idnits 2.17.1 draft-ietf-bmwg-ospfconv-applicability-02.txt: ** The Abstract section seems to be numbered Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- ** Looks like you're using RFC 2026 boilerplate. This must be updated to follow RFC 3978/3979, as updated by RFC 4748. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- ** The document seems to lack a 1id_guidelines paragraph about Internet-Drafts being working documents. ** The document seems to lack a 1id_guidelines paragraph about 6 months document validity. == No 'Intended status' indicated for this document; assuming Proposed Standard == The page length should not exceed 58 lines per page, but there was 10 longer pages, the longest (page 2) being 60 lines == It seems as if not all pages are separated by form feeds - found 0 form feeds but 11 pages Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack an Introduction section. ** The document seems to lack a Security Considerations section. ** The document seems to lack an IANA Considerations section. (See Section 2.2 of https://www.ietf.org/id-info/checklist for how to handle the case when there are no actions for IANA.) ** The abstract seems to contain references ([BENCHMARK], [TERM]), which it shouldn't. Please replace those with straight textual mentions of the documents in question. Miscellaneous warnings: ---------------------------------------------------------------------------- -- The document seems to lack a disclaimer for pre-RFC5378 work, but may have content which was first submitted before 10 November 2008. If you have contacted all the original authors and they are all willing to grant the BCP78 rights to the IETF Trust, then this is fine, and you can ignore this comment. If not, you may need to add the pre-RFC5378 disclaimer. (See the Legal Provisions document at https://trustee.ietf.org/license-info for more information.) -- The document date (March 2003) is 7706 days in the past. Is this intentional? Checking references for intended status: Proposed Standard ---------------------------------------------------------------------------- (See RFCs 3967 and 4897 for information about using normative references to lower-maturity documents in RFCs) -- Missing reference section? 'BENCHMARK' on line 386 looks like a reference -- Missing reference section? 'TERM' on line 37 looks like a reference -- Missing reference section? 'INTERCONNECT' on line 395 looks like a reference -- Missing reference section? 'FIB-TERM' on line 399 looks like a reference Summary: 8 errors (**), 0 flaws (~~), 3 warnings (==), 6 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group Vishwas Manral 3 Internet Draft Netplane Systems 4 Russ White 5 Cisco Systems 6 Aman Shaikh 7 Expiration Date: September 2003 University of California 8 File Name: draft-ietf-bmwg-ospfconv-applicability-02.txt March 2003 10 Benchmarking Applicability for Basic OSPF Convergence 11 draft-ietf-bmwg-ospfconv-applicability-02.txt 13 1. Status of this Memo 15 This document is an Internet-Draft and is in full conformance with 16 all provisions of Section 10 of RFC2026. 18 Internet Drafts are working documents of the Internet Engineering 19 Task Force (IETF), its Areas, and its Working Groups. Note that other 20 groups may also distribute working documents as Internet Drafts. 22 Internet Drafts are draft documents valid for a maximum of six 23 months. Internet Drafts may be updated, replaced, or obsoleted by 24 other documents at any time. It is not appropriate to use Internet 25 Drafts as reference material or to cite them other than as a "working 26 draft" or "work in progress". 28 The list of current Internet-Drafts can be accessed at 29 http://www.ietf.org/ietf/1id-abstracts.txt 31 The list of Internet-Draft Shadow Directories can be accessed at 32 http://www.ietf.org/shadow.html. 34 2. Abstract 36 This draft describes the applicability of [BENCHMARK] and similar 37 work which may be done in the future. Refer to [TERM] for terminology 38 used in this draft and [BENCHMARK]. The draft defines the advantages 39 as well as limitations of using the method defined in [BENCHMARK], 40 besides describing the pitfalls to avoid during measurement. 42 3. Motivation 44 There is a growing interest in testing SR-Convergence for routing 45 protocols, with many people looking at testing methodologies which 46 can provide information on how long it takes for a network to 47 converge after various network events occur. It is important to 48 consider the framework within which any given convergence test is 49 executed when attempting to apply the results of the testing, since 50 the framework can have a major impact on the results. For instance, 51 determining when a network is converged, what parts of the router's 52 operation are considered within the testing, and other such things 53 will have a major impact on what apparent performance routing 54 protocols provide. 56 This document describes in detail the various benefits and pitfalls 57 of tests described in [BENCHMARK]. It also explains how such 58 measurements can be useful for providers and the research community. 60 4. Advantages of Such Measurement 62 o To be able to compare the iterations of a protocol implemen- 63 tation. It is often useful to be able to compare the perfor- 64 mance of two iterations of a given implementation of a proto- 65 col to determine where improvements have been made and where 66 further improvements can be made. 68 o To understand, given a set parameters (network conditions), 69 how a particular implementation on a particular device is 70 going to perform. For instance, if you were trying to decide 71 the processing power (size of device) required in a certain 72 location within a network, you can emulate the conditions 73 which are going to exist at that point in the network and use 74 the test described to measure the perfomance of several dif- 75 ferent routers. The results of these tests can provide one 76 possible data point for an intelligent decision. 78 If the device being tested is to be deployed in a running 79 network, using routes taken from the network where the equip- 80 ment is to be deployed rather than some generated topology in 81 these tests will give results which are closer to the real 82 preformance of the device. Care should be taken to emulate or 83 take routes from the actual location in the network where the 84 device will be (or would be) deployed. For instance, one set 85 of routes may be taken from an abr, one set from an area 0 86 only router, various sets from stub area, another set from 87 various normal areas, etc. 89 o To measure the performance of an OSPF implementation in a 90 wide variety of scenarios. 92 o To be used as parameters in OSPF simulations by researchers. 93 It may some times be required for certain kinds of research 94 to measure the individual delays of each parameter within an 95 OSPF implementation. These delays can be measured using the 96 methods defined in [BENCHMARK]. 98 o To help optimize certain configurable parameters. It may some 99 times be helpful for operators to know the delay required for 100 individual tasks so as to optimize the resource usage in the 101 network i.e. if it is found that the processing time is x 102 seconds on an router, it would be helpful to determine the 103 rate at which to flood LSA's to that router so as to not 104 overload the network. 106 5. Assumptions Made and Limitations of such measurements 108 o The interactions of SR-Convergence and forwarding; testing is res- 109 tricted to events occurring within the control plane. Forwarding 110 performance is the primary focus in [INTERCONNECT] and it is 111 expected to be dealt with in work that ensues from [FIB-TERM]. 113 o Duplicate LSAs are Acknowledged Immediately. A few tests rely on 114 the property that duplicate LSA Acknowledgements are not delayed 115 but are done immediately. However if some implementation does not 116 acknowledge duplicate LSAs immediately on receipt, the testing 117 methods presented in [BENCHMARK] could give inaccurate measure- 118 ments. 120 o It is assumed that SPF is non-preemptive. If SPF is implemented so 121 that it can (and will be) preempted, the SPF measurements taken in 122 [BENCHMARK] would include the times that the SPF process is not 123 running ([BENCHMARK] measures the total time taken for SPF to run, 124 not the amount of time that SPF actually spends on the device's 125 processor), thus giving inaccurate measurements. 127 o Some implementations may be multithreaded or use a 128 multiprocess/multirouter model of OSPF. If because of this any of 129 the assumptions taken in measurement are violated in such a model, 130 it could lead to inaccurate measurements. 132 o The measurements resulting from the tests in [BENCHMARK] may not 133 provide the information required to deploy a device in a large 134 scale network. The tests described focus on individual components 135 of an OSPF implementation's performance, and it may be difficult 136 to combine the measurements in a way which accurately depicts a 137 device's performance in a large scale network. Further research is 138 required in this area. 140 o The measurements described in [BENCHMARK] should be used with 141 great care when comparing two different implementations of OSPF 142 from two different vendors. For instance, there are many other 143 factors than convergence speed which must be taken into considera- 144 tion when comparing different vendor's products, and it's diffi- 145 cult to align the resources available on one device to the 146 resources available on another device. 148 6. Observations on the Tests Described in [BENCHMARK] 150 Some observations taken while implementing the tests described in 151 [BENCHMARK] are noted in this section. 153 6.1. Measuring the SPF Processing Time Externally 155 The most difficult test to perform is the external measurement of the 156 time required to perform an SPF calculation, since the amount of time 157 between the first LSA which indicates a topology change and the 158 duplicate LSA is critical. If the duplicate LSA is sent too quickly, 159 it may be received before the device under test actually begins run- 160 ning SPF on the network change information. If the delay between the 161 two LSAs is too long, the device under test may finish SPF processing 162 before receiving the duplicate LSA. It is important to closely inves- 163 tigate any delays between the receipt of an LSA and the beginning of 164 an SPF calculation in the device under test; multiple tests with 165 various delays might be required to determine what delay needs to be 166 used to accurately measure the SPF calculation time. 168 6.2. Noise in the Measurement Device 170 The device on which measurements are taken (not the device under 171 test) also adds noise to the test results, primarily in the form of 172 delay in packet processing and measurement output. The largest source 173 of noise is generally the delay between the receipt of packets by the 174 measuring device and the information about the packet reaching the 175 device's output, where the event can be measured. The following steps 176 may be taken to reduce this sampling noise: 178 o Increasing the number of samples taken will generally improve 179 the tester's ability to determine what is noise, and remove it 180 from the results. 182 o Try to take time-stamp for a packet as early as possible. 183 Depending on the operating system being used on the box, one 184 can instrument the kernel to take the time-stamp when the 185 interrupt is processed. This does not eliminate the noise com- 186 pletely, but at least reduces it. 188 o Keep the measurement box as lightly loaded as possible. 190 o Having an estimate of noise can also be useful. 192 The DUT also adds noise to the measurement. Points (a) and (c) 193 apply to the DUT as well. 195 6.3. Gaining an Understanding of the Implementation Improves Measure- 196 ments 198 While the tester will (generally) not have access to internal infor- 199 mation about the OSPF implementation being tested using [BENCHMARK], 200 the more thorough the tester's knowledge of the implementation is, 201 the more accurate the results of the tests will be. For instance, in 202 some implementations, the installation of routes in local routing 203 tables may occur while the SPF is being calculated, dramatically 204 impacting the time required to calculate the SPF. 206 6.4. Gaining an Understanding of the Tests Improves Measurements 208 One method which can be used to become familiar with the tests 209 described in [BENCHMARK] is to perform the tests on an OSPF implemen- 210 tation for which all the internal details are available, such as 211 GateD. While there is no assurance that any two implementations will 212 be similar, this will provide a better understanding of the tests 213 themselves. 215 7. LSA and Destination mix 217 In many OSPF benchmark tests, a generator injecting a number of LSAs 218 is called for. There are several areas in which injected LSAs can be 219 varied in testing: 221 o The number of destinations represented by the injected LSAs 223 Each destination represents a single reachable IP network; 224 these will be leaf nodes on the shortest path tree. The pri- 225 mary impact to performance should be the time required to 226 insert destinations in the local routing table and handling 227 the memory required to store the data. 229 o The types of LSAs injected 231 There are several types of LSAs which would be acceptable 232 under different situations; within an area, for instance, 233 type 1, 2, 3, 4, and 5 are likely to be received by a router. 234 Within a not-so-stubby area, however, type 7 LSAs would 235 replace the type 5 LSAs received. These sorts of characteri- 236 zations are important to note in any test results. 238 o The Number of LSAs injected 240 Within any injected set of information, the number of each 241 type of LSA injected is also important. This will impact the 242 shortest path algorithms ability to handle large numbers of 243 nodes, large shortest path first trees, etc. 245 o The Order of LSA Injection 247 The order in which LSAs are injected should not favor any 248 given data structure used for storing the LSA database on the 249 device under test. For instance, AS-External LSA's have AS 250 wide flooding scope; any Type-5 LSA originated is immediately 251 flooded to all neighbors. However the Type-4 LSA which 252 announces the ASBR as a border router is originated in an 253 area at SPF time (by ABR's on the edge of the area in which 254 the ASBR is). If SPF isn't scheduled immediately on the ABRs 255 originating the type 4 LSA, the type-4 LSA is sent after the 256 type-5 LSA's reach a router in the adjacent area. So routes 257 to the external destinations aren't immediately added to the 258 routers in the other areas. When the routers which already 259 have the type 5's receive the type-4 LSA, all the external 260 routes are added to the tree at the same time. This timing 261 could produce different results than a router receiving a 262 type 4 indicating the presence of a border router, followed 263 by the type 5's originated by that border router. 265 The ordering can be changed in various tests to provide 266 insight on the efficiency of storage within the DUT. Any such 267 changes in ordering should be noted in test results. 269 8. Tree Shape and the SPF Algorithm 271 The complexity of Dijkstra's algo depends on the data structure used 272 for storing vertices with their current minimum distances from the 273 source. The simplest structure is a list of vertices currently reach- 274 able from the source. Finding the minimum cost vertex then would take 275 O(size of the list). There will be O(n) such operations if we assume 276 that all the vertices are ultimately reachable from the source. More- 277 over, after the vertex with min cost is found, the algo iterates thru 278 all the edges of the vertex and updates cost of other vertices. With 279 an adjacency list representation, this step when iterated over all 280 the vertices, would take O(E) time. Thus, overall running time is: 282 O(sum(i:1, n)(size(list at level i) + E). 284 So, everything boils down to the size(list at level i). 286 If the graph is linear: 288 root 289 | 290 1 291 | 292 2 293 | 294 3 295 | 296 4 297 | 298 5 299 | 300 6 302 and source is a vertex on the end, then size(list at level i) 303 = 1 for all i. Moreover, E = n - 1. Therefore, running time 304 is O(n). 306 If graph is a balanced binary tree: 308 root 309 / \ 310 1 2 311 / \ / \ 312 3 4 5 6 314 size(list at level i) is a little complicated. First it 315 increases by 1 at each level upto a certain number, and then 316 goes down by 1. If we assumed that tree is a complete tree 317 (like the one in the draft) with k levels (1 to k), then 318 size(list) goes on like this: 1, 2, 3, 320 Then the number of edges E is still n - 1. It then turns out 321 that the run-time is O(n^2) for such a tree. 323 If graph is a complete graph (fully-connected mesh), then 324 size(list at level i) = n - i. Number of edges E = O(n^2). 325 Therefore, run-time is O(n^2). 327 shortest path first algorithm to compute the best paths 328 through the network need to be aware that the construction of 329 the tree may impact the performance of the algorithm. Best 330 practice would be to try and make any emulated network look 331 as much like a real network as possible, especially in the 332 area of the tree depth, the meshiness of the network, the 333 number of stub links verses transit links, and the number of 334 connections and nodes to process at each level within the 335 original tree. 337 9. Topology Generation 339 As the size of networks grows, it becomes more and more difficult to 340 actually create a large scale network on which to test the properties 341 of routing protocols and their implementations. In general, network 342 emulators are used to provide emulated topologies which can be adver- 343 tised to a device with varying conditions. Route generators either 344 tend to be a specialized device, a piece of software which runs on a 345 router, or a process that runs on another operating system, such as 346 Linux or another variant of Unix. 348 Some of the characteristics of this device should be: 350 o The ability to connect to the several devices using both point- 351 to-point and broadcast high speed media. Point-to-point links can 352 be emulated with high speed Ethernet as long as there is no hub or 353 other device in between the DUT and the route generator, and the 354 link is configured as a point-to-point link within OSPF. 356 o The ability to create a set of LSAs which appear to be a logical, 357 realistic topology. For instance, the generator should be able to 358 mix the number of point-to-point and broadcast links within the 359 emulated topology, and should be able to inject varying numbers of 360 externally reachable destinations. 362 o The ability to withdraw and add routing information into and from 363 the emulated topology to emulate links flapping. 365 o The ability to randomly order the LSAs representing the emulated 366 topology as they are advertised. 368 o The ability to log or otherwise measure the time between packets 369 transmitted and received. 371 o The ability to change the rate at which OSPF LSAs are transmitted. 373 o The generator and the collector should be fast enough so that they 374 are not bottle necks. The devices should also have a degree of 375 granularity of measurement atleast as small as desired from the 376 test results. 378 10. Acknowledgements 380 Thanks to Howard Berkowitz, (hcb@clark.net) and the rest of the BGP 381 benchmarking team for their support and to Kevin 382 Dubray(kdubray@juniper.net) who realized the need of this draft. 384 11. Normative References 386 [BENCHMARK] 387 Manral, V., "Benchmarking Methodology for Basic OSPF Convergence", 388 draft-bmwg-ospfconv-intraarea-04, March 2003 390 [TERM]Manral, V., "OSPF Convergence Testing Terminiology and Concepts", 391 draft-bmwg-ospfconv-term-03.txt, March 2003 393 12. Informative References 395 [INTERCONNECT] 396 Bradner, S., McQuaid, J., "Benchmarking Methodology for Network 397 Interconnect Devices", RFC2544, March 1999. 399 [FIB-TERM] 400 Trotter, G., "Terminology for Forwarding Information Base (FIB) 401 based Router Performance", RFC3222, October 2001. 403 13. Authors' Addresses 404 Vishwas Manral 405 Netplane Systems 406 189 Prashasan Nagar 407 Road number 72 408 Jubilee Hills 409 Hyderabad, India 411 vmanral@netplane.com 413 Russ White 414 Cisco Systems, Inc. 415 7025 Kit Creek Rd. 416 Research Triangle Park, NC 27709 418 riw@cisco.com 420 Aman Shaikh 421 University of California 422 School of Engineering 423 1156 High Street 424 Santa Cruz, CA 95064 425 aman@soe.ucsc.edu