idnits 2.17.1 draft-ietf-p2psip-self-tuning-11.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (May 9, 2014) is 3638 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) ** Obsolete normative reference: RFC 5245 (Obsoleted by RFC 8445, RFC 8839) ** Obsolete normative reference: RFC 5389 (Obsoleted by RFC 8489) == Outdated reference: A later version (-09) exists of draft-ietf-p2psip-concepts-05 Summary: 2 errors (**), 0 flaws (~~), 2 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 P2PSIP Working Group J. Maenpaa 3 Internet-Draft G. Camarillo 4 Intended status: Standards Track Ericsson 5 Expires: November 10, 2014 May 9, 2014 7 Self-tuning Distributed Hash Table (DHT) for REsource LOcation And 8 Discovery (RELOAD) 9 draft-ietf-p2psip-self-tuning-11.txt 11 Abstract 13 REsource LOcation And Discovery (RELOAD) is a peer-to-peer (P2P) 14 signaling protocol that provides an overlay network service. Peers 15 in a RELOAD overlay network collectively run an overlay algorithm to 16 organize the overlay, and to store and retrieve data. This document 17 describes how the default topology plugin of RELOAD can be extended 18 to support self-tuning, that is, to adapt to changing operating 19 conditions such as churn and network size. 21 Status of This Memo 23 This Internet-Draft is submitted in full conformance with the 24 provisions of BCP 78 and BCP 79. 26 Internet-Drafts are working documents of the Internet Engineering 27 Task Force (IETF). Note that other groups may also distribute 28 working documents as Internet-Drafts. The list of current Internet- 29 Drafts is at http://datatracker.ietf.org/drafts/current/. 31 Internet-Drafts are draft documents valid for a maximum of six months 32 and may be updated, replaced, or obsoleted by other documents at any 33 time. It is inappropriate to use Internet-Drafts as reference 34 material or to cite them other than as "work in progress." 36 This Internet-Draft will expire on November 10, 2014. 38 Copyright Notice 40 Copyright (c) 2014 IETF Trust and the persons identified as the 41 document authors. All rights reserved. 43 This document is subject to BCP 78 and the IETF Trust's Legal 44 Provisions Relating to IETF Documents 45 (http://trustee.ietf.org/license-info) in effect on the date of 46 publication of this document. Please review these documents 47 carefully, as they describe your rights and restrictions with respect 48 to this document. Code Components extracted from this document must 49 include Simplified BSD License text as described in Section 4.e of 50 the Trust Legal Provisions and are provided without warranty as 51 described in the Simplified BSD License. 53 Table of Contents 55 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 56 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 3 57 3. Introduction to Stabilization in DHTs . . . . . . . . . . . . 5 58 3.1. Reactive vs. Periodic Stabilization . . . . . . . . . . . 5 59 3.2. Configuring Periodic Stabilization . . . . . . . . . . . 6 60 3.3. Adaptive Stabilization . . . . . . . . . . . . . . . . . 7 61 4. Introduction to Chord . . . . . . . . . . . . . . . . . . . . 8 62 5. Extending Chord-reload to Support Self-tuning . . . . . . . . 9 63 5.1. Update Requests . . . . . . . . . . . . . . . . . . . . . 10 64 5.2. Neighbor Stabilization . . . . . . . . . . . . . . . . . 10 65 5.3. Finger Stabilization . . . . . . . . . . . . . . . . . . 11 66 5.4. Adjusting Finger Table Size . . . . . . . . . . . . . . . 11 67 5.5. Detecting Partitioning . . . . . . . . . . . . . . . . . 12 68 5.6. Leaving the Overlay . . . . . . . . . . . . . . . . . . . 12 69 6. Self-tuning Chord Parameters . . . . . . . . . . . . . . . . 12 70 6.1. Estimating Overlay Size . . . . . . . . . . . . . . . . . 12 71 6.2. Determining Routing Table Size . . . . . . . . . . . . . 13 72 6.3. Estimating Failure Rate . . . . . . . . . . . . . . . . . 13 73 6.3.1. Detecting Failures . . . . . . . . . . . . . . . . . 14 74 6.4. Estimating Join Rate . . . . . . . . . . . . . . . . . . 15 75 6.5. Estimate Sharing . . . . . . . . . . . . . . . . . . . . 15 76 6.6. Calculating the Stabilization Interval . . . . . . . . . 17 77 7. Overlay Configuration Document Extension . . . . . . . . . . 17 78 8. Security Considerations . . . . . . . . . . . . . . . . . . . 18 79 9. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 18 80 9.1. Message Extensions . . . . . . . . . . . . . . . . . . . 18 81 10. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 19 82 11. References . . . . . . . . . . . . . . . . . . . . . . . . . 19 83 11.1. Normative References . . . . . . . . . . . . . . . . . . 19 84 11.2. Informative References . . . . . . . . . . . . . . . . . 19 85 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 21 87 1. Introduction 89 REsource LOcation And Discovery (RELOAD) [RFC6940] is a peer-to-peer 90 signaling protocol that can be used to maintain an overlay network, 91 and to store data in and retrieve data from the overlay. For 92 interoperability reasons, RELOAD specifies one overlay algorithm, 93 called chord-reload, that is mandatory to implement. This document 94 extends the chord-reload algorithm by introducing self-tuning 95 behavior. 97 Distributed Hash Table (DHT) based overlay networks are self- 98 organizing, scalable and reliable. However, these features come at a 99 cost: peers in the overlay network need to consume network bandwidth 100 to maintain routing state. Most DHTs use a periodic stabilization 101 routine to counter the undesirable effects of churn on routing. To 102 configure the parameters of a DHT, some characteristics such as churn 103 rate and network size need to be known in advance. These 104 characteristics are then used to configure the DHT in a static 105 fashion by using fixed values for parameters such as the size of the 106 successor set, size of the routing table, and rate of maintenance 107 messages. The problem with this approach is that it is not possible 108 to achieve a low failure rate and a low communication overhead by 109 using fixed parameters. Instead, a better approach is to allow the 110 system to take into account the evolution of network conditions and 111 adapt to them. This document extends the mandatory-to-implement 112 chord-reload algorithm by making it self-tuning. Two main advantages 113 of self-tuning are that users no longer need to tune every DHT 114 parameter correctly for a given operating environment and that the 115 system adapts to changing operating conditions. 117 The remainder of this document is structured as follows: Section 2 118 provides definitions of terms used in this document. Section 3 119 discusses alternative approaches to stabilization operations in DHTs, 120 including reactive stabilization, periodic stabilization, and 121 adaptive stabilization. Section 4 gives an introduction to the Chord 122 DHT algorithm. Section 5 describes how this document extends the 123 stabilization routine of the chord-reload algorithm. Section 6 124 describes how the stabilization rate and routing table size are 125 calculated in an adaptive fashion. 127 2. Terminology 129 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 130 "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and 131 "OPTIONAL" in this document are to be interpreted as described in RFC 132 2119 [RFC2119]. 134 This document uses the terminology and definitions from the Concepts 135 and Terminology for Peer to Peer SIP [I-D.ietf-p2psip-concepts] 136 draft. 138 numBitsInNodeId: Specifies the number of bits in a RELOAD Node-ID. 140 DHT: Distributed Hash Tables (DHTs) are a class of decentralized 141 distributed systems that provide a lookup service similar to a 142 regular hash table. Given a key, any peer participating in the 143 system can retrieve the value associated with that key. The 144 responsibility for maintaining the mapping from keys to values is 145 distributed among the peers. 147 Chord Ring: The Chord DHT uses ring topology and orders identifiers 148 on an identifier circle of size 2^numBitsInNodeId. This 149 identifier circle is called the Chord ring. On the Chord ring, 150 the responsibility for a key k is assigned to the node whose 151 identifier equals to or immediately follows k. 153 Finger Table: A data structure with up to (but typically less than) 154 numBitsInNodeId entries maintained by each peer in a Chord-based 155 overlay. The ith entry in the finger table of peer n contains the 156 identity of the first peer that succeeds n by at least 157 2^(numBitsInNodeId-i) on the Chord ring. This peer is called the 158 ith finger of peer n. As an example, the first entry in the finger 159 table of peer n contains a peer half-way around the Chord ring 160 from peer n. The purpose of the finger table is to accelerate 161 lookups. 163 n.id: An abbreviation that is in this document used refer to the 164 Node-ID of peer n. 166 O(g(n)): Informally, saying that some equation f(n) = O(g(n)) means 167 that f(n) is less than some constant multiple of g(n). For the 168 formal definition, please refer to [weiss1998]. 170 Omega(g(n)): Informally, saying that some equation f(n) = 171 Omega(g(n)) means that f(n) is more than some constant multiple of 172 g(n). For the formal definition, please refer to [weiss1998] 174 Percentile: The Pth (0<=P<=100) percentile of N values arranged in 175 ascending order is obtained by first calculating the (ordinal) 176 rank n=(P/100)*N, rounding the result to the nearest integer, and 177 then taking the value corresponding to that rank. 179 Predecessor List: A data structure containing the first r 180 predecessors of a peer on the Chord ring. 182 Successor List: A data structure containing the first r successors 183 of a peer on the Chord ring. 185 Neighborhood Set: A term used to refer to the set of peers included 186 in the successor and predecessor lists of a given peer. 188 Routing Table: Contents of a given peer's routing table include the 189 set of peers that the peer can use to route overlay messages. The 190 routing table is made up of the finger table, successor list and 191 predecessor list. 193 3. Introduction to Stabilization in DHTs 195 DHTs use stabilization routines to counter the undesirable effects of 196 churn on routing. The purpose of stabilization is to keep the 197 routing information of each peer in the overlay consistent with the 198 constantly changing overlay topology. There are two alternative 199 approaches to stabilization: periodic and reactive [rhea2004]. 200 Periodic stabilization can either use a fixed stabilization rate or 201 calculate the stabilization rate in an adaptive fashion. 203 3.1. Reactive vs. Periodic Stabilization 205 In reactive stabilization, a peer reacts to the loss of a peer in its 206 neighborhood set or to the appearance of a new peer that should be 207 added to its neighborhood set by sending a copy of its neighbor table 208 to all peers in the neighborhood set. Periodic recovery, in 209 contrast, takes place independently of changes in the neighborhood 210 set. In periodic recovery, a peer periodically shares its 211 neighborhood set with each or a subset of the members of that set. 213 The chord-reload algorithm [RFC6940] supports both reactive and 214 periodic stabilization. It has been shown in [rhea2004] that 215 reactive stabilization works well for small neighborhood sets (i.e., 216 small overlays) and moderate churn. However, in large-scale (e.g., 217 1000 peers or more [rhea2004]) or high-churn overlays, reactive 218 stabilization runs the risk of creating a positive feedback cycle, 219 which can eventually result in congestion collapse. In [rhea2004], 220 it is shown that a 1000-peer overlay under churn uses significantly 221 less bandwidth and has lower latencies when periodic stabilization is 222 used than when reactive stabilization is used. Although in the 223 experiments carried out in [rhea2004], reactive stabilization 224 performed well when there was no churn, its bandwidth use was 225 observed to jump dramatically under churn. At higher churn rates and 226 larger scale overlays, periodic stabilization uses less bandwidth and 227 the resulting lower contention for the network leads to lower 228 latencies. For this reason, most DHTs such as CAN [CAN], Chord 229 [Chord], Pastry [Pastry], Bamboo [rhea2004], etc. use periodic 230 stabilization [ghinita2006]. As an example, the first version of 231 Bamboo used reactive stabilization, which caused Bamboo to suffer 232 from degradation in performance under churn. To fix this problem, 233 Bamboo was modified to use periodic stabilization. 235 In Chord, periodic stabilization is typically done both for 236 successors and fingers. An alternative strategy is analyzed in 237 [krishnamurthy2008]. In this strategy, called the correction-on- 238 change maintenance strategy, a peer periodically stabilizes its 239 successors but does not do so for its fingers. Instead, finger 240 pointers are stabilized in a reactive fashion. The results obtained 241 in [krishnamurthy2008] imply that although the correction-on-change 242 strategy works well when churn is low, periodic stabilization 243 outperforms the correction-on-change strategy when churn is high. 245 3.2. Configuring Periodic Stabilization 247 When periodic stabilization is used, one faces the problem of 248 selecting an appropriate execution rate for the stabilization 249 procedure. If the execution rate of periodic stabilization is high, 250 changes in the system can be quickly detected, but at the 251 disadvantage of increased communication overhead. Alternatively, if 252 the stabilization rate is low and the churn rate is high, routing 253 tables become inaccurate and DHT performance deteriorates. Thus, the 254 problem is setting the parameters so that the overlay achieves the 255 desired reliability and performance even in challenging conditions, 256 such as under heavy churn. This naturally results in high cost 257 during periods when the churn level is lower than expected, or 258 alternatively, poor performance or even network partitioning in worse 259 than expected conditions. 261 In addition to selecting an appropriate stabilization interval, 262 regardless of whether periodic stabilization is used or not, an 263 appropriate size needs to be selected for the neighborhood set and 264 for the finger table. 266 The current approach is to configure overlays statically. This works 267 in situations where perfect information about the future is 268 available. In situations where the operating conditions of the 269 network are known in advance and remain static throughout the 270 lifetime of the system, it is possible to choose fixed optimal values 271 for parameters such as stabilization rate, neighborhood set size and 272 routing table size. However, if the operating conditions (e.g., the 273 size of the overlay and its churn rate) do not remain static but 274 evolve with time, it is not possible to achieve both a low lookup 275 failure rate and a low communication overhead by using fixed 276 parameters [ghinita2006]. 278 As an example, to configure the Chord DHT algorithm, one needs to 279 select values for the following parameters: size of successor list, 280 stabilization interval, and size of the finger table. To select an 281 appropriate value for the stabilization interval, one needs to know 282 the expected churn rate and overlay size. According to 283 [liben-nowell2002], a Chord network in a ring-like state remains in a 284 ring-like state as long as peers send Omega(square(log(N))) messages 285 before N new peers join or N/2 peers fail. Thus, in a 500-peer 286 overlay churning at a rate such that one peer joins and one peer 287 leaves the network every 30 seconds, an appropriate stabilization 288 interval would be on the order of 93s. According to [Chord], the size 289 of the successor list and finger table should be on the order of 290 log(N). Already a successor list of a modest size (e.g., log2(N) or 291 2*log2(N), which is the successor list size used in [Chord]) makes it 292 very unlikely that a peer will lose all of its successors, which 293 would cause the Chord ring to become disconnected. Thus, in a 294 500-peer network each peer should maintain on the order of nine 295 successors and fingers. However, if the churn rate doubles and the 296 network size remains unchanged, the stabilization rate should double 297 as well. That is, the appropriate maintenance interval would now be 298 on the order of 46s. On the other hand, if the churn rate becomes 299 e.g. six-fold and the size of the network grows to 2000 peers, on the 300 order of eleven fingers and successors should be maintained and the 301 stabilization interval should be on the order of 42s. If one 302 continued using the old values, this could result in inaccurate 303 routing tables, network partitioning, and deteriorating performance. 305 3.3. Adaptive Stabilization 307 A self-tuning DHT takes into consideration the continuous evolution 308 of network conditions and adapts to them. In a self-tuning DHT, each 309 peer collects statistical data about the network and dynamically 310 adjusts its stabilization rate, neighborhood set size, and finger 311 table size based on the analysis of the data [ghinita2006]. 312 Reference [mahajan2003] shows that by using self-tuning, it is 313 possible to achieve high reliability and performance even in adverse 314 conditions with low maintenance cost. Adaptive stabilization has 315 been shown to outperform periodic stabilization in terms of both 316 lookup failures and communication overhead [ghinita2006]. 318 4. Introduction to Chord 320 Chord [Chord] is a structured P2P algorithm that uses consistent 321 hashing to build a DHT out of several independent peers. Consistent 322 hashing assigns each peer and resource a fixed-length identifier. 323 Peers use SHA-1 as the base hash fuction to generate the identifiers. 324 As specified in RELOAD base, the length of the identifiers is 325 numBitsInNodeId=128 bits. The identifiers are ordered on an 326 identifier circle of size 2^numBitsInNodeId. On the identifier 327 circle, key k is assigned to the first peer whose identifier equals 328 or follows the identifier of k in the identifier space. The 329 identifier circle is called the Chord ring. 331 Different DHTs differ significantly in performance when bandwidth is 332 limited. It has been shown that when compared to other DHTs, the 333 advantages of Chord include that it uses bandwidth efficiently and 334 can achieve low lookup latencies at little cost [li2004]. 336 A simple lookup mechanism could be implemented on a Chord ring by 337 requiring each peer to only know how to contact its current successor 338 on the identifier circle. Queries for a given identifier could then 339 be passed around the circle via the successor pointers until they 340 encounter the first peer whose identifier is equal to or larger than 341 the desired identifier. Such a lookup scheme uses a number of 342 messages that grows linearly with the number of peers. To reduce the 343 cost of lookups, Chord maintains also additional routing information; 344 each peer n maintains a data structure with up to numBitsInNodeId 345 entries, called the finger table. The first entry in the finger 346 table of peer n contains the peer half-way around the ring from peer 347 n. The second entry contains the peer that is 1/4th of the way 348 around, the third entry the peer that is 1/8th of the way around, 349 etc. In other words, the ith entry in the finger table at peer n 350 contains the identity of the first peer s that succeeds n by at least 351 2^(numBitsInNodeId-i) on the Chord ring. This peer is called the ith 352 finger of peer n. The interval between two consecutive fingers is 353 called a finger interval. The ith finger interval of peer n covers 354 the range [n.id + 2^(numBitsInNodeId-i), n.id + 355 2^(numBitsInNodeId-i+1)) on the Chord ring. In an N-peer network, 356 each peer maintains information about O(log(N)) other peers in its 357 finger table. As an example, if N=100000, it is sufficient to 358 maintain 17 fingers. 360 Chord needs all peers' successor pointers to be up to date in order 361 to ensure that lookups produce correct results as the set of 362 participating peers changes. To achieve this, peers run a 363 stabilization protocol periodically in the background. The 364 stabilization protocol of the original Chord algorithm uses two 365 operations: successor stabilization and finger stabilization. 366 However, the Chord algorithm of RELOAD base defines two additional 367 stabilization components, as will be discussed below. 369 To increase robustness in the event of peer failures, each Chord peer 370 maintains a successor list of size r, containing the peer's first r 371 successors. The benefit of successor lists is that if each peer 372 fails independently with probability p, the probability that all r 373 successors fail simultaneously is only p^r. 375 The original Chord algorithm maintains only a single predecessor 376 pointer. However, multiple predecessor pointers (i.e., a predecessor 377 list) can be maintained to speed up recovery from predecessor 378 failures. The routing table of a peer consists of the successor 379 list, finger table, and predecessor list. 381 5. Extending Chord-reload to Support Self-tuning 383 This section describes how the mandatory-to-implement chord-reload 384 algorithm defined in RELOAD base [RFC6940] can be extended to support 385 self-tuning. 387 The chord-reload algorithm supports both reactive and periodic 388 recovery strategies. When the self-tuning mechanisms defined in this 389 document are used, the periodic recovery strategy MUST be used. 390 Further, chord-reload specifies that at least three predecessors and 391 three successors need to be maintained. When the self-tuning 392 mechanisms are used, the appropriate sizes of the successor list and 393 predecessor list are determined in an adaptive fashion based on the 394 estimated network size, as will be described in Section 6. 396 As specified in RELOAD base, each peer MUST maintain a stabilization 397 timer. When the stabilization timer fires, the peer MUST restart the 398 timer and carry out the overlay stabilization routine. Overlay 399 stabilization has four components in chord-reload: 401 1. Update the neighbor table. We refer to this as neighbor 402 stabilization. 404 2. Refreshing the finger table. We refer to this as finger 405 stabilization. 407 3. Adjusting finger table size. 409 4. Detecting partitioning. We refer to this as strong 410 stabilization. 412 As specified in RELOAD base [RFC6940], a peer sends periodic messages 413 as part of the neighbor stabilization, finger stabilization, and 414 strong stabilization routines. In neighbor stabilization, a peer 415 periodically sends an Update request to every peer in its Connection 416 Table. The default time is every ten minutes. In finger 417 stabilization, a peer periodically searches for new peers to include 418 in its finger table. This time defaults to one hour. This document 419 specifies how the neighbor stabilization and finger stabilization 420 intervals can be determined in an adaptive fashion based on the 421 operating conditions of the overlay. The subsections below describe 422 how this document extends the four components of stabilization. 424 5.1. Update Requests 426 As described in RELOAD base [RFC6940], the neighbor and finger 427 stabilization procedures are implemented using Update requests. 428 RELOAD base defines three types of Update requests: 'peer_ready', 429 'neighbors', and 'full'. Regardless of the type, all Update requests 430 include an 'uptime' field. Since the self-tuning extensions require 431 information on the uptimes of peers in the routing table, the sender 432 of an Update request MUST include its current uptime in seconds in 433 the 'uptime' field. 435 When self-tuning is used, each peer decides independently the 436 appropriate size for the successor list, predecessor list and finger 437 table. Thus, the 'predecessors', 'successors', and 'fingers' fields 438 included in RELOAD Update requests are of variable length. As 439 specified in RELOAD [RFC6940], variable length fields are on the wire 440 preceded by length bytes. In the case of the successor list, 441 predecessor list, and finger table, there are two length bytes 442 (allowing lengths up to 2^16-1). The number of NodeId structures 443 included in each field can be calculated based on the length bytes 444 since the size of a single NodeId structure is 16 bytes. If a peer 445 receives more entries than fit into its successor list, predecessor 446 list or finger table, the peer MUST ignore the extra entries. If a 447 peer receives less entries than it currently has in its own data 448 structure, the peer MUST NOT drop the extra entries from its data 449 structure. 451 5.2. Neighbor Stabilization 453 In the neighbor stabilization operation of chord-reload, a peer 454 periodically sends an Update request to every peer in its Connection 455 Table. In a small, low-churn overlay, the amount of traffic this 456 process generates is typically acceptable. However, in a large-scale 457 overlay churning at a moderate or high churn rate, the traffic load 458 may no longer be acceptable since the size of the connection table is 459 large and the stabilization interval relatively short. The self- 460 tuning mechanisms described in this document are especially designed 461 for overlays of the latter type. Therefore, when the self-tuning 462 mechanisms are used, each peer MUST send a periodic Update request 463 only to its first predecessor and first successor on the Chord ring. 465 The neighbor stabilization routine MUST be executed when the 466 stabilization timer fires. To begin the neighbor stabilization 467 routine, a peer MUST send an Update request to its first successor 468 and its first predecessor. The type of the Update request MUST be 469 'neighbors'. The Update request MUST include the successor and 470 predecessor lists of the sender. If a peer receiving such an Update 471 request learns from the predecessor and successor lists included in 472 the request that new peers can be included in its neighborhood set, 473 it MUST send Attach requests to the new peers. 475 After a new peer has been added to the predecessor or successor list, 476 an Update request of type 'peer_ready' MUST be sent to the new peer. 477 This allows the new peer to insert the sender into its neighborhood 478 set. 480 5.3. Finger Stabilization 482 Chord-reload specifies two alternative methods for searching for new 483 peers to the finger table. Both of the alternatives can be used with 484 the self-tuning extensions defined in this document. 486 Immediately after a new peer has been added to the finger table, a 487 Probe request MUST be sent to the new peer to fetch its uptime. The 488 requested_info field of the Probe request MUST be set to contain the 489 ProbeInformationType 'uptime' defined in RELOAD base [RFC6940]. 491 5.4. Adjusting Finger Table Size 493 The chord-reload algorithm defines how a peer can make sure that the 494 finger table is appropriately sized to allow for efficient routing. 495 Since the self-tuning mechanisms specified in this document produce a 496 network size estimate, this estimate can be directly used to 497 calculate the optimal size for the finger table. This mechanism MUST 498 be used instead of the one specified by chord-reload. A peer MUST 499 use the network size estimate to determine whether it needs to adjust 500 the size of its finger table each time when the stabilization timer 501 fires. The way this is done is explained in Section 6.2. 503 5.5. Detecting Partitioning 505 This document does not require any changes to the mechanism chord- 506 reload uses to detect network partitioning. 508 5.6. Leaving the Overlay 510 As specified in RELOAD base [RFC6940], a leaving peer SHOULD send a 511 Leave request to all members of its neighbor table prior to leaving 512 the overlay. The overlay_specific_data field MUST contain the 513 ChordLeaveData structure. The Leave requests that are sent to 514 successors MUST contain the predecessor list of the leaving peer. 515 The Leave requests that are sent to the predecessors MUST contain the 516 successor list of the leaving peer. If a given successor can 517 identify better predecessors than are already included in its 518 predecessor lists by investigating the predecessor list it receives 519 from the leaving peer, it MUST send Attach requests to them. 520 Similarly, if a given predecessor identifies better successors by 521 investigating the successor list it receives from the leaving peer, 522 it MUST send Attach requests to them. 524 6. Self-tuning Chord Parameters 526 This section specifies how to determine an appropriate stabilization 527 rate and routing table size in an adaptive fashion. The proposed 528 mechanism is based on [mahajan2003], [liben-nowell2002], and 529 [ghinita2006]. To calculate an appropriate stabilization rate, the 530 values of three parameters MUST be estimated: overlay size N, failure 531 rate U, and join rate L. To calculate an appropriate routing table 532 size, the estimated network size N can be used. Peers in the overlay 533 MUST re-calculate the values of the parameters to self-tune the 534 chord-reload algorithm at the end of each stabilization period before 535 re-starting the stabilization timer. 537 6.1. Estimating Overlay Size 539 Techniques for estimating the size of an overlay network have been 540 proposed for instance in [mahajan2003], [horowitz2003], 541 [kostoulas2005], [binzenhofer2006], and [ghinita2006]. In Chord, the 542 density of peer identifiers in the neighborhood set can be used to 543 produce an estimate of the size of the overlay, N [mahajan2003]. 544 Since peer identifiers are picked randomly with uniform probability 545 from the numBitsInNodeId-bit identifier space, the average distance 546 between peer identifiers in the successor set is (2^numBitsInNodeId)/ 547 N. 549 To estimate the overlay network size, a peer MUST compute the average 550 inter-peer distance d between the successive peers starting from the 551 most distant predecessor and ending to the most distant successor in 552 the successor list. The estimated network size MUST be calculated 553 as: 555 2^numBitsInNodeId 556 N = ------------------- 557 d 559 This estimate has been found to be accurate within 15% of the real 560 network size [ghinita2006]. Of course, the size of the neighborhood 561 set affects the accuracy of the estimate. 563 During the join process, a joining peer fills its routing table by 564 sending a series of Ping and Attach requests, as specified in RELOAD 565 base [RFC6940]. Thus, a joining peer immediately has enough 566 information at its disposal to calculate an estimate of the network 567 size. 569 6.2. Determining Routing Table Size 571 As specified in RELOAD base, the finger table must contain at least 572 16 entries. When the self-tuning mechanisms are used, the size of 573 the finger table MUST be set to max(ceiling(log2(N)), 16) using the 574 estimated network size N. 576 The size of the successor list MUST be set to ceiling(log2(N)). An 577 implementation MAY place a lower limit on the size of the successor 578 list. As an example, the implementation might require the size of 579 the successor list to be always at least three. 581 A peer MAY choose to maintain a fixed-size predecessor list with only 582 three entries as specified in RELOAD base. However, it is 583 RECOMMENDED that a peer maintains ceiling(log2(N)) predecessors. 585 6.3. Estimating Failure Rate 587 A typical approach is to assume that peers join the overlay according 588 to a Poisson process with rate L and leave according to a Poisson 589 process with rate parameter U [mahajan2003]. The value of U can be 590 estimated using peer failures in the finger table and neighborhood 591 set [mahajan2003]. If peers fail with rate U, a peer with M unique 592 peer identifiers in its routing table should observe K failures in 593 time K/(M*U). Every peer in the overlay MUST maintain a history of 594 the last K failures. The current time MUST be inserted into the 595 history when the peer joins the overlay. The estimate of U MUST be 596 calculated as: 598 k 599 U = --------, 600 M * Tk 602 where M is the number of unique peer identifiers in the routing 603 table, Tk is the time between the first and the last failure in the 604 history, and k is the number of failures in the history. If k is 605 smaller than K, the estimate MUST be computed as if there was a 606 failure at the current time. It has been shown that an estimate 607 calculated in a similar manner is accurate within 17% of the real 608 value of U [ghinita2006]. 610 The size of the failure history K affects the accuracy of the 611 estimate of U. One can increase the accuracy by increasing K. 612 However, this has the side effect of decreasing responsiveness to 613 changes in the failure rate. On the other hand, a small history size 614 may cause a peer to overreact each time a new failure occurs. In 615 [ghinita2006], K is set 25% of the routing table size. Use of this 616 approach is RECOMMENDED. 618 6.3.1. Detecting Failures 620 A new failure MUST be inserted to the failure history in the 621 following cases: 623 1. A Leave request is received from a neigbhor. 625 2. A peer fails to reply to a Ping request sent in the situation 626 explained below. If no packets have been received on a 627 connection during the past 2*Tr seconds (where Tr is the 628 inactivity timer defined by ICE [RFC5245]), a RELOAD Ping request 629 MUST be sent to the remote peer. RELOAD mandates the use of STUN 630 [RFC5389] for keepalives. STUN keepalives take the form of STUN 631 Binding Indication transactions. As specified in ICE [RFC5245], 632 a peer sends a STUN Binding Indication if there has been no 633 packet sent on a connection for Tr seconds. Tr is configurable 634 and has a default of 15 seconds. Although STUN Binding 635 Indications do not generate a response, the fact that a peer has 636 failed can be learned from the lack of packets (Binding 637 Indications or application protocol packets) received from the 638 peer. If the remote peer fails to reply to the Ping request, the 639 sender MUST consider the remote peer to have failed. 641 As an alternative to relying on STUN keepalives to detect peer 642 failure, a peer could send additional, frequent RELOAD messages to 643 every peer in its Connection Table. These messages could be Update 644 requests, in which case they would serve two purposes: detecting peer 645 failure and stabilization. However, as the cost of this approach can 646 be very high in terms of bandwidth consumption and traffic load, 647 especially in large-scale overlays experiencing churn, its use is NOT 648 RECOMMENDED. 650 6.4. Estimating Join Rate 652 Reference [ghinita2006] proposes that a peer can estimate the join 653 rate based on the uptime of the peers in its routing table. An 654 increase in peer join rate will be reflected by a decrease in the 655 average age of peers in the routing table. Thus, each peer MUST 656 maintain an array of the ages of the peers in its routing table 657 sorted in increasing order. Using this information, an estimate of 658 the global peer join rate L MUST be calculated as: 660 N 661 L = ----------------------, 662 Ages[floor(rsize/2)] 664 where Ages is an array containing the ages of the peers in the 665 routing table sorted in increasing order and rsize is the size of the 666 routing table. It has been shown that the estimate obtained by using 667 this method is accurate within 22% of the real join rate 668 [ghinita2006]. Of course, the size of the routing table affects the 669 accuracy. 671 In order for this mechanism to work, peers need to exchange 672 information about the time they have been present in the overlay. 673 Peers receive the uptimes of their successors and predecessors during 674 the stabilization operations since all Update requests carry uptime 675 values. A joining peer learns the uptime of the admitting peer since 676 it receives an Update from the admitting peer during the join 677 procedure. Peers learn the uptimes of new fingers since they can 678 fetch the uptime using a Probe request after having attached to the 679 new finger. 681 6.5. Estimate Sharing 683 To improve the accuracy of network size, join rate, and leave rate 684 estimates, peers MUST share their estimates. When the stabilization 685 timer fires, a peer MUST select number-of-peers-to-probe random peers 686 from its finger table and send each of them a Probe request. The 687 targets of Probe requests are selected from the finger table rather 688 than from the neighbor table since neighbors are likely to make 689 similar errors when calculating their estimates. number-of-peers-to- 690 probe is a new element in the overlay configuration document. It is 691 defined in Section 7 and has a default value of 4. Both the Probe 692 request and the answer returned by the target peer MUST contain a new 693 message extension whose MessageExtensionType is 'self_tuning_data'. 695 This extension type is defined in Section 9.1. The 696 extension_contents field of the MessageExtension structure MUST 697 contain a SelfTuningData structure: 699 struct { 700 uint32 network_size; 701 uint32 join_rate; 702 uint32 leave_rate; 703 } SelfTuningData; 705 The contents of the SelfTuningData structure are as follows: 707 network_size 709 The latest network size estimate calculated by the sender. 711 join_rate 713 The latest join rate estimate calculated by the sender. 715 leave_rate 717 The latest leave rate estimate calculated by the sender. 719 The join and leave rates are expressed as joins or failures per 24 720 hours. As an example, if the global join rate estimate a peer has 721 calculated is 0.123 peers/s, it would include in the join_rate field 722 the ceiling of the value 10627.2 (24*60*60*0.123 = 10627.2), that is, 723 the value 10628. 725 The 'type' field of the MessageExtension structure MUST be set to 726 contain the value 'self_tuning_data'. The 'critical' field of the 727 structure MUST be set to False. 729 A peer MUST store all estimates it receives in Probe requests and 730 answers during a stabilization interval. When the stabilization 731 timer fires, the peer MUST calculate the estimates to be used during 732 the next stabilization interval by taking the 75th percentile (i.e., 733 third quartile) of a data set containing its own estimate and the 734 received estimates. 736 6.6. Calculating the Stabilization Interval 738 According to [liben-nowell2002], a Chord network in a ring-like state 739 remains in a ring-like state as long as peers send 740 Omega(square(log(N))) messages before N new peers join or N/2 peers 741 fail. We can use the estimate of peer failure rate, U, to calculate 742 the time Tf in which N/2 peers fail: 744 1 745 Tf = ------ 746 2*U 748 Based on this estimate, a stabilization interval Tstab-1 MUST be 749 calculated as: 751 Tf 752 Tstab-1 = ----------------- 753 square(log2(N)) 755 On the other hand, the estimated join rate L can be used to calculate 756 the time in which N new peers join the overlay. Based on the 757 estimate of L, a stabilization interval Tstab-2 MUST be calculated 758 as: 760 N 761 Tstab-2 = --------------------- 762 L * square(log2(N)) 764 Finally, the actual stabilization interval Tstab that MUST be used 765 can be obtained by taking the minimum of Tstab-1 and Tstab-2. 767 The results obtained in [maenpaa2009] indicate that making the 768 stabilization interval too small has the effect of making the overlay 769 less stable (e.g., in terms of detected loops and path failures). 770 Thus, a lower limit should be used for the stabilization period. 771 Based on the results in [maenpaa2009], a lower limit of 15s is 772 RECOMMENDED, since using a stabilization period smaller than this 773 will with a high probability cause too much traffic in the overlay. 775 7. Overlay Configuration Document Extension 777 This document extends the RELOAD overlay configuration document by 778 adding one new element, "number-of-peers-to-probe", inside each 779 "configuration" element. 781 self-tuning:number-of-peers-to-probe: The number of fingers to which 782 Probe requests are sent to obtain their network size, join rate, 783 and leave rate estimates. The default value is 4. 785 This new element is formally defined as follows: 787 namespace self-tuning = "urn:ietf:params:xml:ns:p2p:self-tuning" 789 parameter &= element self-tuning:number-of-peers-to-probe { 790 xsd:unsignedInt } 792 This namespace is added into the element in the 793 overlay configuration file. 795 8. Security Considerations 797 In the same way as malicious or compromised peers implementing the 798 RELOAD base protocol [RFC6940] can advertise false network metrics or 799 distribute false routing table information for instance in RELOAD 800 Update messages, malicious peers implementing this specification may 801 share false join rate, leave rate, and network size estimates. For 802 such attacks, the same security concerns apply as in the RELOAD base 803 specification. In addition, as long as the amount of malicious peers 804 in the overlay remains modest, the statistical mechanisms applied in 805 Section 6.5 (i.e., the use of 75th percentiles) to process the shared 806 estimates a peer obtains help ensuring that estimates that are 807 clearly different from (i.e., larger or smaller than) other received 808 estimates will not significantly influence the process of adapting 809 the stabilization interval and routing table size. However, it 810 should be noted that if an attacker is able to impersonate a high 811 number of other peers in the overlay in strategic locations, it may 812 be able to send a high enough number of false estimates to a victim 813 and therefore influence the victim's choice of a stabilization 814 interval. 816 9. IANA Considerations 818 9.1. Message Extensions 820 This document introduces one additional extension to the "RELOAD 821 Extensions" Registry: 823 +------------------+-------+---------------+ 824 | Extension Name | Code | Specification | 825 +------------------+-------+---------------+ 826 | self_tuning_data | 3 | RFC-AAAA | 827 +------------------+-------+---------------+ 829 The contents of the extension are defined in Section 6.5. 831 Note to RFC Editor: please replace AAAA with the RFC number for this 832 specification. 834 10. Acknowledgments 836 The authors would like to thank Jani Hautakorpi for his contributions 837 to the document. The authors would also like to thank Carlos 838 Bernardos and Martin Durst for their comments on the document. 840 11. References 842 11.1. Normative References 844 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 845 Requirement Levels", BCP 14, RFC 2119, March 1997. 847 [RFC5245] Rosenberg, J., "Interactive Connectivity Establishment 848 (ICE): A Protocol for Network Address Translator (NAT) 849 Traversal for Offer/Answer Protocols", RFC 5245, April 850 2010. 852 [RFC5389] Rosenberg, J., Mahy, R., Matthews, P., and D. Wing, 853 "Session Traversal Utilities for NAT (STUN)", RFC 5389, 854 October 2008. 856 [RFC6940] Jennings, C., Lowekamp, B., Rescorla, E., Baset, S., and 857 H. Schulzrinne, "REsource LOcation And Discovery (RELOAD) 858 Base Protocol", RFC 6940, January 2014. 860 11.2. Informative References 862 [CAN] Ratnasamy, S., Francis, P., Handley, M., Karp, R., and S. 863 Schenker, "A Scalable Content-Addressable Network", In 864 Proceedings of the 2001 Conference on Applications, 865 Technologies, Architectures and Protocols for Computer 866 Communications pp. 161-172, August 2001. 868 [Chord] Stoica, I., Morris, R., Liben-Nowell, D., Karger, D., 869 Kaashoek, M., Dabek, F., and H. Balakrishnan, "Chord: A 870 Scalable Peer-to-peer Lookup Service for Internet 871 Applications", IEEE/ACM Transactions on Networking Volume 872 11, Issue 1, pp. 17-32, February 2003. 874 [I-D.ietf-p2psip-concepts] 875 Bryan, D., Matthews, P., Shim, E., Willis, D., and S. 876 Dawkins, "Concepts and Terminology for Peer to Peer SIP", 877 draft-ietf-p2psip-concepts-05 (work in progress), July 878 2013. 880 [Pastry] Rowstron, A. and P. Druschel, "Pastry: Scalable, 881 Decentralized Object Location and Routing for Large-Scale 882 Peer-to-Peer Systems", In Proceedings of the IFIP/ACM 883 International Conference on Distribued Systems Platforms 884 pp. 329-350, November 2001. 886 [binzenhofer2006] 887 Binzenhofer, A., Kunzmann, G., and R. Henjes, "A Scalable 888 Algorithm to Monitor Chord-Based P2P Systems at Runtime", 889 In Proceedings of the 20th IEEE International Parallel and 890 Distributed Processing Symposium (IPDPS) pp. 1-8, April 891 2006. 893 [ghinita2006] 894 Ghinita, G. and Y. Teo, "An Adaptive Stabilization 895 Framework for Distributed Hash Tables", In Proceedings of 896 the 20th IEEE International Parallel and Distributed 897 Processing Symposium (IPDPS) pp. 29-38, April 2006. 899 [horowitz2003] 900 Horowitz, K. and D. Malkhi, "Estimating Network Size from 901 Local Information", Information Processing Letters Volume 902 88, Issue 5, pp. 237-243, December 2003. 904 [kostoulas2005] 905 Kostoulas, D., Psaltoulis, D., Gupta, I., Birman, K., and 906 A. Demers, "Decentralized Schemes for Size Estimation in 907 Large and Dynamic Groups", In Proceedings of the 4th IEEE 908 International Symposium on Network Computing and 909 Applications pp. 41-48, July 2005. 911 [krishnamurthy2008] 912 Krishnamurthy, S., El-Ansary, S., Aurell, E., and S. 913 Haridi, "Comparing Maintenance Strategies for Overlays", 914 In Proceedings of the 16th Euromicro Conference on 915 Parallel, Distributed and Network-Based Processing pp. 916 473-482, February 2008. 918 [li2004] Li, J., Strinbling, J., Gil, T., Morris, R., and M. 919 Kaashoek, "Comparing the Performance of Distributed Hash 920 Tables Under Churn", Peer-to-Peer Systems III, volume 3279 921 of Lecture Notes in Computer Science Springer, pp. 87-99, 922 February 2005. 924 [liben-nowell2002] 925 Liben-Nowell, D., Balakrishnan, H., and D. Karger, 926 "Observations on the Dynamic Evolution of Peer-to-Peer 927 Networks", In Proceedings of the 1st International 928 Workshop on Peer-to-Peer Systems (IPTPS) pp. 22-33, March 929 2002. 931 [maenpaa2009] 932 Maenpaa, J. and G. Camarillo, "A Study on Maintenance 933 Operations in a Chord-Based Peer-to-Peer Session 934 Initiation Protocol Overlay Network", In Proceedings of 935 the 23rd IEEE International Parallel and Distributed 936 Processing Symposium (IPDPS) pp. 1-9, May 2009. 938 [mahajan2003] 939 Mahajan, R., Castro, M., and A. Rowstron, "Controlling the 940 Cost of Reliability in Peer-to-Peer Overlays", In 941 Proceedings of the 2nd International Workshop on Peer-to- 942 Peer Systems (IPTPS) pp. 21-32, February 2003. 944 [rhea2004] 945 Rhea, S., Geels, D., Roscoe, T., and J. Kubiatowicz, 946 "Handling Churn in a DHT", In Proceedings of the USENIX 947 Annual Technical Conference pp. 127-140, June 2004. 949 [weiss1998] 950 Weiss, M., "Data Structures and Algorithm Analysis in 951 C++", Addison-Wesley Longman Publishin Co., Inc. 2nd 952 Edition, ISBN:0201361221, 1998. 954 Authors' Addresses 956 Jouni Maenpaa 957 Ericsson 958 Hirsalantie 11 959 Jorvas 02420 960 Finland 962 Email: Jouni.Maenpaa@ericsson.com 964 Gonzalo Camarillo 965 Ericsson 966 Hirsalantie 11 967 Jorvas 02420 968 Finland 970 Email: Gonzalo.Camarillo@ericsson.com