetings across the Internet have formed a
67 large-scale experiment in sending digitized voice and video through a
68 packet-switched infrastructure. These highly-visible experiments
69 have depended upon three enabling technologies. (1) Many modern
70 workstations now come equipped with built-in multimedia hardware,
71 including audio codecs and video frame-grabbers, and the necessary
72 video gear is now inexpensive. (2) IP multicasting, which is not yet
73 generally available in commercial routers, is being provided by the
74 MBONE, a temporary "multicast backbone". (3) Highly-sophisticated
75 digital audio and video applications have been developed.
76
77 These experiments also showed that an important technical element is
78 still missing: real-time applications often do not work well across
79 the Internet because of variable queueing delays and congestion
80 losses. The Internet, as originally conceived, offers only a very
81 simple quality of service (QoS), point-to-point best-effort data
82 delivery. Before real-time applications such as remote video,
83 multimedia conferencing, visualization, and virtual reality can be
84 broadly used, the Internet infrastructure must be modified to support
85 real-time QoS, which provides some control over end-to-end packet
86 delays. This extension must be designed from the beginning for
87 multicasting; simply generalizing from the unicast (point-to-point)
88 case does not work.
89
90 Real-time QoS is not the only issue for a next generation of traffic
91 management in the Internet. Network operators are requesting the
92 ability to control the sharing of bandwidth on a particular link
93 among different traffic classes. They want to be able to divide
94 traffic into a few administrative classes and assign to each a
95 minimum percentage of the link bandwidth under conditions of
96 overload, while allowing "unused" bandwidth to be available at other
97 times. These classes may represent different user groups or
98 different protocol families, for example. Such a management facility
99 is commonly called controlled link-sharing. We use the term
100 integrated services (IS) for an Internet service model that includes
101 best-effort service, real-time service, and controlled link sharing.
102
103 The requirements and mechanisms for integrated services have been the
104 subjects of much discussion and research over the past several years
105
106
107
108 <span class="grey">Braden, Clark &amp; Shenker [Page 2]</span></pre>
109 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-3" ></span>
110 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
111
112
113 (the literature is much too large to list even a representative
114 sample here; see the references in [CSZ92, Floyd92, Jacobson91,
115 JSCZ93, Partridge92, SCZ93, RSVP93a] for a partial list). This work
116 has led to the unified approach to integrated services support that
117 is described in this memo. We believe that it is now time to begin
118 the engineering that must precede deployment of integrated services
119 in the Internet.
120
121 <a href="#section-2">Section 2</a> of this memo introduces the elements of an IS extension of
122 the Internet. <a href="#section-3">Section 3</a> discusses real-time service models [SCZ93a,
123 SCZ93b]. <a href="#section-4">Section 4</a> discusses traffic control, the forwarding
124 algorithms to be used in routers [<a href="#ref-CSZ92" title="&quot;Supporting Real-Time Applications in an Integrated Services Packet Network: Architecture and Mechanisms&quot;">CSZ92</a>]. <a href="#section-5">Section 5</a> discusses the
125 design of RSVP, a resource setup protocol compatible with the
126 assumptions of our IS model [<a href="#ref-RSVP93a" title="&quot;RSVP: A New Resource ReSerVation Protocol&quot;">RSVP93a</a>, <a href="#ref-RSVP93b" title="&quot;Resource ReSerVation Protocol (RSVP) - Version 1 Functional Specification&quot;">RSVP93b</a>].
127
128 <span class="h2"><a class="selflink" id="section-2" href="#section-2">2</a>. Elements of the Architecture</span>
129
130 The fundamental service model of the Internet, as embodied in the
131 best-effort delivery service of IP, has been unchanged since the
132 beginning of the Internet research project 20 years ago [<a href="#ref-CerfKahn74" title="&quot;A Protocol for Packet Network Intercommunication&quot;">CerfKahn74</a>].
133 We are now proposing to alter that model to encompass integrated
134 service. From an academic viewpoint, changing the service model of
135 the Internet is a major undertaking; however, its impact is mitigated
136 by the fact that we wish only to extend the original architecture.
137 The new components and mechanisms to be added will supplement but not
138 replace the basic IP service.
139
140 Abstractly, the proposed architectural extension is comprised of two
141 elements: (1) an extended service model, which we call the IS model,
142 and (2) a reference implementation framework, which gives us a set of
143 vocabulary and a generic program organization to realize the IS
144 model. It is important to separate the service model, which defines
145 the externally visible behavior, from the discussion of the
146 implementation, which may (and should) change during the life of the
147 service model. However, the two are related; to make the service
148 model credible, it is useful to provide an example of how it might be
149 realized.
150
151 2.1 Integrated Services Model
152
153 The IS model we are proposing includes two sorts of service
154 targeted towards real-time traffic: guaranteed and predictive
155 service. It integrates these services with controlled link-
156 sharing, and it is designed to work well with multicast as well as
157 unicast. Deferring a summary of the IS model to <a href="#section-3">Section 3</a>, we
158 first discuss some key assumptions behind the model.
159
160
161
162
163
164 <span class="grey">Braden, Clark &amp; Shenker [Page 3]</span></pre>
165 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-4" ></span>
166 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
167
168
169 The first assumption is that resources (e.g., bandwidth) must be
170 explicitly managed in order to meet application requirements.
171 This implies that "resource reservation" and "admission control"
172 are key building blocks of the service. An alternative approach,
173 which we reject, is to attempt to support real-time traffic
174 without any explicit changes to the Internet service model.
175
176 The essence of real-time service is the requirement for some
177 service guarantees, and we argue that guarantees cannot be
178 achieved without reservations. The term "guarantee" here is to be
179 broadly interpreted; they may be absolute or statistical, strict
180 or approximate. However, the user must be able to get a service
181 whose quality is sufficiently predictable that the application can
182 operate in an acceptable way over a duration of time determined by
183 the user. Again, "sufficiently" and "acceptable" are vague terms.
184 In general, stricter guarantees have a higher cost in resources
185 that are made unavailable for sharing with others.
186
187 The following arguments have been raised against resource
188 guarantees in the Internet.
189
190 o "Bandwidth will be infinite."
191
192 The incredibly large carrying capacity of an optical fiber
193 leads some to conclude that in the future bandwidth will be
194 so abundant, ubiquitous, and cheap that there will be no
195 communication delays other than the speed of light, and
196 therefore there will be no need to reserve resources.
197 However, we believe that this will be impossible in the short
198 term and unlikely in the medium term. While raw bandwidth
199 may seem inexpensive, bandwidth provided as a network service
200 is not likely to become so cheap that wasting it will be the
201 most cost-effective design principle. Even if low-cost
202 bandwidth does eventually become commonly available, we do
203 not accept that it will be available "everywhere" in the
204 Internet. Unless we provide for the possibility of dealing
205 with congested links, then real-time services will simply be
206 precluded in those cases. We find that restriction
207 unacceptable.
208
209 o "Simple priority is sufficient."
210
211 It is true that simply giving higher priority to real-time
212 traffic would lead to adequate real-time service at some
213 times and under some conditions. But priority is an
214 implementation mechanism, not a service model. If we define
215 the service by means of a specific mechanism, we may not get
216 the exact features we want. In the case of simple priority,
217
218
219
220 <span class="grey">Braden, Clark &amp; Shenker [Page 4]</span></pre>
221 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-5" ></span>
222 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
223
224
225 the issue is that as soon as there are too many real-time
226 streams competing for the higher priority, every stream is
227 degraded. Restricting our service to this single failure
228 mode is unacceptable. In some cases, users will demand that
229 some streams succeed while some new requests receive a "busy
230 signal".
231
232 o "Applications can adapt."
233
234 The development of adaptive real-time applications, such as
235 Jacobson's audio program VAT, does not eliminate the need to
236 bound packet delivery time. Human requirements for
237 interaction and intelligibility limit the possible range of
238 adaptation to network delays. We have seen in real
239 experiments that, while VAT can adapt to network delays of
240 many seconds, the users find that interaction is impossible
241 in these cases.
242
243 We conclude that there is an inescapable requirement for routers
244 to be able to reserve resources, in order to provide special QoS
245 for specific user packet streams, or "flows". This in turn
246 requires flow-specific state in the routers, which represents an
247 important and fundamental change to the Internet model. The
248 Internet architecture was been founded on the concept that all
249 flow-related state should be in the end systems [<a href="#ref-Clark88" title="&quot;The Design Philosophy of the DARPA Internet Protocols&quot;">Clark88</a>].
250 Designing the TCP/IP protocol suite on this concept led to a
251 robustness that is one of the keys to its success. In <a href="#section-5">section 5</a>
252 we discuss how the flow state added to the routers for resource
253 reservation can be made "soft", to preserve the robustness of the
254 Internet protocol suite.
255
256 There is a real-world side effect of resource reservation in
257 routers. Since it implies that some users are getting privileged
258 service, resource reservation will need enforcement of policy and
259 administrative controls. This in turn will lead to two kinds of
260 authentication requirements: authentication of users who make
261 reservation requests, and authentication of packets that use the
262 reserved resources. However, these issues are not unique to "IS";
263 other aspects of the evolution of the Internet, including
264 commercialization and commercial security, are leading to the same
265 requirements. We do not discuss the issues of policy or security
266 further in this memo, but they will require attention.
267
268 We make another fundamental assumption, that it is desirable to
269 use the Internet as a common infrastructure to support both non-
270 real-time and real-time communication. One could alternatively
271 build an entirely new, parallel infrastructure for real-time
272 services, leaving the Internet unchanged. We reject this
273
274
275
276 <span class="grey">Braden, Clark &amp; Shenker [Page 5]</span></pre>
277 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-6" ></span>
278 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
279
280
281 approach, as it would lose the significant advantages of
282 statistical sharing between real-time and non-real-time traffic,
283 and it would be much more complex to build and administer than a
284 common infrastructure.
285
286 In addition to this assumption of common infrastructure, we adopt
287 a unified protocol stack model, employing a single internet-layer
288 protocol for both real-time and non-real-time service. Thus, we
289 propose to use the existing internet-layer protocol (e.g., IP or
290 CLNP) for real-time data. Another approach would be to add a new
291 real-time protocol in the internet layer [<a href="#ref-ST2-90" title="&quot;Experimental Internet Stream Protocol: Version 2 (ST-II)&quot;">ST2-90</a>]. Our unified
292 stack approach provides economy of mechanism, and it allows us to
293 fold controlled link-sharing in easily. It also handles the
294 problem of partial coverage, i.e., allowing interoperation between
295 IS-capable Internet systems and systems that have not been
296 extended, without the complexity of tunneling.
297
298 We take the view that there should be a single service model for
299 the Internet. If there were different service models in different
300 parts of the Internet, it is very difficult to see how any end-
301 to-end service quality statements could be made. However, a
302 single service model does not necessarily imply a single
303 implementation for packet scheduling or admission control.
304 Although specific packet scheduling and admission control
305 mechanisms that satisfy our service model have been developed, it
306 is quite possible that other mechanisms will also satisfy the
307 service model. The reference implementation framework, introduced
308 below, is intended to allow discussion of implementation issues
309 without mandating a single design.
310
311 Based upon these considerations, we believe that an IS extension
312 that includes additional flow state in routers and an explicit
313 setup mechanism is necessary to provide the needed service. A
314 partial solution short of this point would not be a wise
315 investment. We believe that the extensions we propose preserve
316 the essential robustness and efficiency of the Internet
317 architecture, and they allow efficient management of the network
318 resources; these will be important goals even if bandwidth becomes
319 very inexpensive.
320
321 2.2 Reference Implementation Framework
322
323 We propose a reference implementation framework to realize the IS
324 model. This framework includes four components: the packet
325 scheduler, the admission control routine, the classifier, and the
326 reservation setup protocol. These are discussed briefly below and
327 more fully in Sections <a href="#section-4">4</a> and <a href="#section-5">5</a>.
328
329
330
331
332 <span class="grey">Braden, Clark &amp; Shenker [Page 6]</span></pre>
333 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-7" ></span>
334 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
335
336
337 In the ensuing discussion, we define the "flow" abstraction as a
338 distinguishable stream of related datagrams that results from a
339 single user activity and requires the same QoS. For example, a
340 flow might consist of one transport connection or one video stream
341 between a given host pair. It is the finest granularity of packet
342 stream distinguishable by the IS. We define a flow to be simplex,
343 i.e., to have a single source but N destinations. Thus, an N-way
344 teleconference will generally require N flows, one originating at
345 each site.
346
347 In today's Internet, IP forwarding is completely egalitarian; all
348 packets receive the same quality of service, and packets are
349 typically forwarded using a strict FIFO queueing discipline. For
350 integrated services, a router must implement an appropriate QoS
351 for each flow, in accordance with the service model. The router
352 function that creates different qualities of service is called
353 "traffic control". Traffic control in turn is implemented by
354 three components: the packet scheduler, the classifier, and
355 admission control.
356
357 o Packet Scheduler
358
359 The packet scheduler manages the forwarding of different
360 packet streams using a set of queues and perhaps other
361 mechanisms like timers. The packet scheduler must be
362 implemented at the point where packets are queued; this is
363 the output driver level of a typical operating system, and
364 corresponds to the link layer protocol. The details of the
365 scheduling algorithm may be specific to the particular output
366 medium. For example, the output driver will need to invoke
367 the appropriate link-layer controls when interfacing to a
368 network technology that has an internal bandwidth allocation
369 mechanism.
370
371 An experimental packet scheduler has been built that
372 implements the IS model described in <a href="#section-3">Section 3</a> and [SCZ93];
373 this is known as the CSZ scheduler and is discussed further
374 in <a href="#section-4">Section 4</a>. We note that the CSZ scheme is not mandatory
375 to accomplish our service model; indeed for parts of the
376 network that are known always to be underloaded, FIFO will
377 deliver satisfactory service.
378
379 There is another component that could be considered part of
380 the packet scheduler or separate: the estimator [<a href="#ref-Jacobson91" title="&quot;Private Communication&quot;">Jacobson91</a>].
381 This algorithm is used to measure properties of the outgoing
382 traffic stream, to develop statistics that control packet
383 scheduling and admission control. This memo will consider
384 the estimator to be a part of the packet scheduler.
385
386
387
388 <span class="grey">Braden, Clark &amp; Shenker [Page 7]</span></pre>
389 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-8" ></span>
390 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
391
392
393 o Classifier
394
395 For the purpose of traffic control (and accounting), each
396 incoming packet must be mapped into some class; all packets
397 in the same class get the same treatment from the packet
398 scheduler. This mapping is performed by the classifier.
399 Choice of a class may be based upon the contents of the
400 existing packet header(s) and/or some additional
401 classification number added to each packet.
402
403 A class might correspond to a broad category of flows, e.g.,
404 all video flows or all flows attributable to a particular
405 organization. On the other hand, a class might hold only a
406 single flow. A class is an abstraction that may be local to
407 a particular router; the same packet may be classified
408 differently by different routers along the path. For
409 example, backbone routers may choose to map many flows into a
410 few aggregated classes, while routers nearer the periphery,
411 where there is much less aggregation, may use a separate
412 class for each flow.
413
414 o Admission Control
415
416 Admission control implements the decision algorithm that a
417 router or host uses to determine whether a new flow can be
418 granted the requested QoS without impacting earlier
419 guarantees. Admission control is invoked at each node to
420 make a local accept/reject decision, at the time a host
421 requests a real-time service along some path through the
422 Internet. The admission control algorithm must be consistent
423 with the service model, and it is logically part of traffic
424 control. Although there are still open research issues in
425 admission control, a first cut exists [<a href="#ref-JCSZ92" title="&quot;An Admission Control Algorithm for Predictive Real-Time Service&quot;">JCSZ92</a>].
426
427 Admission control is sometimes confused with policing or
428 enforcement, which is a packet-by-packet function at the
429 "edge" of the network to ensure that a host does not violate
430 its promised traffic characteristics. We consider policing
431 to be one of the functions of the packet scheduler.
432
433 In addition to ensuring that QoS guarantees are met,
434 admission control will be concerned with enforcing
435 administrative policies on resource reservations. Some
436 policies will demand authentication of those requesting
437 reservations. Finally, admission control will play an
438
439
440
441
442
443
444 <span class="grey">Braden, Clark &amp; Shenker [Page 8]</span></pre>
445 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-9" ></span>
446 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
447
448
449 important role in accounting and administrative reporting.
450
451 The fourth and final component of our implementation framework is
452 a reservation setup protocol, which is necessary to create and
453 maintain flow-specific state in the endpoint hosts and in routers
454 along the path of a flow. Section discusses a reservation setup
455 protocol called RSVP (for "ReSerVation Protocol") [RSVP93a,
456 RSVP93b]. It may not be possible to insist that there be only one
457 reservation protocol in the Internet, but we will argue that
458 multiple choices for reservation protocols will cause confusion.
459 We believe that multiple protocols should exist only if they
460 support different modes of reservation.
461
462 The setup requirements for the link-sharing portion of the service
463 model are far less clear than those for resource reservations.
464 While we expect that much of this can be done through network
465 management interfaces, and thus need not be part of the overall
466 architecture, we may also need RSVP to play a role in providing
467 the required state.
468
469 In order to state its resource requirements, an application must
470 specify the desired QoS using a list of parameters that is called
471 a "flowspec" [<a href="#ref-Partridge92" title="&quot;A Proposed Flow Specification&quot;">Partridge92</a>]. The flowspec is carried by the
472 reservation setup protocol, passed to admission control for to
473 test for acceptability, and ultimately used to parametrize the
474 packet scheduling mechanism.
475
476 Figure shows how these components might fit into an IP router
477 that has been extended to provide integrated services. The router
478 has two broad functional divisions: the forwarding path below the
479 double horizontal line, and the background code above the line.
480
481 The forwarding path of the router is executed for every packet and
482 must therefore be highly optimized. Indeed, in most commercial
483 routers, its implementation involves a hardware assist. The
484 forwarding path is divided into three sections: input driver,
485 internet forwarder, and output driver. The internet forwarder
486 interprets the internetworking protocol header appropriate to the
487 protocol suite, e.g., the IP header for TCP/IP, or the CLNP header
488 for OSI. For each packet, an internet forwarder executes a
489 suite-dependent classifier and then passes the packet and its
490 class to the appropriate output driver. A classifier must be both
491 general and efficient. For efficiency, a common mechanism should
492 be used for both resource classification and route lookup.
493
494 The output driver implements the packet scheduler. (Layerists
495 will observe that the output driver now has two distinct sections:
496 the packet scheduler that is largely independent of the detailed
497
498
499
500 <span class="grey">Braden, Clark &amp; Shenker [Page 9]</span></pre>
501 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-10" ></span>
502 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
503
504
505 mechanics of the interface, and the actual I/O driver that is only
506 concerned with the grittiness of the hardware. The estimator
507 lives somewhere in between. We only note this fact, without
508 suggesting that it be elevated to a principle.).
509
510
511 _____________________________________________________________
512 | ____________ ____________ ___________ |
513 | | | | Reservation| | | |
514 | | Routing | | Setup | | Management| |
515 | | Agent | | Agent | | Agent | |
516 | |______._____| |______._____| |_____._____| |
517 | . . | . |
518 | . . _V________ . |
519 | . . | Admission| . |
520 | . . | Control | . |
521 | V . |__________| . |
522 | [Routing ] V V |
523 | [Database] [Traffic Control Database] |
524 |=============================================================|
525 | | | _______ |
526 | | __________ | |_|_|_|_| =&gt; o |
527 | | | | | Packet | _____ |
528 | ====&gt; |Classifier| =====&gt; Scheduler |===&gt;|_|_|_| ===&gt;
529 | | |__________| | _______ | |
530 | | | |_|_|_|_| =&gt; o |
531 | Input | Internet | |
532 | Driver | Forwarder | O u t p u t D r i v e r |
533 |________|__________________|_________________________________|
534
535 Figure 1: Implementation Reference Model for Routers
536
537
538 The background code is simply loaded into router memory and
539 executed by a general-purpose CPU. These background routines
540 create data structures that control the forwarding path. The
541 routing agent implements a particular routing protocol and builds
542 a routing database. The reservation setup agent implements the
543 protocol used to set up resource reservations; see Section . If
544 admission control gives the "OK" for a new request, the
545 appropriate changes are made to the classifier and packet
546 scheduler database to implement the desired QoS. Finally, every
547 router supports an agent for network management. This agent must
548 be able to modify the classifier and packet scheduler databases to
549 set up controlled link-sharing and to set admission control
550 policies.
551
552
553
554
555
556 <span class="grey">Braden, Clark &amp; Shenker [Page 10]</span></pre>
557 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-11" ></span>
558 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
559
560
561 The implementation framework for a host is generally similar to
562 that for a router, with the addition of applications. Rather than
563 being forwarded, host data originates and terminates in an
564 application. An application needing a real-time QoS for a flow
565 must somehow invoke a local reservation setup agent. The best way
566 to interface to applications is still to be determined. For
567 example, there might be an explicit API for network resource
568 setup, or the setup might be invoked implicitly as part of the
569 operating system scheduling function. The IP output routine of a
570 host may need no classifier, since the class assignment for a
571 packet can be specified in the local I/O control structure
572 corresponding to the flow.
573
574 In routers, integrated service will require changes to both the
575 forwarding path and the background functions. The forwarding
576 path, which may depend upon hardware acceleration for performance,
577 will be the more difficult and costly to change. It will be vital
578 to choose a set of traffic control mechanisms that is general and
579 adaptable to a wide variety of policy requirements and future
580 circumstances, and that can be implemented efficiently.
581
582 <span class="h2"><a class="selflink" id="section-3" href="#section-3">3</a>. Integrated Services Model</span>
583
584 A service model is embedded within the network service interface
585 invoked by applications to define the set of services they can
586 request. While both the underlying network technology and the
587 overlying suite of applications will evolve, the need for
588 compatibility requires that this service interface remain relatively
589 stable (or, more properly, extensible; we do expect to add new
590 services in the future but we also expect that it will be hard to
591 change existing services). Because of its enduring impact, the
592 service model should not be designed in reference to any specific
593 network artifact but rather should be based on fundamental service
594 requirements.
595
596 We now briefly describe a proposal for a core set of services for the
597 Internet; this proposed core service model is more fully described in
598 [<a href="#ref-SCZ93a" title="&quot;A Scheduling Service Model and a Scheduling Architecture for an Integrated Services Packet Network&quot;">SCZ93a</a>, <a href="#ref-SCZ93b" title="&quot;A Service Model for the Integrated Services Internet&quot;">SCZ93b</a>]. This core service model addresses those services
599 which relate most directly to the time-of-delivery of packets. We
600 leave the remaining services (such as routing, security, or stream
601 synchronization) for other standardization venues. A service model
602 consists of a set of service commitments; in response to a service
603 request the network commits to deliver some service. These service
604 commitments can be categorized by the entity to whom they are made:
605 they can be made to either individual flows or to collective entities
606 (classes of flows). The service commitments made to individual flows
607 are intended to provide reasonable application performance, and thus
608 are driven by the ergonomic requirements of the applications; these
609
610
611
612 <span class="grey">Braden, Clark &amp; Shenker [Page 11]</span></pre>
613 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-12" ></span>
614 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
615
616
617 service commitments relate to the quality of service delivered to an
618 individual flow. The service commitments made to collective entities
619 are driven by resource-sharing, or economic, requirements; these
620 service commitments relate to the aggregate resources made available
621 to the various entities.
622
623 In this section we start by exploring the service requirements of
624 individual flows and propose a corresponding set of services. We
625 then discuss the service requirements and services for resource
626 sharing. Finally, we conclude with some remarks about packet
627 dropping.
628
629 3.1 Quality of Service Requirements
630
631 The core service model is concerned almost exclusively with the
632 time-of-delivery of packets. Thus, per-packet delay is the
633 central quantity about which the network makes quality of service
634 commitments. We make the even more restrictive assumption that
635 the only quantity about which we make quantitative service
636 commitments are bounds on the maximum and minimum delays.
637
638 The degree to which application performance depends on low delay
639 service varies widely, and we can make several qualitative
640 distinctions between applications based on the degree of their
641 dependence. One class of applications needs the data in each
642 packet by a certain time and, if the data has not arrived by then,
643 the data is essentially worthless; we call these real-time
644 applications. Another class of applications will always wait for
645 data to arrive; we call these " elastic" applications. We now
646 consider the delay requirements of these two classes separately.
647
648 3.1.1 Real-Time Applications
649
650 An important class of such real-time applications, which are
651 the only real-time applications we explicitly consider in the
652 arguments that follow, are "playback" applications. In a
653 playback application, the source takes some signal, packetizes
654 it, and then transmits the packets over the network. The
655 network inevitably introduces some variation in the delay of
656 the delivered packets. The receiver depacketizes the data and
657 then attempts to faithfully play back the signal. This is done
658 by buffering the incoming data and then replaying the signal at
659 some fixed offset delay from the original departure time; the
660 term "playback point" refers to the point in time which is
661 offset from the original departure time by this fixed delay.
662 Any data that arrives before its associated playback point can
663 be used to reconstruct the signal; data arriving after the
664 playback point is essentially useless in reconstructing the
665
666
667
668 <span class="grey">Braden, Clark &amp; Shenker [Page 12]</span></pre>
669 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-13" ></span>
670 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
671
672
673 real-time signal.
674
675 In order to choose a reasonable value for the offset delay, an
676 application needs some "a priori" characterization of the
677 maximum delay its packets will experience. This "a priori"
678 characterization could either be provided by the network in a
679 quantitative service commitment to a delay bound, or through
680 the observation of the delays experienced by the previously
681 arrived packets; the application needs to know what delays to
682 expect, but this expectation need not be constant for the
683 entire duration of the flow.
684
685 The performance of a playback application is measured along two
686 dimensions: latency and fidelity. Some playback applications,
687 in particular those that involve interaction between the two
688 ends of a connection such as a phone call, are rather sensitive
689 to the latency; other playback applications, such as
690 transmitting a movie or lecture, are not. Similarly,
691 applications exhibit a wide range of sensitivity to loss of
692 fidelity. We will consider two somewhat artificially
693 dichotomous classes: intolerant applications, which require an
694 absolutely faithful playback, and tolerant applications, which
695 can tolerate some loss of fidelity. We expect that the vast
696 bulk of audio and video applications will be tolerant, but we
697 also suspect that there will be other applications, such as
698 circuit emulation, that are intolerant.
699
700 Delay can affect the performance of playback applications in
701 two ways. First, the value of the offset delay, which is
702 determined by predictions about the future packet delays,
703 determines the latency of the application. Second, the delays
704 of individual packets can decrease the fidelity of the playback
705 by exceeding the offset delay; the application then can either
706 change the offset delay in order to play back late packets
707 (which introduces distortion) or merely discard late packets
708 (which creates an incomplete signal). The two different ways
709 of coping with late packets offer a choice between an
710 incomplete signal and a distorted one, and the optimal choice
711 will depend on the details of the application, but the
712 important point is that late packets necessarily decrease
713 fidelity.
714
715 Intolerant applications must use a fixed offset delay, since
716 any variation in the offset delay will introduce some
717 distortion in the playback. For a given distribution of packet
718 delays, this fixed offset delay must be larger than the
719 absolute maximum delay, to avoid the possibility of late
720 packets. Such an application can only set its offset delay
721
722
723
724 <span class="grey">Braden, Clark &amp; Shenker [Page 13]</span></pre>
725 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-14" ></span>
726 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
727
728
729 appropriately if it is given a perfectly reliable upper bound
730 on the maximum delay of each packet. We call a service
731 characterized by a perfectly reliable upper bound on delay "
732 guaranteed service", and propose this as the appropriate
733 service model for intolerant playback applications.
734
735 In contrast, tolerant applications need not set their offset
736 delay greater than the absolute maximum delay, since they can
737 tolerate some late packets. Moreover, instead of using a
738 single fixed value for the offset delay, they can attempt to
739 reduce their latency by varying their offset delays in response
740 to the actual packet delays experienced in the recent past. We
741 call applications which vary their offset delays in this manner
742 "adaptive" playback applications.
743
744 For tolerant applications we propose a service model called "
745 predictive service" which supplies a fairly reliable, but not
746 perfectly reliable, delay bound. This bound, in contrast to
747 the bound in the guaranteed service, is not based on worst case
748 assumptions on the behavior of other flows. Instead, this
749 bound might be computed with properly conservative predictions
750 about the behavior of other flows. If the network turns out to
751 be wrong and the bound is violated, the application's
752 performance will perhaps suffer, but the users are willing to
753 tolerate such interruptions in service in return for the
754 presumed lower cost of the service. Furthermore, because many
755 of the tolerant applications are adaptive, we augment the
756 predictive service to also give "minimax" service, which is to
757 attempt to minimize the ex post maximum delay. This service is
758 not trying to minimize the delay of every packet, but rather is
759 trying to pull in the tail of the delay distribution.
760
761 It is clear that given a choice, with all other things being
762 equal, an application would perform no worse with absolutely
763 reliable bounds than with fairly reliable bounds. Why, then,
764 do we offer predictive service? The key consideration here is
765 efficiency; when one relaxes the service requirements from
766 perfectly to fairly reliable bounds, this increases the level
767 of network utilization that can be sustained, and thus the
768 price of the predictive service will presumably be lower than
769 that of guaranteed service. The predictive service class is
770 motivated by the conjecture that the performance penalty will
771 be small for tolerant applications but the overall efficiency
772 gain will be quite large.
773
774 In order to provide a delay bound, the nature of the traffic
775 from the source must be characterized, and there must be some
776 admission control algorithm which insures that a requested flow
777
778
779
780 <span class="grey">Braden, Clark &amp; Shenker [Page 14]</span></pre>
781 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-15" ></span>
782 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
783
784
785 can actually be accommodated. A fundamental point of our
786 overall architecture is that traffic characterization and
787 admission control are necessary for these real-time delay bound
788 services. So far we have assumed that an application's data
789 generation process is an intrinsic property unaffected by the
790 network. However, there are likely to be many audio and video
791 applications which can adjust their coding scheme and thus can
792 alter the resulting data generation process depending on the
793 network service available. This alteration of the coding
794 scheme will present a tradeoff between fidelity (of the coding
795 scheme itself, not of the playback process) and the bandwidth
796 requirements of the flow. Such "rate-adaptive" playback
797 applications have the advantage that they can adjust to the
798 current network conditions not just by resetting their playback
799 point but also by adjusting the traffic pattern itself. For
800 rate-adaptive applications, the traffic characterizations used
801 in the service commitment are not immutable. We can thus
802 augment the service model by allowing the network to notify
803 (either implicitly through packet drops or explicitly through
804 control packets) rate-adaptive applications to change their
805 traffic characterization.
806
807 3.1.2 Elastic Applications
808
809 While real-time applications do not wait for late data to
810 arrive, elastic applications will always wait for data to
811 arrive. It is not that these applications are insensitive to
812 delay; to the contrary, significantly increasing the delay of a
813 packet will often harm the application's performance. Rather,
814 the key point is that the application typically uses the
815 arriving data immediately, rather than buffering it for some
816 later time, and will always choose to wait for the incoming
817 data rather than proceed without it. Because arriving data can
818 be used immediately, these applications do not require any a
819 priori characterization of the service in order for the
820 application to function. Generally speaking, it is likely that
821 for a given distribution of packet delays, the perceived
822 performance of elastic applications will depend more on the
823 average delay than on the tail of the delay distribution. One
824 can think of several categories of such elastic applications:
825 interactive burst (Telnet, X, NFS), interactive bulk transfer
826 (FTP), and asynchronous bulk transfer (electronic mail, FAX).
827 The delay requirements of these elastic applications vary from
828 rather demanding for interactive burst applications to rather
829 lax for asynchronous bulk transfer, with interactive bulk
830 transfer being intermediate between them.
831
832
833
834
835
836 <span class="grey">Braden, Clark &amp; Shenker [Page 15]</span></pre>
837 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-16" ></span>
838 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
839
840
841 An appropriate service model for elastic applications is to
842 provide "as-soon-as-possible", or ASAP service. (For
843 compatibility with historical usage, we will use the term
844 best-effort service when referring to ASAP service.). We
845 furthermore propose to offer several classes of best-effort
846 service to reflect the relative delay sensitivities of
847 different elastic applications. This service model allows
848 interactive burst applications to have lower delays than
849 interactive bulk applications, which in turn would have lower
850 delays than asynchronous bulk applications. In contrast to the
851 real-time service models, applications using this service are
852 not subject to admission control.
853
854 The taxonomy of applications into tolerant playback, intolerant
855 playback, and elastic is neither exact nor complete, but was
856 only used to guide the development of the core service model.
857 The resulting core service model should be judged not on the
858 validity of the underlying taxonomy but rather on its ability
859 to adequately meet the needs of the entire spectrum of
860 applications. In particular, not all real-time applications
861 are playback applications; for example, one might imagine a
862 visualization application which merely displayed the image
863 encoded in each packet whenever it arrived. However, non-
864 playback applications can still use either the guaranteed or
865 predictive real-time service model, although these services are
866 not specifically tailored to their needs. Similarly, playback
867 applications cannot be neatly classified as either tolerant or
868 intolerant, but rather fall along a continuum; offering both
869 guaranteed and predictive service allows applications to make
870 their own tradeoff between fidelity, latency, and cost.
871 Despite these obvious deficiencies in the taxonomy, we expect
872 that it describes the service requirements of current and
873 future applications well enough so that our core service model
874 can adequately meet all application needs.
875
876 3.2 Resource-Sharing Requirements and Service Models
877
878 The last section considered quality of service commitments; these
879 commitments dictate how the network must allocate its resources
880 among the individual flows. This allocation of resources is
881 typically negotiated on a flow-by-flow basis as each flow requests
882 admission to the network, and does not address any of the policy
883 issues that arise when one looks at collections of flows. To
884 address these collective policy issues, we now discuss resource-
885 sharing service commitments. Recall that for individual quality
886 of service commitments we focused on delay as the only quantity of
887 interest. Here, we postulate that the quantity of primary
888 interest in resource-sharing is aggregate bandwidth on individual
889
890
891
892 <span class="grey">Braden, Clark &amp; Shenker [Page 16]</span></pre>
893 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-17" ></span>
894 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
895
896
897 links. Thus, this component of the service model, called "link-
898 sharing", addresses the question of how to share the aggregate
899 bandwidth of a link among various collective entities according to
900 some set of specified shares. There are several examples that are
901 commonly used to explain the requirement of link-sharing among
902 collective entities.
903
904 Multi-entity link-sharing. -- A link may be purchased and used
905 jointly by several organizations, government agencies or the like.
906 They may wish to insure that under overload the link is shared in
907 a controlled way, perhaps in proportion to the capital investment
908 of each entity. At the same time, they might wish that when the
909 link is underloaded, any one of the entities could utilize all the
910 idle bandwidth.
911
912 Multi-protocol link-sharing -- In a multi-protocol Internet, it
913 may be desired to prevent one protocol family (DECnet, IP, IPX,
914 OSI, SNA, etc.) from overloading the link and excluding the other
915 families. This is important because different families may have
916 different methods of detecting and responding to congestion, and
917 some methods may be more "aggressive" than others. This could lead
918 to a situation in which one protocol backs off more rapidly than
919 another under congestion, and ends up getting no bandwidth.
920 Explicit control in the router may be required to correct this.
921 Again, one might expect that this control should apply only under
922 overload, while permitting an idle link to be used in any
923 proportion.
924
925 Multi-service sharing -- Within a protocol family such as IP, an
926 administrator might wish to limit the fraction of bandwidth
927 allocated to various service classes. For example, an
928 administrator might wish to limit the amount of real-time traffic
929 to some fraction of the link, to avoid preempting elastic traffic
930 such as FTP.
931
932 In general terms, the link-sharing service model is to share the
933 aggregate bandwidth according to some specified shares. We can
934 extend this link-sharing service model to a hierarchical version.
935 For instance, a link could be divided between a number of
936 organizations, each of which would divide the resulting allocation
937 among a number of protocols, each of which would be divided among
938 a number of services. Here, the sharing is defined by a tree with
939 shares assigned to each leaf node.
940
941 An idealized fluid model of instantaneous link-sharing with
942 proportional sharing of excess is the fluid processor sharing
943 model (introduced in [<a href="#ref-DKS89" title="&quot;Analysis and Simulation of a Fair Queueing Algorithm&quot;">DKS89</a>] and further explored in [<a href="#ref-Parekh92" title="&quot;A Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks&quot;">Parekh92</a>]
944 and generalized to the hierarchical case) where at every instant
945
946
947
948 <span class="grey">Braden, Clark &amp; Shenker [Page 17]</span></pre>
949 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-18" ></span>
950 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
951
952
953 the available bandwidth is shared between the active entities
954 (i.e., those having packets in the queue) in proportion to the
955 assigned shares of the resource. This fluid model exhibits the
956 desired policy behavior but is, of course, an unrealistic
957 idealization. We then propose that the actual service model
958 should be to approximate, as closely as possible, the bandwidth
959 shares produced by this ideal fluid model. It is not necessary to
960 require that the specific order of packet departures match those
961 of the fluid model since we presume that all detailed per-packet
962 delay requirements of individual flows are addressed through
963 quality of service commitments and, furthermore, the satisfaction
964 with the link-sharing service delivered will probably not depend
965 very sensitively on small deviations from the scheduling implied
966 by the fluid link-sharing model.
967
968 We previously observed that admission control was necessary to
969 ensure that the real-time service commitments could be met.
970 Similarly, admission control will again be necessary to ensure
971 that the link-sharing commitments can be met. For each entity,
972 admission control must keep the cumulative guaranteed and
973 predictive traffic from exceeding the assigned link-share.
974
975 3.3 Packet Dropping
976
977 So far, we have implicitly assumed that all packets within a flow
978 were equally important. However, in many audio and video streams,
979 some packets are more valuable than others. We therefore propose
980 augmenting the service model with a "preemptable" packet service,
981 whereby some of the packets within a flow could be marked as
982 preemptable. When the network was in danger of not meeting some
983 of its quantitative service commitments, it could exercise a
984 certain packet's "preemptability option" and discard the packet
985 (not merely delay it, since that would introduce out-of-order
986 problems). By discarding these preemptable packets, a router can
987 reduce the delays of the not-preempted packets.
988
989 Furthermore, one can define a class of packets that is not subject
990 to admission control. In the scenario described above where
991 preemptable packets are dropped only when quantitative service
992 commitments are in danger of being violated, the expectation is
993 that preemptable packets will almost always be delivered and thus
994 they must included in the traffic description used in admission
995 control. However, we can extend preemptability to the extreme
996 case of "expendable" packets (the term expendable is used to
997 connote an extreme degree of preemptability), where the
998 expectation is that many of these expendable packets may not be
999 delivered. One can then exclude expendable packets from the
1000 traffic description used in admission control; i.e., the packets
1001
1002
1003
1004 <span class="grey">Braden, Clark &amp; Shenker [Page 18]</span></pre>
1005 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-19" ></span>
1006 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1007
1008
1009 are not considered part of the flow from the perspective of
1010 admission control, since there is no commitment that they will be
1011 delivered.
1012
1013 3.4 Usage Feedback
1014
1015 Another important issue in the service is the model for usage
1016 feedback, also known as "accounting", to prevent abuse of network
1017 resources. The link-sharing service described earlier can be
1018 used to provide administratively-imposed limits on usage.
1019 However, a more free-market model of network access will require
1020 back-pressure on users for the network resources they reserve.
1021 This is a highly contentious issue, and we are not prepared to say
1022 more about it at this time.
1023
1024 3.5 Reservation Model
1025
1026 The "reservation model" describes how an application negotiates
1027 for a QoS level. The simplest model is that the application asks
1028 for a particular QoS and the network either grants it or refuses.
1029 Often the situation will be more complex. Many applications will
1030 be able to get acceptable service from a range of QoS levels, or
1031 more generally, from anywhere within some region of the multi-
1032 dimensional space of a flowspec.
1033
1034 For example, rather than simply refusing the request, the network
1035 might grant a lower resource level and inform the application of
1036 what QoS has been actually granted. A more complex example is the
1037 "two-pass" reservation model, In this scheme, an "offered"
1038 flowspec is propagated along the multicast distribution tree from
1039 each sender Si to all receivers Rj. Each router along the path
1040 records these values and perhaps adjusts them to reflect available
1041 capacity. The receivers get these offers, generate corresponding
1042 "requested" flowspecs, and propagate them back along the same
1043 routes to the senders. At each node, a local reconciliation must
1044 be performed between the offered and the requested flowspec to
1045 create a reservation, and an appropriately modified requested
1046 flowspec is passed on. This two-pass scheme allows extensive
1047 properties like allowed delay to be distributed across hops in the
1048 path [<a href="#ref-Tenet90" title="&quot;A Scheme for Real-Time Channel Establishment in Wide-Area Networks&quot;">Tenet90</a>, <a href="#ref-ST2-90" title="&quot;Experimental Internet Stream Protocol: Version 2 (ST-II)&quot;">ST2-90</a>]. Further work is needed to define the
1049 amount of generality, with a corresponding level of complexity,
1050 that is required in the reservation model.
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060 <span class="grey">Braden, Clark &amp; Shenker [Page 19]</span></pre>
1061 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-20" ></span>
1062 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1063
1064
1065 <span class="h2"><a class="selflink" id="section-4" href="#section-4">4</a>. Traffic Control Mechanisms</span>
1066
1067 We first survey very briefly the possible traffic control mechanisms.
1068 Then in <a href="#section-4.2">Section 4.2</a> we apply a subset of these mechanisms to support
1069 the various services that we have proposed.
1070
1071 4.1 Basic Functions
1072
1073 In the packet forwarding path, there is actually a very limited
1074 set of actions that a router can take. Given a particular packet,
1075 a router must select a route for it; in addition the router can
1076 either forward it or drop it, and the router may reorder it with
1077 respect to other packets waiting to depart. The router can also
1078 hold the packet, even though the link is idle. These are the
1079 building blocks from which we must fashion the desired behavior.
1080
1081 4.1.1 Packet Scheduling
1082
1083 The basic function of packet scheduling is to reorder the
1084 output queue. There are many papers that have been written on
1085 possible ways to manage the output queue, and the resulting
1086 behavior. Perhaps the simplest approach is a priority scheme,
1087 in which packets are ordered by priority, and highest priority
1088 packets always leave first. This has the effect of giving some
1089 packets absolute preference over others; if there are enough of
1090 the higher priority packets, the lower priority class can be
1091 completely prevented from being sent.
1092
1093 An alternative scheduling scheme is round-robin or some
1094 variant, which gives different classes of packets access to a
1095 share of the link. A variant called Weighted Fair Queueing, or
1096 WFQ, has been demonstrated to allocate the total bandwidth of a
1097 link into specified shares.
1098
1099 There are more complex schemes for queue management, most of
1100 which involve observing the service objectives of individual
1101 packets, such as delivery deadline, and ordering packets based
1102 on these criteria.
1103
1104 4.1.2 Packet Dropping
1105
1106 The controlled dropping of packets is as important as their
1107 scheduling.
1108
1109 Most obviously, a router must drop packets when its buffers are
1110 all full. This fact, however, does not determine which packet
1111 should be dropped. Dropping the arriving packet, while simple,
1112 may cause undesired behavior.
1113
1114
1115
1116 <span class="grey">Braden, Clark &amp; Shenker [Page 20]</span></pre>
1117 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-21" ></span>
1118 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1119
1120
1121 In the context of today's Internet, with TCP operating over
1122 best effort IP service, dropping a packet is taken by TCP as a
1123 signal of congestion and causes it to reduce its load on the
1124 network. Thus, picking a packet to drop is the same as picking
1125 a source to throttle. Without going into any particular
1126 algorithm, this simple relation suggests that some specific
1127 dropping controls should be implemented in routers to improve
1128 congestion control.
1129
1130 In the context of real-time services, dropping more directly
1131 relates to achieving the desired quality of service. If a
1132 queue builds up, dropping one packet reduces the delay of all
1133 the packets behind it in the queue. The loss of one can
1134 contribute to the success of many. The problem for the
1135 implementor is to determine when the service objective (the
1136 delay bound) is in danger of being violated. One cannot look
1137 at queue length as an indication of how long packets have sat
1138 in a queue. If there is a priority scheme in place, packets of
1139 lower priority can be pre-empted indefinitely, so even a short
1140 queue may have very old packets in it. While actual time
1141 stamps could be used to measure holding time, the complexity
1142 may be unacceptable.
1143
1144 Some simple dropping schemes, such as combining all the buffers
1145 in a single global pool, and dropping the arriving packet if
1146 the pool is full, can defeat the service objective of a WFQ
1147 scheduling scheme. Thus, dropping and scheduling must be
1148 coordinated.
1149
1150 4.1.3 Packet Classification
1151
1152 The above discussion of scheduling and dropping presumed that
1153 the packet had been classified into some flow or sequence of
1154 packets that should be treated in a specified way. A
1155 preliminary to this sort of processing is the classification
1156 itself. Today a router looks at the destination address and
1157 selects a route. The destination address is not sufficient to
1158 select the class of service a packet must receive; more
1159 information is needed.
1160
1161 One approach would be to abandon the IP datagram model for a
1162 virtual circuit model, in which a circuit is set up with
1163 specific service attributes, and the packet carries a circuit
1164 identifier. This is the approach of ATM as well as protocols
1165 such as ST-II [<a href="#ref-ST2-90" title="&quot;Experimental Internet Stream Protocol: Version 2 (ST-II)&quot;">ST2-90</a>]. Another model, less hostile to IP, is
1166 to allow the classifier to look at more fields in the packet,
1167 such as the source address, the protocol number and the port
1168 fields. Thus, video streams might be recognized by a
1169
1170
1171
1172 <span class="grey">Braden, Clark &amp; Shenker [Page 21]</span></pre>
1173 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-22" ></span>
1174 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1175
1176
1177 particular well-known port field in the UDP header, or a
1178 particular flow might be recognized by looking at both the
1179 source and destination port numbers. It would be possible to
1180 look even deeper into the packets, for example testing a field
1181 in the application layer to select a subset of a
1182 hierarchically-encoded video stream.
1183
1184 The classifier implementation issues are complexity and
1185 processing overhead. Current experience suggests that careful
1186 implementation of efficient algorithms can lead to efficient
1187 classification of IP packets. This result is very important,
1188 since it allows us to add QoS support to existing applications,
1189 such as Telnet, which are based on existing IP headers.
1190
1191 One approach to reducing the overhead of classification would
1192 be to provide a "flow-id" field in the Internet-layer packet
1193 header. This flow-id would be a handle that could be cached
1194 and used to short-cut classification of the packet. There are
1195 a number of variations of this concept, and engineering is
1196 required to choose the best design.
1197
1198 4.1.4 Admission Control
1199
1200 As we stated in the introduction, real-time service depends on
1201 setting up state in the router and making commitments to
1202 certain classes of packets. In order to insure that these
1203 commitments can be met, it is necessary that resources be
1204 explicitly requested, so that the request can be refused if the
1205 resources are not available. The decision about resource
1206 availability is called admission control.
1207
1208 Admission control requires that the router understand the
1209 demands that are currently being made on its assets. The
1210 approach traditionally proposed is to remember the service
1211 parameters of past requests, and make a computation based on
1212 the worst-case bounds on each service. A recent proposal,
1213 which is likely to provide better link utilization, is to
1214 program the router to measure the actual usage by existing
1215 packet flows, and to use this measured information as a basis
1216 of admitting new flows [<a href="#ref-JCSZ92" title="&quot;An Admission Control Algorithm for Predictive Real-Time Service&quot;">JCSZ92</a>]. This approach is subject to
1217 higher risk of overload, but may prove much more effective in
1218 using bandwidth.
1219
1220 Note that while the need for admission control is part of the
1221 global service model, the details of the algorithm run in each
1222 router is a local matter. Thus, vendors can compete by
1223 developing and marketing better admission control algorithms,
1224 which lead to higher link loadings with fewer service
1225
1226
1227
1228 <span class="grey">Braden, Clark &amp; Shenker [Page 22]</span></pre>
1229 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-23" ></span>
1230 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1231
1232
1233 overloads.
1234
1235 4.2 Applying the Mechanisms
1236
1237 The various tools described above can be combined to support the
1238 services which were discussed in <a href="#section-3">section 3</a>.
1239
1240 o Guaranteed delay bounds
1241
1242 A theoretical result by Parekh [<a href="#ref-Parekh92" title="&quot;A Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks&quot;">Parekh92</a>] shows that if the
1243 router implements a WFQ scheduling discipline, and if the
1244 nature of the traffic source can be characterized (e.g. if it
1245 fits within some bound such as a token bucket) then there
1246 will be an absolute upper bound on the network delay of the
1247 traffic in question. This simple and very powerful result
1248 applies not just to one switch, but to general networks of
1249 routers. The result is a constructive one; that is, Parekh
1250 displays a source behavior which leads to the bound, and then
1251 shows that this behavior is the worst possible. This means
1252 that the bound he computes is the best there can be, under
1253 these assumptions.
1254
1255 o Link sharing
1256
1257 The same WFQ scheme can provide controlled link sharing. The
1258 service objective here is not to bound delay, but to limit
1259 overload shares on a link, while allowing any mix of traffic
1260 to proceed if there is spare capacity. This use of WFQ is
1261 available in commercial routers today, and is used to
1262 segregate traffic into classes based on such things as
1263 protocol type or application. For example, one can allocate
1264 separate shares to TCP, IPX and SNA, and one can assure that
1265 network control traffic gets a guaranteed share of the link.
1266
1267 o Predictive real-time service
1268
1269 This service is actually more subtle than guaranteed service.
1270 Its objective is to give a delay bound which is, on the one
1271 hand, as low as possible, and on the other hand, stable
1272 enough that the receiver can estimate it. The WFQ mechanism
1273 leads to a guaranteed bound, but not necessarily a low bound.
1274 In fact, mixing traffic into one queue, rather than
1275 separating it as in WFQ, leads to lower bounds, so long as
1276 the mixed traffic is generally similar (e.g., mixing traffic
1277 from multiple video coders makes sense, mixing video and FTP
1278 does not).
1279
1280
1281
1282
1283
1284 <span class="grey">Braden, Clark &amp; Shenker [Page 23]</span></pre>
1285 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-24" ></span>
1286 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1287
1288
1289 This suggests that we need a two-tier mechanism, in which the
1290 first tier separates traffic which has different service
1291 objectives, and the second tier schedules traffic within each
1292 first tier class in order to meet its service objective.
1293
1294 4.3 An example: The CSZ scheme
1295
1296 As a proof of concept, a code package has been implemented which
1297 realizes the services discussed above. It actually uses a number
1298 of the basic tools, combined in a way specific to the service
1299 needs. We describe in general terms how it works, to suggest how
1300 services can be realized. We stress that there are other ways of
1301 building a router to meet the same service needs, and there are in
1302 fact other implementations being used today.
1303
1304
1305 At the top level, the CSZ code uses WFQ as an isolation mechanism
1306 to separate guaranteed flows from each other, as well as from the
1307 rest of the traffic. Guaranteed service gets the highest priority
1308 when and only when it needs the access to meets its deadline. WFQ
1309 provides a separate guarantee for each and every guaranteed flow.
1310
1311 Predictive service and best effort service are separated by
1312 priority. Within the predictive service class, a further priority
1313 is used to provide sub-classes with different delay bounds.
1314 Inside each predictive sub-class, simple FIFO queueing is used to
1315 mix the traffic, which seems to produce good overall delay
1316 behavior. This works because the top-tier algorithm has separated
1317 out the best effort traffic such as FTP.
1318
1319 Within the best-effort class, WFQ is used to provide link sharing.
1320 Since there is a possible requirement for nested shares, this WFQ
1321 code can be used recursively. There are thus two different uses
1322 of WFQ in this code, one to segregate the guaranteed classes, and
1323 one to segregate the link shares. They are similar, but differ in
1324 detail.
1325
1326 Within each link share of the best effort class, priority is used
1327 to permit more time-sensitive elastic traffic to precede other
1328 elastic traffic, e.g., to allow interactive traffic to precede
1329 asynchronous bulk transfers.
1330
1331 The CSZ code thus uses both WFQ and priority in an alternating
1332 manner to build a mechanism to support a range of rather
1333 sophisticated service offerings. This discussion is very brief,
1334 and does not touch on a number of significant issues, such as how
1335 the CSZ code fits real time traffic into the link sharing
1336 objectives. But the basic building blocks are very simple, and
1337
1338
1339
1340 <span class="grey">Braden, Clark &amp; Shenker [Page 24]</span></pre>
1341 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-25" ></span>
1342 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1343
1344
1345 very powerful. In particular, while priority has been proposed as
1346 a key to real-time services, WFQ may be the more general and
1347 powerful of the two schemes. It, rather than priority, supports
1348 guaranteed service and link sharing.
1349
1350 <span class="h2"><a class="selflink" id="section-5" href="#section-5">5</a>. Reservation Setup Protocol</span>
1351
1352 There are a number of requirements to be met by the design of a
1353 reservation setuop protocol. It should be fundamentally designed for
1354 a multicast environment, and it must accommodate heterogeneous
1355 service needs. It must give flexible control over the manner in
1356 which reservations can be shared along branches of the multicast
1357 delivery trees. It should be designed around the elementary action
1358 of adding one sender and/or receiver to an existing set, or deleting
1359 one. It must be robust and scale well to large multicast groups.
1360 Finally, it must provide for advance reservation of resources, and
1361 for the preemption that this implies. The reservation setup protocol
1362 RSVP has been designed to meet these requirements [<a href="#ref-RSVP93a" title="&quot;RSVP: A New Resource ReSerVation Protocol&quot;">RSVP93a</a>, <a href="#ref-RSVP93b" title="&quot;Resource ReSerVation Protocol (RSVP) - Version 1 Functional Specification&quot;">RSVP93b</a>].
1363 This section gives an overview of the design of RSVP.
1364
1365 5.1 RSVP Overview
1366
1367 Figure shows multi-source, multi-destination data delivery for a
1368 particular shared, distributed application. The arrows indicate
1369 data flow from senders S1 and S2 to receivers R1, R2, and R3, and
1370 the cloud represents the distribution mesh created by the
1371 multicast routing protocol. Multicasting distribution replicates
1372 each data packet from a sender Si, for delivery to every receiver
1373 Rj. We treat uncast delivery from S1 to R1 as a special case, and
1374 we call this multicast distribution mesh a session. A session is
1375 defined by the common IP (multicast) destination address of the
1376 receiver(s).
1377
1378
1379 Senders Receivers
1380 _____________________
1381 ( ) ===&gt; R1
1382 S1 ===&gt; ( Multicast )
1383 ( ) ===&gt; R2
1384 ( distribution )
1385 S2 ===&gt; ( )
1386 ( ) ===&gt; R3
1387 (_____________________)
1388
1389 Figure 2: Multicast Distribution Session
1390
1391
1392
1393
1394
1395
1396 <span class="grey">Braden, Clark &amp; Shenker [Page 25]</span></pre>
1397 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-26" ></span>
1398 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1399
1400
1401 5.1.1 Flowspecs and Filter Specs
1402
1403 In general, an RSVP reservation request specifies the amount of
1404 resources to be reserved for all, or some subset of, the
1405 packets in a particular session. The resource quantity is
1406 specified by a flowspec, while the packet subset to receive
1407 those resources is specified by a filter spec. Assuming
1408 admission control succeeds, the flowspec will be used to
1409 parametrize a resource class in the packet scheduler, and the
1410 filter spec will be instantiated in the packet classifier to
1411 map the appropriate packets into this class. The subset of the
1412 classifier state that selects a particular class is referred to
1413 in RSVP documentation as a (packet) "filter".
1414
1415 The RSVP protocol mechanisms provide a very general facility
1416 for creating and maintaining distributed reservation state
1417 across the mesh of multicast delivery paths. These mechanisms
1418 treat flowspecs and filter specs as mostly opaque binary data,
1419 handing them to the local traffic control machinery for
1420 interpretation. Of course, the service model presented to an
1421 application must specify how to encode flowspecs and filter
1422 specs.
1423
1424 5.1.2 Reservation Styles
1425
1426 RSVP offers several different reservation "styles", which
1427 determine the manner in which the resource requirements of
1428 multiple receivers are aggregated in the routers. These styles
1429 allow the reserved resources to more efficiently meet
1430 application requirements. Currently there are three
1431 reservation styles, "wildcard", "fixed-filter", and " dynamic-
1432 filter". A wildcard reservation uses a filter spec that is not
1433 source-specific, so all packets destined for the associated
1434 destination (session) may use a common pool of reserved
1435 resources. This allows a single resource allocation to be made
1436 across all distribution paths for the group. The wildcard
1437 reservation style is useful in support of an audio conference,
1438 where at most a small number of sources are active
1439 simultaneously and may share the resource allocation.
1440
1441 The other two styles use filter specs that select particular
1442 sources. A receiver may desire to receive from a fixed set of
1443 sources, or instead it may desire the network to switch between
1444 different source, by changing its filter spec(s) dymamically.
1445 A fixed-filter style reservation cannot be changed during its
1446 lifetime without re-invoking admission control. Dynamic-filter
1447 reservations do allow a receiver to modify its choice of
1448 source(s) over time without additional admission control;
1449
1450
1451
1452 <span class="grey">Braden, Clark &amp; Shenker [Page 26]</span></pre>
1453 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-27" ></span>
1454 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1455
1456
1457 however, this requires that sufficient resources be allocated
1458 to handle the worst case when all downstream receivers take
1459 input from different sources.
1460
1461 5.1.3 Receiver Initiation
1462
1463 An important design question is whether senders or receivers
1464 should have responsibility for initiating reservations. A
1465 sender knows the qualities of the traffic stream it can send,
1466 while a receiver knows what it wants to (or can) receive.
1467 Perhaps the most obvious choice is to let the sender initiate
1468 the reservation. However, this scales poorly for large,
1469 dynamic multicast delivery trees and for heterogeneous
1470 receivers.
1471
1472 Both of these scaling problems are solved by making the
1473 receiver responsible for initiating a reservation. Receiver
1474 initiation handles heterogeneous receivers easily; each
1475 receiver simply asks for a reservation appropriate to itself,
1476 and any differences among reservations from different receivers
1477 are resolved ("merged") within the network by RSVP. Receiver
1478 initiation is also consisent with IP multicast, in which a
1479 multicast group is created implicitly by receivers joining it.
1480
1481 Although receiver-initiated reservation is the natural choice
1482 for multicast sessions, the justification for receiver
1483 initiateion may appear weaker for unicast sessions, where the
1484 sender may be the logical session initiator. However, we
1485 expect that every realtime application will have its higher-
1486 level signalling and control protocol, and this protocol can be
1487 used to signal the receiver to initiate a reservation (and
1488 perhaps indicate the flowspec to be used). For simplicity and
1489 economy, a setup protocol should support only one direction of
1490 initiation, and, and receiver initiation appears to us to be
1491 the clear winner.
1492
1493 RSVP uses receiver-initiation of rservations [<a href="#ref-RSVP93b" title="&quot;Resource ReSerVation Protocol (RSVP) - Version 1 Functional Specification&quot;">RSVP93b</a>]. A
1494 receiver is assumed to learn the senders' offered flowspecs by
1495 a higher-level mechanism ("out of band"), it then generates its
1496 own desired flowspec and propagates it towards the senders,
1497 making reservations in each router along the way.
1498
1499 5.1.4 Soft State
1500
1501 There are two different possible styles for reservation setup
1502 protocols, the "hard state" (HS) approach (also called
1503 "connection-oriented"), and the "soft state" (SS) approach
1504 (also called "connectionless"). In both approaches, multicast
1505
1506
1507
1508 <span class="grey">Braden, Clark &amp; Shenker [Page 27]</span></pre>
1509 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-28" ></span>
1510 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1511
1512
1513 distribution is performed using flow-specific state in each
1514 router along the path. Under the HS approach, this state is
1515 created and deleted in a fully deterministic manner by
1516 cooperation among the routers. Once a host requests a session,
1517 the "network" takes responsibility for creating and later
1518 destroying the necessary state. ST-II is an example of the HS
1519 approach [<a href="#ref-ST2-90" title="&quot;Experimental Internet Stream Protocol: Version 2 (ST-II)&quot;">ST2-90</a>]. Since management of HS session state is
1520 completely deterministic, the HS setup protocol must be
1521 reliable, with acknowledgments and retransmissions. In order
1522 to achieve deterministic cleanup of state after a failure,
1523 there must be some mechanism to detect failures, i.e., an
1524 "up/down" protocol. The router upstream (towards the source)
1525 from a failure takes responsibility for rebuilding the
1526 necessary state on the router(s) along an alternate route.
1527
1528 RSVP takes the SS approach, which regards the reservation state
1529 as cached information that is installed and periodically
1530 refreshed by the end hosts. Unused state is timed out by the
1531 routers. If the route changes, the refresh messages
1532 automatically install the necessary state along the new route.
1533 The SS approach was chosen to obtain the simplicity and
1534 robustness that have been demonstrated by connectionless
1535 protocols such as IP [<a href="#ref-Clark88" title="&quot;The Design Philosophy of the DARPA Internet Protocols&quot;">Clark88</a>].
1536
1537 5.2 Routing and Reservations
1538
1539 There is a fundamental interaction between resource reservation
1540 set up and routing, since reservation requires the installation of
1541 flow state along the route of data packets. If and when a route
1542 changes, there must be some mechanism to set up a reservation
1543 along the new route.
1544
1545 Some have suggested that reservation setup necessarily requires
1546 route set up, i.e., the imposition of a virtual-circuit internet
1547 layer. However, our goal is to simply extend the Internet
1548 architecture, not replace it. The fundamental connectionless
1549 internet layer [<a href="#ref-Clark88" title="&quot;The Design Philosophy of the DARPA Internet Protocols&quot;">Clark88</a>] has been highly successful, and we wish
1550 to retain it as an architectural foundation. We propose instead
1551 to modify somewhat the pure datagram forwarding mechanism of the
1552 present Internet to accomodate "IS".
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564 <span class="grey">Braden, Clark &amp; Shenker [Page 28]</span></pre>
1565 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-29" ></span>
1566 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1567
1568
1569 There are four routing issues faced by a reservation setup
1570 protocol such as RSVP.
1571
1572 1. Find a route that supports resource reservation.
1573
1574 This is simply "type-of-service" routing, a facility that is
1575 already available in some modern routing protocols.
1576
1577 2. Find a route that has sufficient unreserved capacity for a
1578 new flow.
1579
1580 Early experiments on the ARPANET showed that it is difficult
1581 to do load-dependent dynamic routing on a packet-by-packet
1582 basis without instability problems. However, instability
1583 should not be a problem if load-dependent routing is
1584 performed only at reservation setup time.
1585
1586 Two different approaches might be taken to finding a route
1587 with enough capacity. One could modify the routing
1588 protocol(s) and interface them to the traffic control
1589 mechanism, so the route computation can consider the average
1590 recent load. Alternatively, the routing protocol could be
1591 (re-)designed to provide multiple alternative routes, and
1592 reservation setup could be attempted along each in turn.
1593
1594 3. Adapt to a route failure
1595
1596 When some node or link fails, adaptive routing finds an
1597 alternate path. The periodic refresh messages of RSVP will
1598 automatically request a reservation along the new path. Of
1599 course, this reservation may fail because there is
1600 insufficienct available capacity on the new path. This is a
1601 problem of provisioning and network engineering, which cannot
1602 be solved by the routing or setup protocols.
1603
1604 There is a problem of timeliness of establishing reservation
1605 state on the new path. The end-to-end robustness mechanism
1606 of refreshes is limited in frequency by overhead, which may
1607 cause a gap in realtime service when an old route breaks and
1608 a new one is chosen. It should be possible to engineer RSVP
1609 to sypplement the global refresh mechanism with a local
1610 repair mechanism, using hints about route changes from the
1611 routing mechanism.
1612
1613 4. Adapt to a route change (without failure)
1614
1615 Route changes may occur even without failure in the affected
1616 path. Although RSVP could use the same repair techniques as
1617
1618
1619
1620 <span class="grey">Braden, Clark &amp; Shenker [Page 29]</span></pre>
1621 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-30" ></span>
1622 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1623
1624
1625 those described in (3), this case raises a problem with the
1626 robustness of the QoS guarantees. If it should happen that
1627 admission control fails on the new route, the user will see
1628 service degradation unnecessarily and capriciously, since the
1629 orginal route is still functional.
1630
1631 To avoid this problem, a mechanism called "route pinning" has
1632 been suggested. This would modify the routing protocol
1633 implementation and the interface to the classifier, so that
1634 routes associated with resource reservations would be
1635 "pinned". The routing prootocol would not change a pinned
1636 route if it was still viable.
1637
1638 It may eventually be possible to fold together the routing and
1639 reservation setup problems, but we do not yet understand enough to
1640 do that. Furthermore, the reservation protocol needs to coexist
1641 with a number of different routing protocols in use in the
1642 Internet. Therefore, RSVP is currently designed to work with any
1643 current-generation routing protocol without modification. This is
1644 a short-term compromise, which may result in an occasional failure
1645 to create the best, or even any, real-time session, or an
1646 occasional service degradation due to a route change. We expect
1647 that future generations of routing protocols will remove this
1648 compromise, by including hooks and mechanisms that, in conjunction
1649 with RSVP, will solve the problems (1) through (4) just listed.
1650 They will support route pinning, notification of RSVP to trigger
1651 local repair, and selection of routes with "IS" support and
1652 adequate capacity.
1653
1654 The last routing-related issue is provided by mobile hosts. Our
1655 conjecture is that mobility is not essentially different from
1656 other route changes, so that the mechanism suggested in (3) and
1657 (4) will suffice. More study and experimentation is needed to
1658 prove or disprove this conjecture.
1659
1660 <span class="h2"><a class="selflink" id="section-6" href="#section-6">6</a>. ACKNOWLEDGMENTS</span>
1661
1662 Many Internet researchers have contributed to the work described in
1663 this memo. We want to especially acknowledge, Steve Casner, Steve
1664 Deering, Deborah Estrin, Sally Floyd, Shai Herzog, Van Jacobson,
1665 Sugih Jamin, Craig Partridge, John Wroclawski, and Lixia Zhang. This
1666 approach to Internet integrated services was initially discussed and
1667 organized in the End-to-End Research Group of the Internet Research
1668 Taskforce, and we are grateful to all members of that group for their
1669 interesting (and sometimes heated) discussions.
1670
1671
1672
1673
1674
1675
1676 <span class="grey">Braden, Clark &amp; Shenker [Page 30]</span></pre>
1677 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-31" ></span>
1678 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1679
1680
1681 REFERENCES
1682
1683 [<a id="ref-CerfKahn74">CerfKahn74</a>] Cerf, V., and R. Kahn, "A Protocol for Packet Network
1684 Intercommunication", IEEE Trans on Comm., Vol. Com-22, No. 5, May
1685 1974.
1686
1687 [<a id="ref-Clark88">Clark88</a>] Clark, D., "The Design Philosophy of the DARPA Internet
1688 Protocols", ACM SIGCOMM '88, August 1988.
1689
1690 [<a id="ref-CSZ92">CSZ92</a>] Clark, D., Shenker, S., and L. Zhang, "Supporting Real-Time
1691 Applications in an Integrated Services Packet Network: Architecture
1692 and Mechanisms", Proc. SIGCOMM '92, Baltimore, MD, August 1992.
1693
1694 [<a id="ref-DKS89">DKS89</a>] Demers, A., Keshav, S., and S. Shenker. "Analysis and
1695 Simulation of a Fair Queueing Algorithm", Journal of
1696 Internetworking: Research and Experience, 1, pp. 3-26, 1990. Also
1697 in Proc. ACM SIGCOMM '89, pp 3-12.
1698
1699 [<a id="ref-SCZ93a">SCZ93a</a>] Shenker, S., Clark, D., and L. Zhang, "A Scheduling Service
1700 Model and a Scheduling Architecture for an Integrated Services
1701 Packet Network", submitted to ACM/IEEE Trans. on Networking.
1702
1703 [<a id="ref-SCZ93b">SCZ93b</a>] Shenker, S., Clark, D., and L. Zhang, "A Service Model for the
1704 Integrated Services Internet", Work in Progress, October 1993.
1705
1706 [<a id="ref-Floyd92">Floyd92</a>] Floyd, S., "Issues in Flexible Resource Management for
1707 Datagram Networks", Proceedings of the 3rd Workshop on Very High
1708 Speed Networks, March 1992.
1709
1710 [<a id="ref-Jacobson91">Jacobson91</a>] Jacobson, V., "Private Communication", 1991.
1711
1712 [<a id="ref-JCSZ92">JCSZ92</a>] Jamin, S., Shenker, S., Zhang, L., and D. Clark, "An Admission
1713 Control Algorithm for Predictive Real-Time Service", Extended
1714 abstract, in Proc. Third International Workshop on Network and
1715 Operating System Support for Digital Audio and Video, San Diego, CA,
1716 Nov. 1992, pp. 73-91.
1717
1718 [<a id="ref-Parekh92">Parekh92</a>] Parekh, A., "A Generalized Processor Sharing Approach to
1719 Flow Control in Integrated Services Networks", Technical Report
1720 LIDS-TR-2089, Laboratory for Information and Decision Systems,
1721 Massachusetts Institute of Technology, 1992.
1722
1723 [<a id="ref-Partridge92">Partridge92</a>] Partridge, C., "A Proposed Flow Specification", <a href="./rfc1363">RFC 1363</a>,
1724 BBN, July 1992.
1725
1726 [<a id="ref-RSVP93a">RSVP93a</a>] Zhang, L., Deering, S., Estrin, D., Shenker, S., and D.
1727 Zappala, "RSVP: A New Resource ReSerVation Protocol", Accepted for
1728 publication in IEEE Network, 1993.
1729
1730
1731
1732 <span class="grey">Braden, Clark &amp; Shenker [Page 31]</span></pre>
1733 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-32" ></span>
1734 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1735
1736
1737 [<a id="ref-RSVP93b">RSVP93b</a>] Zhang, L., Braden, R., Estrin, D., Herzog, S., and S. Jamin,
1738 "Resource ReSerVation Protocol (RSVP) - Version 1 Functional
1739 Specification", Work in Progress, 1993.
1740
1741 [<a id="ref-ST2-90">ST2-90</a>] Topolcic, C., "Experimental Internet Stream Protocol: Version
1742 2 (ST-II)", <a href="./rfc1190">RFC 1190</a>, BBN, October 1990.
1743
1744 [<a id="ref-Tenet90">Tenet90</a>] Ferrari, D., and D. Verma, "A Scheme for Real-Time Channel
1745 Establishment in Wide-Area Networks", IEEE JSAC, Vol. 8, No. 3, pp
1746 368-379, April 1990.
1747
1748 Security Considerations
1749
1750 As noted in <a href="#section-2.1">Section 2.1</a>, the ability to reserve resources will create
1751 a requirement for authentication, both of users requesting resource
1752 guarantees and of packets that claim to have the right to use those
1753 guarantees. These authentication issues are not otherwise addressed
1754 in this memo, but are for further study.
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788 <span class="grey">Braden, Clark &amp; Shenker [Page 32]</span></pre>
1789 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-33" ></span>
1790 <span class="grey"><a href="./rfc1633">RFC 1633</a> Integrated Services Architecture June 1994</span>
1791
1792
1793 Authors' Addresses
1794
1795 Bob Braden
1796 USC Information Sciences Institute
1797 4676 Admiralty Way
1798 Marina del Rey, CA 90292
1799
1800 Phone: (310) 822-1511
1801 EMail: [email protected]
1802
1803
1804 David Clark
1805 MIT Laboratory for Computer Science
1806 545 Technology Square
1807 Cambridge, MA 02139-1986
1808
1809 Phone: (617) 253-6003
1810 EMail: [email protected]
1811
1812
1813 Scott Shenker
1814 Xerox Palo Alto Research Center
1815 3333 Coyote Hill Road
1816 Palo Alto, CA 94304
1817
1818 Phone: (415) 812-4840
1819 EMail: [email protected]
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844 Braden, Clark &amp; Shenker [Page 33]
1845 </pre>