Debian Codegraph @

rfc1104.html

raw
1 <pre>Network Working Group H-W. Braun
2 Request for Comments: 1104 Merit/NSFNET
3 June 1989
4
5
6 <span class="h1">Models of Policy Based Routing</span>
7
8 <span class="h2"><a class="selflink" id="section-1" href="#section-1">1</a>. Status of this Memo</span>
9
10 The purpose of this RFC is to outline a variety of models for policy
11 based routing. The relative benefits of the different approaches are
12 reviewed. Discussions and comments are explicitly encouraged to move
13 toward the best policy based routing model that scales well within a
14 large internetworking environment.
15
16 Distribution of this memo is unlimited.
17
18 <span class="h2"><a class="selflink" id="section-2" href="#section-2">2</a>. Acknowledgements</span>
19
20 Specific thanks go to Yakov Rekhter (IBM Research), Milo Medin
21 (NASA), Susan Hares (Merit/NSFNET), Jessica Yu (Merit/NSFNET) and
22 Dave Katz (Merit/NSFNET) for extensively contributing to and
23 reviewing this document.
24
25 <span class="h2"><a class="selflink" id="section-3" href="#section-3">3</a>. Overview</span>
26
27 To evaluate the methods and models for policy based routing, it is
28 necessary to investigate the context into which the model is to be
29 used, as there are a variety of different methods to introduce
30 policies. Most frequently the following three models are referenced:
31
32 Policy based distribution of routing information
33 Policy based packet filtering/forwarding
34 Policy based dynamic allocation of network resources (e.g.,
35 bandwidth, buffers, etc.)
36
37 The relative properties of those methods need to be evaluated to find
38 their merits for a specific application. In some cases, more than
39 one method needs to be implemented.
40
41 While comparing different models for policy based routing, it is
42 important to realize that specific models have been designed to
43 satisfy a certain set of requirements. For different models these
44 requirements may or may not overlap. Even if they overlap, they may
45 have a different degree of granularity. In the first model, the
46 requirements can be formulated at the Administrative Domain or
47 network number level. In the second model, the requirements can be
48 formulated at the end system level or probably even at the level of
49
50
51
52 <span class="grey">Braun [Page 1]</span></pre>
53 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-2" ></span>
54 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
55
56
57 individual users. In the third model, the requirements need to be
58 formulated at both the end system and local router level, as well as
59 at the level of Routing Domains and Administrative Domains.
60
61 Each of these models looks at the power of policy based routing in a
62 different way. They may be implemented separately or in combination
63 with other methods. The model to describe policy based dynamic
64 allocation of network resources is orthogonal to the model of policy
65 based distribution of routing information. However, in an actual
66 implementation each of these models may interact.
67
68 It is important to realize that the use of a policy based scheme for
69 individual network applications requires that the actual effects as
70 well as the interaction of multiple methods need to be determined
71 ahead of time by policy.
72
73 While uncontrolled dynamic routing and allocation of resources may
74 have a better real time behavior, the use of policy based routing
75 will provide a predictable, stable result based on the desires of the
76 administrator. In a production network, it is imperative to provide
77 continuously consistent and acceptable services.
78
79 <span class="h2"><a class="selflink" id="section-4" href="#section-4">4</a>. Policy based distribution of routing information</span>
80
81 Goals:
82
83 The goal of this model is to enforce certain flows by means of
84 policy based distribution of routing information. This
85 enforcement allows control over who can and who can not use
86 specific network resources.
87
88 Enforcement is done at the network or Administrative Domain (AD)
89 level - macroscopic policies.
90
91 Description:
92
93 A good example of policy based routing based on the distribution
94 of routing information is the NSFNET with its interfaces to mid-
95 level networks [<a href="#ref-1" title="&quot;EGP and Policy Based Routing in the New NSFNET Backbone&quot;">1</a>], [<a href="#ref-2" title="&quot;The NSFNET Routing Architecture&quot;">2</a>]. At the interface into the NSFNET, the
96 routing information is authenticated and controlled by four means:
97
98 1. Routing peer authentication based on the source address.
99
100 2. Verification of the Administrative Domain identification
101 (currently EGP Autonomous System numbers).
102
103 3. Verification of Internet network numbers which are
104 advertised via the routing peer.
105
106
107
108 <span class="grey">Braun [Page 2]</span></pre>
109 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-3" ></span>
110 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
111
112
113 4. Control of metrics via a Routing Policy Data Base for the
114 announced Internet network numbers to allow for primary
115 paths to the NSFNET as well as for paths of a lesser
116 degree.
117
118 At the interfaces that pass routing traffic out of the NSFNET, the
119 NSS routing code authenticates the router acting as an EGP peer by
120 its address as well as the Administrative Domain identification
121 (Autonomous System Number).
122
123 Outbound announcements of network numbers via the EGP protocol are
124 controlled on the basis of Administrative Domains or individual
125 network numbers by the NSFNET Routing Policy Data Base.
126
127 The NSFNET routing policy implementation has been in place since
128 July 1988 and the NSFNET community has significant experience with
129 its application.
130
131 Another example of policy controlled dissimination of routing
132 information is a method proposed for ESNET in [<a href="#ref-3" title="&quot;ESNET Routing&quot;">3</a>].
133
134 Benefits:
135
136 A major merit of the control of routing information flow is that
137 it enables the engineering of large wide area networks and allows
138 for a more meshed environment than would be possible without tight
139 control. Resource allocation in a non-hostile environment is
140 possible by filtering specific network numbers or Administrative
141 Domains on a per need basis. Another important benefit of this
142 scheme is that it allows for network policy control with virtually
143 no performance degradation, as only the routing packets themselves
144 are relevant for policy control. Routing tables are generated as
145 a result of these interactions. This means that this scheme
146 imposes only very little impact on packet switching performance at
147 large.
148
149 Concerns:
150
151 Policy based routing information distribution does not address
152 packet based filtering. An example is the inability to prevent
153 malicious attacks by introduced source routed IP packets. While
154 resource allocation is possible, it extends largely to filtering
155 on network numbers or whole Administrative Domains, but it would
156 not extend to end systems or individual users.
157
158 Costs:
159
160 Policy based routing in the NSFNET is implemented in a series of
161
162
163
164 <span class="grey">Braun [Page 3]</span></pre>
165 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-4" ></span>
166 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
167
168
169 configuration files. These configuration files are in turn
170 generated from a routing information database. The careful
171 creation of this routing information database requires knowledge
172 of the Internet at large. Because the Internet is changing
173 constantly, the upkeep of this routing information database is a
174 continuous requirement. However, the effort of collecting and
175 maintaining an accurate view of the Internet at large can be
176 distributed.
177
178 Since policy controlled distribution of routing information allows
179 for filtering on the basis of network numbers or Administrative
180 Domains, the routing information database only needs to collect
181 information for the more than 1300 networks within the Internet
182 today.
183
184 <span class="h2"><a class="selflink" id="section-5" href="#section-5">5</a>. Policy based packet filtering/forwarding</span>
185
186 Goals:
187
188 The goal of the model of policy based packet filtering/forwarding
189 is to allow the enforcement of certain flows of network traffic on
190 a per packet basis. This enforcement allows the network
191 administrator to control who can and who can not use specific
192 network resources.
193
194 Enforcement may be done at the end system or even individual user
195 level - microscopic policies.
196
197 Description:
198
199 An example of packet/flow based policies is outlined in [<a href="#ref-4" title="&quot;Policy Routing in Internet Protocols&quot;">4</a>]. In a
200 generic sense, policy based packet filtering/forwarding allows
201 very tight control of the distribution of packet traffic. An
202 implemented example of policy based filtering/forwarding is a
203 protection mechanism built into the NSFNET NSS structure, whereby
204 the nodes can protect themselves against packets targeted at the
205 NSFNET itself by filtering according to IP destination. While this
206 feature has so far not been enabled, it is fully implemented and
207 can be turned on within a matter of seconds.
208
209 Benefits:
210
211 The principal merit of this scheme is that it allows the
212 enforcement of packet policies and resource allocation down to
213 individual end systems and perhaps even individual end users. It
214 does not address a sane distribution of routing information. If
215 policies are contained in the packets themselves it could identify
216 users, resulting in the ability of users to move between
217
218
219
220 <span class="grey">Braun [Page 4]</span></pre>
221 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-5" ></span>
222 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
223
224
225 locations.
226
227 Concerns:
228
229 The major concern would be the potentially significant impact on
230 the performance of the routers, as, at least for tight policy
231 enforcements, each packet to be forwarded would need to be
232 verified against a policy data base. This limitation makes the
233 application of this scheme questionable using current Internet
234 technology, but it may be very applicable to circuit switched
235 environments (with source-routed IP packets being similar to a
236 circuit switched environment). Another difficulty could be the
237 sheer number of potential policies to be enforced, which could
238 result in a very high administrative effort. This could result
239 from the creation of policies at the per-user level. Furthermore,
240 the overhead of carrying policy information in potentially every
241 packet could result in additional burdens on resource
242 availabilities. This again is more applicable to connection-
243 oriented networks, such as public data networks, where the policy
244 would only need to be verified at the call setup time. It is an
245 open question how well packet based policies will scale in a large
246 and non homogeneous Internet environment, where policies may be
247 created by all of the participants. These creations of policy
248 types of services may have to be doable in real time.
249
250 Scaling may require hierarchy. Hierarchy may conflict with
251 arbitrary Type of Service (TOS) routing, which is one of the
252 benefits of this model.
253
254 Costs of implementation:
255
256 A large scale implemention of packet based policy routing would
257 require a routing information base that would contain information
258 down to the end system level and possibly end users. If one would
259 assume that for each of the 1300 networks there is an average of
260 200 end systems, this would result in over 260000 end systems
261 Internet wide. Each end system in turn could further contribute
262 some information on the type of traffic desired, including types
263 of service (issues like agency network selection), potentially on
264 a per-user basis. The effort for the routing policy data base
265 could be immense, in particular if there is a scaling requirement
266 towards a variety of policies for backbones, mid-level networks,
267 campus networks, subnets, hosts, and users. The administration of
268 this "packet routing" database could be distributed. However,
269 with a fully distributed database of this size several consistency
270 checks would have to be built into the system.
271
272
273
274
275
276 <span class="grey">Braun [Page 5]</span></pre>
277 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-6" ></span>
278 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
279
280
281 <span class="h2"><a class="selflink" id="section-6" href="#section-6">6</a>. Policy based dynamic allocation of network resources (e.g.,</span>
282 <span class="h2"> bandwidth, buffers, etc.).</span>
283
284 Goals:
285
286 Flexible and economical allocation of network resources based on
287 current needs and certain policies. Policies may be formulated at
288 the network or Administrative Domain (AD) levels. It is also
289 possible to formulate policies which will regulate resource
290 allocation for different types of traffic (e.g., Telnet, FTP,
291 precedence indicators, network control traffic).
292
293 Enforcement of policy based allocation of network resources might
294 be implemented within the following parts of the network:
295
296 routers for networks and Administrative Domain (AD) levels
297 circuit switches for networks
298 end systems establishing network connections
299
300 Description:
301
302 Policy based allocation of bandwidth could allow the modulation of
303 the circuits of the networking infrastructure according to real
304 time needs. Assuming that available resources are limited towards
305 an upper bound, the allocation of bandwidth would need to be
306 controlled by policy. One example might be a single end system
307 that may or may not be allowed to, perhaps even automatically,
308 take resources away from other end systems or users. An example
309 of dynamic bandwidth allocation is the currently implemented
310 circuit switched IDNX component of the NSFNET, as well as the MCI
311 Digital Reconfiguration Service (DRS) which is planned for the
312 NSFNET later this year.
313
314 Another model for resource allocation occurs at the packet level,
315 where the allocation is controlled by multiple packet queues.
316 This could allow for precedence queuing, with preferences based on
317 some type of service and preferred forwarding of recognized
318 critical data, such as network monitoring, control and routing.
319 An example can be found in the NSFNET, where the NSFNET nodes
320 prefer traffic affiliated with the NSFNET backbone network number
321 over all other traffic, to allow for predictable passing of
322 routing information as well as effective network monitoring and
323 control. At the other end of the spectrum, an implementation
324 could also allow for queues of most deferrable traffic (such as
325 large background file transfers).
326
327
328
329
330
331
332 <span class="grey">Braun [Page 6]</span></pre>
333 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-7" ></span>
334 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
335
336
337 Benefits:
338
339 Dynamic allocation of bandwidth could allow for a truly flexible
340 environment where the networking infrastructure could create
341 bandwidth on a per need basis. This could result in significant
342 cost reductions during times when little bandwidth is needed.
343 This method could potentially accommodate real time transient high
344 bandwidth requirements, potentially by reducing the bandwidth
345 available to other parts of the infrastructure. A positive aspect
346 is that the bandwidth allocation could be protocol independent,
347 with no impact on routing protocols or packet forwarding
348 performance.
349
350 Policy based allocation of bandwidth can provide a predictable
351 dynamic environment. The rules about allocation of bandwidth at
352 the circuit level or at the packet level need to be determined by
353 a consistent and predictable policy, so that other networks or
354 Administrative Domains can tune their allocation of networking
355 resources at the same time.
356
357 Concerns:
358
359 The policies involved in making dynamic bandwidth allocation in a
360 largely packet switching environment possible are still in the
361 development phase. Even the technical implications of
362 infrastructure reconfiguration in result of events happening on a
363 higher level still requires additional research.
364
365 A policy based allocation of bandwidth could tune the network to
366 good performance, but could cause networks located in other
367 Administrative Domains to pass traffic poorly. It is important
368 that network resource policy information for a network be
369 discussed within the context of its Administrative Domain.
370 Administrative Domains need to discuss their network resource
371 allocation policies with other Administrative Domains.
372
373 The technical problem of sharing network resource policy
374 information could be solved by a making a "network resource policy
375 information" database available to all administrators of networks
376 and Administrative Domains. However, the political problems
377 involved in creating a network resource policy with impact on
378 multiple Administrative Domains does still require additional
379 study.
380
381 <span class="h2"><a class="selflink" id="section-7" href="#section-7">7</a>. Discussion</span>
382
383 Both the first and the second model of policy based routing are
384 similar in the sense that their goal is to enforce certain flows.
385
386
387
388 <span class="grey">Braun [Page 7]</span></pre>
389 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-8" ></span>
390 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
391
392
393 This enforcement allows the control of access to scarce network
394 resources (if the resource is not scarce, there is no performance
395 reason to control access to it). The major difference is the level
396 of enforcement: macroscopic level versus microscopic level control.
397
398 Associated with the enforcement for a certain network resource is the
399 cost. If this cost is higher than the cost required to make a
400 particular resource less scarce, then the feasibility of enforcement
401 may be questionable.
402
403 If portions of the Internet find that microscopic enforcement of
404 policy is necessary, then this will need to be implementable without
405 significant performance degradation to the networking environment at
406 large. Local policies within specific Routing Domains or
407 Administrative Domains should not affect global Internet traffic or
408 routing. Policies within Administrative Domains which act as traffic
409 transit systems (such as the NSFNET) should not be affected by
410 policies a single network imposes for its local benefit.
411
412 Some models of policy routing are trying to deal with cases where
413 network resources require rather complex usage policies. One of
414 scenarios in [<a href="#ref-4" title="&quot;Policy Routing in Internet Protocols&quot;">4</a>] is one in which a specific agency may have some
415 network resource (in the example it is a link) which is sometimes
416 underutilized. The goal is to sell this resource to other agencies
417 during the underutilization period to recover expenses. This
418 situation is equivalent to the problem of finding optimum routes,
419 with respect to a certain TOS, in the presence of network resources
420 (e.g., links) with variable characteristics. Any proposed solution
421 to this problem should address such issues as network and route
422 stability. More feasibility study is necessary for the whole
423 approach where links used for global communication are also subject
424 to arbitrary local policies. An alternative approach would be to
425 reconfigure the network topology so that underutilized links will be
426 dropped and possibly returned to the phone company. This is
427 comparable to what the NSFNET is planning on doing with the MCI
428 Digital Reconfiguration Service (DRS). A DRS model may appear
429 cleaner and more easy to implement than a complicated model like the
430 one outlined in [<a href="#ref-4" title="&quot;Policy Routing in Internet Protocols&quot;">4</a>].
431
432 The models for policy based routing emphasize that careful
433 engineering of the Internet needs to decided upon the profile of
434 traffic during normal times, outage periods, and peak loads. This
435 type of engineering is not a new requirement. However, there could
436 potentially be a significant benefit in deciding these policies ahead
437 of time and using policy based routing to implement specific routing
438 policies.
439
440
441
442
443
444 <span class="grey">Braun [Page 8]</span></pre>
445 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-9" ></span>
446 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
447
448
449 <span class="h2"><a class="selflink" id="section-8" href="#section-8">8</a>. Accounting vs. Policy Based Routing</span>
450
451 Quite often Accounting and Policy Based Routing are discussed
452 together. While the application of both Accounting and Policy Based
453 Routing is to control access to scarce network resources, these are
454 separate (but related) issues.
455
456 The chief difference between Accounting and Policy Based Routing is
457 that Accounting combines history information with policy information
458 to track network usage for various purposes. Accounting information
459 may in turn drive policy mechanisms (for instance, one could imagine
460 a policy limiting a certain organization to a fixed aggregate
461 percentage of dynamically shared bandwidth). Conversely, policy
462 information may affect accounting issues. Network accounting
463 typically involves route information (at any level from AD to end
464 system) and volume information (packet, octet counts).
465
466 Accounting may be implemented in conjunction with any of the policy
467 models mentioned above. Similar to the microscopic versus
468 macroscopic policies, accounting may be classified into different
469 levels. One may collect accounting data at the AD level, network
470 level, host level, or even at the individual user level. However,
471 since accounting may be organized hierarchically, microscopic
472 accounting may be supported at the network or host level, while
473 macroscopic accounting may be supported at the network or AD level.
474 An example might be the amount of traffic passed at the interface
475 between the NSFNET and a mid-level network or between a mid-level
476 network and a campus. Furthermore, the NSFNET has facilities
477 implemented to allow for accounting of traffic trends from individual
478 network numbers as well as application-specific information.
479
480 Full-blown accounting schemes suffer the same types of concerns
481 previously discussed, with the added complication of potentially
482 large amounts of additional data gathered that must be reliably
483 retrieved. As pointed out in [<a href="#ref-4" title="&quot;Policy Routing in Internet Protocols&quot;">4</a>], policy issues may impact the way
484 accounting data is collected (one administration billing for packets
485 that were then dropped in the network of another administration).
486 Microscopic accounting may not scale well in a large internet.
487
488 Furthermore, from the standpoint of billing, it is not clear that the
489 services provided at the network layer map well to the sorts of
490 services that network consumers are willing to pay for. In the
491 telephone network (as well as public data networks), users pay for
492 end-to-end service and expect good quality service in terms of error
493 rate and delay (and may be unwilling to pay for service that is
494 viewed as unacceptable). In an internetworking environment, the
495 heterogeneous administrative environment combined with the lack of
496 end-to-end control may make this approach infeasible.
497
498
499
500 <span class="grey">Braun [Page 9]</span></pre>
501 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-10" ></span>
502 <span class="grey"><a href="./rfc1104">RFC 1104</a> Models of Policy Based Routing June 1989</span>
503
504
505 Lightweight approaches to accounting can be used (with less impact)
506 when specific, limited goals are set. One suggested approach
507 involves monitoring traffic patterns. If a pattern of abuse (e.g.,
508 unauthorized use) develops, an accounting system could track this and
509 allow corrective action to be taken, by changing routing policy or
510 imposing access control (blocking hosts or nets). Note that this is
511 much less intrusive into the packet forwarding aspects of the
512 routers, but requires distribution of a policy database that the
513 accounting system can use to reduce the raw information. Because
514 this approach is statistical in nature, it may be slow to react.
515
516 <span class="h2"><a class="selflink" id="section-9" href="#section-9">9</a>. References</span>
517
518 [<a id="ref-1">1</a>] Rekhter, Y., "EGP and Policy Based Routing in the New NSFNET
519 Backbone", <a href="./rfc1092">RFC 1092</a>, IBM Research, February 1989.
520
521 [<a id="ref-2">2</a>] Braun, H-W., "The NSFNET Routing Architecture", <a href="./rfc1093">RFC 1093</a>,
522 Merit/NSFNET Project, February 1989.
523
524 [<a id="ref-3">3</a>] Collins, M., and R. Nitzan, "ESNET Routing", DRAFT Version 1.0,
525 LLNL, May 1989.
526
527 [<a id="ref-4">4</a>] Clark, D., "Policy Routing in Internet Protocols", <a href="./rfc1102">RFC 1102</a>,
528 M.I.T. Laboratory for Computer Science, May 1989.
529
530 Author's Address
531
532 Hans-Werner Braun
533 Merit Computer Network
534 University of Michigan
535 1075 Beal Avenue
536 Ann Arbor, Michigan 48109
537
538 Telephone: 313 763-4897
539 Fax: 313 747-3745
540 EMail: [email protected]
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556 Braun [Page 10]
557 </pre>